<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Adevinta on Medium]]></title>
        <description><![CDATA[Stories by Adevinta on Medium]]></description>
        <link>https://medium.com/@adevinta?source=rss-b8cb87b84913------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*a3qR8Oy66YmcDJXzU_mC3A.png</url>
            <title>Stories by Adevinta on Medium</title>
            <link>https://medium.com/@adevinta?source=rss-b8cb87b84913------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 08:11:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@adevinta/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Lakehouse architecture to data mesh]]></title>
            <link>https://medium.com/adevinta-tech-blog/from-lakehouse-architecture-to-data-mesh-c532c91f7b61?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/c532c91f7b61</guid>
            <category><![CDATA[data-mesh]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[data-product-platform]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 10 Dec 2024 10:05:40 GMT</pubDate>
            <atom:updated>2024-12-10T10:05:40.386Z</atom:updated>
            <content:encoded><![CDATA[<h4>Evolving towards data product creation and automated governance in Adevinta Spain</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fa-FAD9X9KWwfqib" /></figure><p>By <a href="https://www.linkedin.com/in/marcplanaguma/">Marc Planagumà</a></p><p>Over the past three years, the Data Platform team in <a href="https://adevinta.com/brands/">Adevinta Spain</a> (CompaaS), which I have had the pleasure of leading, has undergone a profound transformation of its data infrastructure. This process has been based on the implementation of a lakehouse architecture using Databricks as well as the development of several key initiatives such as data contracts, data product frameworks etc. These initiatives have transformed the platform into an environment that not only adopts the principles of data mesh but also acts as a true catalyst for creating data products.</p><figure><img alt="Data mesh vision in Adevinta with domain driven data ownership architecture and federated computational governance working across the different marketplace verticals i.e. Real estate, jobs, motor, generalist and cross" src="https://cdn-images-1.medium.com/max/1024/1*33eNm7itaT7VAHXDi-zrNA.png" /></figure><p>We must place ourselves in the context of a platform that matured into a multimodal, multisite and multidomain architecture some time ago. It is a unique platform used by different teams to cater to various profiles. This structure fits perfectly into the data mesh paradigm. The four pillars of data mesh, as shown in the above image, align with the vision of a common but distributed platform, providing a unique work ecosystem for different domains and granting them greater autonomy.</p><p>This mesh has become the link that unifies all our domains into a single platform. However, transforming it into a mesh was not only about seeing the fit but also finding transformative initiatives that would make it happen.</p><h3>Transforming from a lakehouse architecture to data mesh</h3><p>The CompaaS team started from an existing layered architecture, which we adapted and evolved to comply with <a href="https://docs.gcp.databricks.com/en/lakehouse/medallion.html">Medallion</a> architecture. This initial architecture already had a layer-based division, but we made a significant effort to transform it into a structure that not only complied with a <a href="https://www.databricks.com/glossary/data-lakehouse">lakehouse</a> but also served as a foundation for implementing data mesh principles. Thus, each layer became a specialised workspace for different types of data products, facilitating greater autonomy and specialisation within the platform.</p><figure><img alt="Layers in the lakehouse, following the medallion architecture. The bronze, silver and gold layers all control different aspects of the architecture and data products" src="https://cdn-images-1.medium.com/max/1024/1*YGDGIVhOP2wpVmmyKhW1nA.png" /><figcaption>Layers in the lakehouse, following the medallion architecture</figcaption></figure><p>This effort allowed us to adapt the lakehouse to the principles of data mesh, representing a significant shift in how we managed data and data products. Each layer was redefined to serve as a focal point for different functions:</p><ul><li>The Bronze layer as a space for <strong>source-aligned data products</strong></li><li>The Silver layer as a space for <strong>aggregated data products</strong></li><li>The Gold layer as a space for <strong>consumer-aligned data products</strong></li></ul><p>The diagram above shows the evolution of the Medallion architecture to a structure that adheres to data mesh principles.</p><p>These layers do not need to be sequential. Instead, they are interconnected, forming a mesh structure that allows greater flexibility and efficiency in data usage. Underneath you can see the various initiatives that have been fundamental to this transformation and how each helps make the lakehouse platform a robust foundation for the different types of data products and specific domains within our company.</p><h3>Key transformative initiatives</h3><p>The transformative initiatives described below have been fundamental in converting Adevinta Spain’s lakehouse architecture into a distributed infrastructure aligned with data mesh principles. These initiatives have enabled the redefinition of each layer of the platform for better governance, efficiency and team autonomy, fostering the creation of data products in a decentralised and agile manner.</p><figure><img alt="Following the discovery framework the bronze layer focuses on source aligned data products, the silver layer works on aggregate data products and the gold layer on consumer aligned data products. They also include their corresponding suites." src="https://cdn-images-1.medium.com/max/1024/1*3REjNAZhhZObJuUNpRp8cQ.png" /><figcaption>Layers details with each of the suites populating them</figcaption></figure><p><strong>The key initiatives include:</strong></p><ul><li>The data ingestion suite with the implementation of data contracts to convert the bronze layer into a container of source-aligned data products</li><li>The domain marketplace to transform the silver layer into aggregated data products</li><li>The data consuming suite along with the discovery framework to convert the gold layer into a consumer-aligned data products zone</li></ul><h3>From bronze layer to source-aligned data products zone thanks to data contracts</h3><p>To achieve a proper bronze layer we created a declarative framework that allows the user to create data contracts, which the process will use to ingest data. These contracts ingest and perform quality checks based on events, creating a raw and event-oriented bronze layer with individualised events.</p><figure><img alt="Bronze layer detail, moving data from landing to history" src="https://cdn-images-1.medium.com/max/1024/1*AbPz87CmpGK-fkqKeaTvXw.png" /><figcaption>Bronze layer detail, moving data from landing</figcaption></figure><p>This layer provides the foundation for source-aligned data products, ensuring traceability and quality from their origin. This approach facilitates the creation of data products with high consistency and a clear mechanism for monitoring and quality verification.</p><p>All the details of the implementation of data contracts as a declarative data entry framework for the lakehouse are detailed in <a href="https://adevinta.com/techblog/creating-source-aligned-data-products-in-adevinta-spain/">this article</a>. These data contracts have been fundamental in ensuring the quality and consistency of data, facilitating interoperability and reuse between different teams. As seen in the diagram, events, objects or incoming data are consumed and automatically controlled based on the information and instructions of the data contract. The contract can, in real-time, determine if the data is correct, trigger errors and alerts, and also handle this received data for PII control and classification, automating GDPR compliance.</p><figure><img alt="Data contracts consumer diagram" src="https://cdn-images-1.medium.com/max/1024/1*FV75ydEyb5JLaQVy1TWP8w.png" /><figcaption>Data contracts consumer diagram</figcaption></figure><p>The article details how data contracts provide a structured means to define explicit agreements between data producers and consumers, establishing specifications regarding data quality, structure and usage conditions. This has been crucial in ensuring data management meets quality and traceability standards from the source, guaranteeing consistent and reliable information for all teams.</p><figure><img alt="Producers, consumers and facilitators relationships; they are all interconnected and the information flows in circles in both directions" src="https://cdn-images-1.medium.com/max/1024/1*zvQDThQoi7zx1CzOoNMWrA.png" /><figcaption>Producers, consumers and facilitators relationships</figcaption></figure><p>The implementation of this technique has allowed us to change data ingestion processes and workflows. We have resolved the classic bottleneck of third-party integrators who don’t know the data well enough to integrate it properly and can’t scale due to the large number of integration sources.</p><p>Being a declarative framework that doesn’t require code to integrate new data into the lakehouse democratises access and ensures that each source has someone capable of creating the contract. This person will be a direct owner of the incoming data and the producer will interact directly with the consumer. This allows them to negotiate, agree and discuss the best data to extract from a source and integrate into the lakehouse. This distribution of tasks shifts the integration effort to each source, allowing the integration to scale in a way that centralised approaches could not. These two effects allow the analytical and quality tasks to start even before data integration. This determines the quality of the raw material for all subsequent analytics. Therefore, this structure allows for greater confidence in data quality, improving interoperability between domains and facilitating data reuse — essential elements for successfully implementing data mesh architecture.</p><h3>From silver layer to aggregated data products zone thanks to semantic one big tables (or big entities)</h3><p>We transformed the silver layer into an aggregated data products area to later convert it into a <a href="https://medium.com/@axel.schwanke/semantic-layer-one-layer-to-serve-them-all-d0ef7eff1ffa">semantic layer</a>, where data aggregation is based on domain semantics rather than specific use cases. This approach makes data classification and organisation follow the business model, so that each aggregation corresponds to a specific domain entity. We adopted the concept of <a href="https://www.ssp.sh/brain/one-big-table/">one big table</a>, or big entities, to ensure that aggregations align with the entities of the business domain.</p><figure><img alt="Silver layer details, moving from bronze to silver" src="https://cdn-images-1.medium.com/max/1024/1*dIihaEnSS2yVKsHUKpne0g.png" /><figcaption>Silver layer details, moving from bronze to silver</figcaption></figure><p>These one big tables are expansive. They can contain multiple columns and be quite sparse, as not all columns are always populated. Each table contains all the objects of a specific domain in a single structure, aggregating multiple source-aligned data products, which are more granular, into a domain entity. In a distributed architecture, this approach <a href="https://blog.dataengineer.io/p/how-to-data-model-correctly-kimball">optimises</a> data access by avoiding the need for complex joins as all relevant information for a domain is found in a single table. This provides greater flexibility in data management while also posing additional privacy and governance challenges as the table grows.</p><figure><img alt="Aggregating data from source to big entity" src="https://cdn-images-1.medium.com/max/1024/1*QKbSi16nHdrTK_zUDaL82g.png" /><figcaption>Aggregating data from source to big entity</figcaption></figure><p>The use of one big table greatly facilitates data discovery and reuse. The integration of the semantic layer as an intermediary between complex data structures and end users allows the use of terminology aligned with the business, making it easier for both technical and non-technical users. This semantic layer helps improve decision-making and ensures effective governance, providing a single source of truth for the entire organisation.</p><p>Regarding the concept of the <a href="https://www.datagalaxy.com/en/blog/using-a-semantic-layer/">semantic layer</a>, the creation of big entities has enabled the development of a business ontology that maps major domain concepts with data modelling. Instead of a low-granularity ontology, which would correspond to the more specific objects of the bronze layer, we work with a medium or high-level ontology. This represents major business concepts aligned with the organisational structure of Adevinta. Each one big table maps one of these domain concepts, providing a clear and coherent view of the most relevant entities for the organisation, such as user, client, user behaviour, buyer, seller, publications etc.</p><figure><img alt="The creation of big entities has enabled the development of a business ontology that maps major domain concepts with data modelling. Instead of a low-granularity ontology, which would correspond to the more specific objects of the bronze layer, we work with a medium or high-level ontology." src="https://cdn-images-1.medium.com/max/1024/1*su24wTLUVZo-AjiyKxVYPA.png" /><figcaption>Modeling and domains</figcaption></figure><p>This combination of one big table and the semantic layer forms a <a href="https://www.dataversity.net/what-is-a-data-marketplace-and-why-does-it-matter/">data marketplace</a> oriented towards domains within the organisation. This facilitates data sharing and availability across domains, providing a structured environment for creating new data products and promoting decentralised governance. The data marketplace also acts as a catalyst for innovation, offering a framework in which teams can experiment in a safe and controlled environment. This functionality is inspired by an online marketplace, where teams can easily find and reuse available data products, promoting transparency and accountability within Adevinta.</p><p>For the creation of big entities, a declarative framework has been developed that allows defining and populating these tables similarly to data contracts. This domain mapping framework specifies in a configuration file which tables and attributes from the bronze layer (source-aligned data products) belong to which domain and how they are integrated. For example, all tracking object tables for behaviour in applications and websites have a configuration file that maps them to a behaviour entity, where all the attributes defining behaviour tracking can be found in a single table.</p><figure><img alt="Market consumer diagram" src="https://cdn-images-1.medium.com/max/1024/1*Mos6C8drCw19ntWwVuIRkQ.png" /><figcaption>Market consumer diagram</figcaption></figure><p>This framework also automates governance aspects, with automatic alerts to validate quality rules and other checks that incoming data must meet. Additionally, it allows for the automation of maintenance, indexing and table optimisation tasks. making its use more efficient, secure and cost-effective. <a href="https://adevinta.com/techblog/architecting-compliance-cost-effective-data-strategies-for-gdpr/">This article</a> describes in detail the optimisation strategies applied and the benefits to performance and security.</p><p>This architecture creates a cohesive data ecosystem that facilitates access and usability at different user levels. Users can make data-driven decisions independently and efficiently, within a governed environment aligned with Adevinta’s objectives.</p><h3>From gold layer to consumer-aligned data products zone through to the consuming suite</h3><p>The gold layer is the most used and diverse because it needs to support all analytical use cases from BI to AI. For this reason, we transformed it into a hosting layer for consumer-aligned data products by implementing a suite of frameworks known as the consuming suite. This suite creates an agile environment where data products can be developed with high governance, for data product production and prototyping.</p><p>Instead of a workspace for analytics assets, we have provided a platform for creating data products with all its attributes [DAUTNIVS]: Discoverable, addressable, understandable, trustworthy, natively accessible, interoperable, valuable (on its own) and secure.</p><figure><img alt="Self serve platform vs. data products platform" src="https://cdn-images-1.medium.com/max/1024/1*t80Keplcf_gq7paXzINmUA.png" /><figcaption>Self serve platform vs. data products platform</figcaption></figure><p>The consuming suite is composed of several frameworks divided into two main categories: data product manufacturing and industrialisation frameworks, and prototyping frameworks.</p><h3>Frameworks for data product manufacturing and industrialisation</h3><p>We have three key frameworks to create data products efficiently and with automated governance: Poseidon, Hermes and Transform It Yourself (TIY):</p><ul><li><strong>Poseidon</strong>: Extraction and construction of distributed pipelines based on Databricks. It includes integrated governance, support and quality components.</li><li><strong>Hermes</strong>: Creation of non-distributed pipelines. It’s ideal for cases where processing needs do not require a distributed architecture.</li><li><strong>Transform It Yourself (TIY)</strong>: Designed to work with data stored in data warehouses, TIY allows teams to create transformation pipelines using SQL/DBT. This facilitates governance and guided data transformation.</li></ul><figure><img alt="Gold layer detail, moving from silver to make the data consumable by people, apps etc." src="https://cdn-images-1.medium.com/max/1024/1*oOCzDrnqRdBcpEP6P5cwDQ.png" /><figcaption>Gold layer detail, moving from silver to make the data consumable</figcaption></figure><p>The different frameworks have the capabilities and permissions to move data between layers. The key advantage of providing complete frameworks, rather than just tools and configurations, is that these frameworks already incorporate governance rules defining which data is accessible and by whom, along with user identification. This allows governance as code, automating the enforcement of governance policies.</p><p>Additionally, these frameworks are adapted to different user experiences and profiles, providing flexibility that ensures different working methods yield a consistent result. When deployed to production, the final output has the same appearance and operation, with unified monitoring, support and troubleshooting., This allows the same support team to maintain various applications with economies of scale, increasing production deployment speed with security and reducing incidents.</p><p>These articles “<a href="https://adevinta.com/techblog/how-we-moved-from-local-scripts-and-spreadsheets-shared-by-email-to-data-products-part-1/">How we moved from local scripts and spreadsheets shared by email to Data Products</a>” and “<a href="https://adevinta.com/techblog/lakehouse-and-warehouse-sharing-data-between-environments/">Lakehouse and Warehouse: sharing data between environments</a>”, show how these frameworks have democratised the use of applications that were previously considered Shadow IT and brought them into production conditions. Additionally, they explain how the consuming suite acts as a bridge between the lakehouse and the data warehouse, facilitating interoperability between these two curated analytics work areas.</p><h3>Framework for data product prototyping</h3><p>The prototyping framework within the consuming suite, known as Atenea, is designed to facilitate the creation of data product prototypes with a relaxed focus on quality standards, allowing for great agility during the testing phase. Unlike production environments, Atenea offers flexibility for high-speed data exploration, accessing all system layers (bronze, silver and gold), including both productive and non-productive data. Prototype results can be persisted, but always in a pre-production or playground environment, without directly impacting production. This environment allows projects to be scalable and operational temporarily, helping to validate ideas without requiring all data to be fully industrialised. This is ideal for analytical scenarios such as dashboards, Machine Learning (ML), or Artificial Intelligence (AI), where ideas can be tested quickly and efficiently.</p><p>One of the key advantages of Atenea is that, if the prototypes use existing productised assets and data, the transition to production will be smoother. However, if the data used is not sufficiently industrialised, the framework still allows for proof of concept, but additional time will be required to bring the prototype to production.</p><figure><img alt="Prototyping in the data products platform looking at all the layers of the architecture" src="https://cdn-images-1.medium.com/max/1024/1*5JwTKc48WzGGZQPE_sP9sw.png" /><figcaption>Prototyping in the data products platform</figcaption></figure><p>Atenea can access all layers of the system, from the initial data sources through to the lakehouse and data warehouse. It stores its results in the gold layer (consumer-aligned data products). This allows Atenea to offer a centralised interface for data discovery and exploration.</p><p>Atenea is primarily based on Databricks technology, using <a href="https://docs.databricks.com/en/notebooks/index.html">Notebooks</a> and <a href="https://docs.databricks.com/en/jobs/index.html">Workflows</a> that enhance user experience and democratise access to data prototyping for everyone. While in production the company only works with pipelines and jobs, the prototyping environment with Databricks provides a highly accessible and flexible notebook interface, which is particularly useful for the quick creation and validation of prototypes. This integration with Databricks makes the framework easily usable for teams with varying technical abilities, thereby fostering the creation of innovative data products without high entry barriers.</p><p>The impact of Atenea has been particularly notable when it is necessary to accelerate development or test ideas before investing in the industrialisation of specific data that may not be used recurrently. This framework allows for one-time analysis (“one-shot”) or rapid hypothesis validation, thus speeding up the initial development phases. Although Atenea’s design is temporary (projects cannot be permanent), and support is limited (it lacks production-level support and cannot write directly to production), these restrictions ensure that prototypes demonstrating value undergo an appropriate industrialisation and production process. This allows them to scale, reduce costs and increase reuse.</p><h3>Conclusion: Automation of governance and scalability</h3><p>The automation of governance has been one of the central pillars of this transformation. The Data Platform team at Adevinta Spain (CompaaS) has worked to ensure that quality, security and compliance policies are applied automatically, reducing the need for manual interventions and minimising the risk of human errors.</p><figure><img alt="Integrated diagram of the whole data products platform including Atenea, all the layers and their corresponding suites and the data product lab and builders" src="https://cdn-images-1.medium.com/max/1024/1*oveo5Z6T0kRWDf6lyKinOA.png" /><figcaption>Integrated diagram of the whole data products platform</figcaption></figure><p>In the diagram, all the layers and data flows of the data platform are represented, as well as the initiatives that have been key in this transformation such as the implementation of a lakehouse and subsequent incorporation of data mesh concepts. This has allowed the lakehouse to function as a “hosting” environment for data products, accompanied by a complete set of tools to build them in various phases: ingestion, exploitation, prototyping and production. Together, these initiatives provide a robust proposal for implementing data mesh and combining a Lakehouse, a data warehouse, and a set of tools that facilitate decentralised, data product-oriented implementation.</p><figure><img alt="Data movements and transformations through the data products platform" src="https://cdn-images-1.medium.com/max/1024/1*1Bknhf166B8a3Jmjhv31Lw.png" /><figcaption>Data movements and transformations through the data products platform</figcaption></figure><p>One of the most important innovations of this architecture is the ability to transform the sequential data flows of the medallion architecture of the lakehouse into a mesh structure with nodes and connections distributed in all directions. Instead of viewing data flows as a unidirectional process, we see them as a mesh in which each area has a specialisation. For example, there is a data input specialised for ingestion, a curated data layer for domain aggregation and data flows focused on consuming and building specific use cases.</p><p>This combination of different flows utilises standard tools like the lakehouse and data warehouse, enabling the construction of a data mesh implementation that supports the four essential pillars: Data as a product, federated automated governance, self-serve platform and domain-driven data ownership architecture.</p><p><a href="https://adevinta.com/techblog/evolving-from-a-data-platform-to-a-data-products-platform/">This article</a> further details how the initiatives implemented in this platform have facilitated the evolution from a Data platform to a data products platform.</p><figure><img alt="Data platform evolution through its users: from no platform to data platform to data products platform" src="https://cdn-images-1.medium.com/max/1024/1*CaLghBjzEo93hF2a9-tuTA.png" /><figcaption>Data platform evolution through its users</figcaption></figure><p>In conclusion, we observe the differences between having no platform, having technology without governance and having a data products platform. In the scenario without a platform, the company faces a large amount of Shadow IT and a lack of control. With a centralised platform, technical divergence is avoided, but there may still be shortcomings in flexibility and team autonomy. Finally, with the data products platform, not only is technical divergence avoided, but processes, working methods and culture are standardised.</p><p>This journey has shown that it’s not only necessary to provide good technology and service, but that the data platform must also be treated as a complete and integral product. The analytical workspace becomes a common, scalable, democratic, secure, economical and high-quality environment, enabling the company to innovate and grow in its data management in a coherent and sustainable manner.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c532c91f7b61" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/from-lakehouse-architecture-to-data-mesh-c532c91f7b61">From Lakehouse architecture to data mesh</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Filters to Phrases: Our AI Revolution in Car Search]]></title>
            <link>https://medium.com/adevinta-tech-blog/from-filters-to-phrases-our-ai-revolution-in-car-search-2d7c73ca4886?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/2d7c73ca4886</guid>
            <category><![CDATA[natural-language-process]]></category>
            <category><![CDATA[search-engine-optimizati]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 12 Nov 2024 09:16:18 GMT</pubDate>
            <atom:updated>2024-11-12T09:16:18.675Z</atom:updated>
            <content:encoded><![CDATA[<h4>How a 15-day sprint transformed the way we connect cars and searches</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6GLjgwhbm1asO8HD" /></figure><p>By <a href="https://www.linkedin.com/in/david-manzano-barba/">David Manzano</a>, Engineering Mind and <a href="https://www.linkedin.com/in/mauro-ferrantelli/">Mauro Ferrantelli</a>, Product Manager and AI Enthusiast</p><h3>Introduction</h3><p>As the first light of dawn broke over Barcelona, our team huddled together to kick off a bold new initiative. We are the minds behind the search experience at <a href="http://coches.net/">coches.net</a>, Spain’s leading vehicle marketplace. Excitement mixed with nervous energy filled the room as we embarked on a mission: to revolutionise our car search functionality with the power of artificial intelligence. We began the project in secret, and gave ourselves just 15 days to bring this vision to life.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4CoUMzMRAj_zghYZ" /><figcaption>Example of an ad on our marketplace coches.net</figcaption></figure><h3>The Spark of Innovation</h3><p>It started with a simple observation. We noticed a colleague struggling with our existing search system, applying multiple filters just to find a specific car model. It became clear that there must be a more efficient way to search our site. Our users deserved the convenience of typing natural phrases and getting accurate results.</p><p>Our car search system faced several challenges:</p><ul><li><strong>Complexity</strong>: Users wanted to search using natural language like “Audi A3 2020 in Barcelona” but had to navigate through several filters instead.</li><li><strong>User Experience</strong>: The system wasn’t intuitive, requiring users to know which specific filters to apply, such as “environmental stickers needed in Madrid”.</li></ul><h3>Discovering the AI Solution</h3><p>While experimenting with Amazon Bedrock for another project, we realised its potential for natural language processing. We developed a proof of concept and shared it with the development team. The results were promising, and we agreed that implementing this feature could significantly improve the user experience.</p><p>Our goal was to enable searches like “family-friendly SUV for mountain trips” or “sporty convertible under €30,000 in Madrid,” making the search process more intuitive.</p><h3>The 15-Day Challenge</h3><p>We set a self-imposed deadline of 15 days for this sprint to develop the AI-powered search. Wanting to avoid spending too much time on an unapproved initiative, we took a slightly unconventional approach. We dedicated all of our efforts to bringing this feature to life without attracting attention until we were sure it would work.</p><p>We worked diligently, refining the system and testing its capabilities. Our approach was collaborative and focused to ensure we stayed aligned with our objectives.</p><h3>Technical Approach</h3><p>We used AI to translate natural language queries into our existing search filters. The AI doesn’t perform the search itself; it converts the user’s input into filter parameters that our system can process.</p><p>Here is a simple flow chart with the steps of the search process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Uaw0MQ_YQoF-ClvF" /><figcaption>Search process flow</figcaption></figure><p>As you can see, the user’s query is first moderated to ensure it’s appropriate and relevant to car search. If it meets the criteria, we instruct the AI model to translate the query into our filters. We provided the AI with context about our filters and examples to enhance accuracy.</p><p>After verifying the AI’s response, we send the filters to the Frontend, where the website applies them to perform the search.</p><h3>Moderation</h3><p>It’s necessary to moderate user questions, as we should not and cannot answer questions that are not ethic, nor questions that have nothing to do with car searches.</p><p>Imagine that the user suggests looking for a car to carry out an illegal activity; we should not pass this query to our search filters. That is why we will moderate all messages before forwarding the question to the AI. Also, this pre-filtering action is quite cheap (in processing terms) and saves us from including the whole context of coches.net.</p><p>How do we do it? We simply send the user’s question to the AI model and ask it to return the following codes in its response:</p><ul><li>0: if the question is valid</li><li>1: if the question is unethical</li><li>2: if the question is invalid</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FbvVyaQ87ZCgpfxT" /><figcaption>Moderation flow</figcaption></figure><h3>AI Engine Selection</h3><p>We chose <a href="https://aws.amazon.com/bedrock/claude/?nc1=h_ls">Amazon Bedrock’s</a> generative AI service, aligning with our AWS infrastructure. Specifically, we used <a href="https://aws.amazon.com/bedrock/claude/?nc1=h_ls">Anthropic’s Claude</a> model, opting for the <strong>Haiku</strong> version due to its balance of speed and cost-effectiveness.</p><h3>Prompt Engineering</h3><p>We employed the <a href="https://www.promptingguide.ai/techniques/fewshot"><strong>Few-shot prompting</strong></a> technique, supplying the AI with context about our filters and examples of searches. This approach helps the AI generate accurate and relevant responses.</p><h3>Cost Considerations</h3><p>In AI processing, “tokens” represent units of text that the AI reads and generates. Each token consumes computational resources, affecting latency and cost.</p><p>Our AI search generates approximately <strong>6,000 input tokens</strong> and <strong>500 output tokens</strong> per query. Using the Haiku model:</p><ul><li><strong>Price per 1,000 input tokens</strong>: \$0.00025</li><li><strong>Price per 1,000 output tokens</strong>: \$0.00125</li></ul><p>With our current traffic on coches.net:</p><ul><li>Around 20,000 model invocations per day</li><li>55 million input tokens</li><li>4 million output tokens</li></ul><p>This results in a cost of about <strong>€19 per day</strong>, which is a reasonable value for the improved user experience.</p><h3>The Unveiling</h3><p>As we concluded our sprint, we prepared to demonstrate the new feature to our stakeholders. The product demo showcased the AI-powered search bar’s capabilities, handling complex queries efficiently. The response was positive, and we were encouraged by the potential benefits for our users.</p><h3>The Impact</h3><p>Implementing AI in our car search system yielded notable improvements:</p><ul><li><strong>Enhanced Efficiency</strong>: Users utilising the AI search generated more value compared to traditional methods.</li><li><strong>Improved User Experience</strong>: The AI search applies necessary filters automatically, including specific requirements like environmental stickers in Madrid.</li></ul><p>Feedback from users has been encouraging. One user mentioned:</p><p><em>“I typed ‘eco-friendly family car for city living,’ and it provided suitable options, considering Madrid’s environmental sticker requirements. It made finding my ideal car much easier!”</em></p><p>However, we recognise that adoption is still limited. A small portion of our traffic uses filters, and fewer are leveraging the new AI-powered search. We are working to improve the user experience and encourage more users to try this feature.</p><h3>Lessons Learned</h3><p>Our experience taught us several valuable lessons:</p><ol><li><strong>Importance of Data</strong>: Quality and quantity of data significantly impact AI performance.</li><li><strong>Team Collaboration</strong>: A focused and dedicated team can achieve meaningful results in a short time.</li><li><strong>Continuous Improvement</strong>: Iterative development is essential. We need to refine the feature based on user feedback and evolving requirements.</li></ol><h3>Looking Ahead</h3><p>We plan to continue enhancing our new AI search function, making it more intuitive and accessible. Our filters are evolving, and we aim to improve usability to increase adoption and impact on our users and metrics.We encourage others in the tech community to share their experiences and insights. If you have questions or feedback, please reach out. Your input is valuable as we refine and develop this technology further.Implementing AI in our car search tool has improved our system and user experience. This project reinforces our commitment to innovation and continuous improvement.</p><p><strong>Thank you to all the contributions of our team:</strong></p><p><a href="https://www.linkedin.com/in/marc-garrido-64159ba/">Marc Garrido, UX</a></p><p><a href="https://www.linkedin.com/in/nacho-torrella/">Nacho Torrella, FE</a></p><p><a href="https://www.linkedin.com/in/emilio-m-villuendas-zambrana-86048b22/">Emilio Villuendas, FE</a></p><p><a href="https://www.linkedin.com/in/ivan-garza-sanchez/">Ivan Garza, And</a></p><p><a href="https://www.linkedin.com/in/jacobo-rodriguez/">Jacobo Rodriguez, iOS</a></p><p><a href="https://www.linkedin.com/in/joan-santaeug%C3%A8nia-2aa300215/">Joan Santaeugenia, BE</a></p><p><a href="https://www.linkedin.com/in/jesusiglesiasiglesias/">Jesus Iglesias, BE</a></p><p><a href="https://www.linkedin.com/in/artxon-larrakoetxea-04a391109/">Artxon Larrakoetxea, DATA</a></p><p><a href="https://www.linkedin.com/in/david-manzano-barba/">David Manzano, BE</a></p><p><a href="https://www.linkedin.com/in/mauro-ferrantelli/">Mauro Ferrantelli, PM</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2d7c73ca4886" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/from-filters-to-phrases-our-ai-revolution-in-car-search-2d7c73ca4886">From Filters to Phrases: Our AI Revolution in Car Search</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artemis: Fostering Autonomy in Monitoring Data Quality]]></title>
            <link>https://medium.com/adevinta-tech-blog/artemis-fostering-autonomy-in-monitoring-data-quality-df59c0d86ef7?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/df59c0d86ef7</guid>
            <category><![CDATA[metrics]]></category>
            <category><![CDATA[data-monitoring]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 25 Jun 2024 08:13:28 GMT</pubDate>
            <atom:updated>2024-06-27T08:16:54.537Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A home-grown tool is revolutionising the monitoring and interpretation of key product metrics at Adevinta Benelux</em></p><p>By <a href="https://www.linkedin.com/in/beratyasa/">Berat Yaṣa</a>, <a href="https://www.linkedin.com/in/kowshihapaulvannan/">Kowshiha Paulvannan</a>, <a href="https://www.linkedin.com/in/riccardo-maria-scotti/">Riccardo Maria Scotti</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GicsZSKYGjKRPeZd" /></figure><p><em>Could a development we’ve made unintentionally break something else without us being aware of it? If a metric we are tracking starts to give incorrect values, how can we quickly identify it?</em></p><p>Within a dynamic working environment like <a href="https://adevinta.com/">Adevinta’s</a>, it’s not uncommon for even the slightest tweak in code to ripple through our metrics, causing unexpected disruptions. Given that we meticulously monitor <strong>over 140 metrics</strong>, it’s naturally challenging to catch every fluctuation. Consequently, deviations, especially in pivotal metrics, might only come to light well after the initial disturbance, casting shadows on our reporting and experimentation efforts.</p><p>This was the case when we encountered a marked decline in the Leads metric on the iOS platform. This wasn’t just a minor dip; we’re talking about a staggering drop of approximately 100K leads daily. To put this into perspective, that’s a <strong>negative</strong> impact to daily <strong>revenue of about 41K euros</strong>.</p><p>Thanks to the alerting and dashboard system from <strong>Artemis</strong>, we were able to detect the drop quickly, and resolve it in <strong>less than a day</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LEFSg5EgP1CQx9Fn1WRUdA.png" /><figcaption>Observed drop in Total Leads on iOS</figcaption></figure><h3>Who are we?</h3><p>We are a passionate group of analysts at <a href="https://adevinta.com/">Adevinta Benelux</a>, focused on data quality. We started Artemis as a side project in Q3, 2023. As the main source of quantitative insights in our cross-functional product teams, we are heavily relied on to identify and resolve data quality issues. However, having access to clean data, and subsequent insights, is key to driving success for the team on the whole. So we started asking ourselves: <em>How might we encourage ownership of data quality across the whole team? How can we identify issues and resolve them in a timely way? </em>Driven by an Adevinta-wide push for improving data quality, we set out to create a tool that is easy to use, managed centrally, and prompts data-driven thinking.</p><h3>Introducing Artemis</h3><p>Artemis, at its core, is about turning the data we navigate daily into actionable insights, ensuring our marketplaces like <a href="https://adevinta.com/brand/marktplaats/">Marktplaats</a> and others within Adevinta remain at the pinnacle of operational excellence. Its vigilance follows a cyclical path — from observing deviations in our metrics to alerting our teams to instigating a swift investigation and steering towards a solution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MK5rSzlUdj1KMvQM" /><figcaption>Image 2: Lifecycle of the Artemis system</figcaption></figure><p>When there is an observed issue, we send customised alerts on Slack, making sure to tag the right team, indicating the platform that’s flagged and the metrics that are affected.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KESaCA4s_seXowJv" /><figcaption>Image 3: Customised &amp; team-specific Artemis alert</figcaption></figure><p>This helps teams to swiftly address the problem; pinpointing the core issue and releasing a fix. This ability to respond rapidly underscores the tool’s indispensable role in maintaining the smooth operation and resilience of our systems.</p><p><em>Artemis isn’t just another tool; it’s a paradigm shift in how we approach data health within our ecosystems</em><strong><em>.</em></strong></p><h3>The Heart of Artemis: Predictive Analysis and Proactive Health</h3><p>You might be wondering: <em>How exactly does Artemis work? How accurate are the alerts?</em></p><h4>Predicting with Prophet</h4><p>At the helm of Artemis’ predictive prowess is <a href="https://facebook.github.io/prophet/">Prophet</a>, the open-source software released by Meta’s Core Data Science team.</p><p>Prophet is a tool for forecasting time series data based on an additive model. Non-linear trends are fit with yearly, weekly and daily seasonality, as well as holiday effects. It works best with time series that have strong seasonal effects. Prophet is robust to missing data and shifts in the trend and typically handles outliers well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VEp1jXnAPPkMvJUG4akVSg.png" /><figcaption>Image 4: Prophet’s main features based on Meta’s official website.</figcaption></figure><p>The model enables Artemis to look into the future, predicting the upper and lower bounds of our key metrics, ensuring we’re always a step ahead.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nrC5KrpRiAGEAgSH" /><figcaption>Image 5: Artemis data model using Prophet ML flow</figcaption></figure><h4>Optimising operations and model execution</h4><p>The model operates within the Databricks scheduling ecosystem and its codebase is securely housed in a GitHub repository. This setup not only grants us oversight of versioning history but also harmonises the execution schedule within an AWS-compatible framework.</p><p>To streamline the model’s operations and maintain its autonomy, metric names and definitions are dynamically sourced from a JSON file curated by the Adevinta Benelux teams. This system simplifies the process of modifying the tool’s metric roster, eliminating the need for manual adjustments and enhancing our agility in responding to evolving data landscapes.</p><p>For optimal efficiency and cost-effectiveness, the tool processes only the most recent day’s metric data, which is then stored in a partitioned table within S3. The model is trained daily using an extensive historical dataset — excluding the latest 10 days to mitigate the impact of any recent anomalies — and sets its sights on forecasting the next 40 days.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/980/0*rLny_23K_bePzAtK" /><figcaption>Image 6: Artemis alerting process flow</figcaption></figure><p>To elevate the model’s performance, we’ve fined-tuned it with some manual adjustments:</p><ol><li><strong>Mitigating Outliers: </strong>We’ve instituted a method where any data points straying beyond 3 standard deviations are tempered, and replaced by the average of the previous 28 days. This strategy significantly diminishes bias from temporary data issues (e.g. a sudden metric drop) over the model’s future performance.</li><li><strong>Selective Historical Data Curation: </strong>In instances of substantial transitions (such as a major change on the ad page), we manually remove the affected historical data pre-transition from our predictions. Though this method reduces our data history, it has demonstrably sharpened the model’s predictive accuracy.</li><li><strong>Enhanced Predictive Buffer:</strong> Despite Prophet’s design for precision forecasting, our primary objective leans towards identifying metric disruptions. To this end, we’ve opted to increase the predicted boundaries by 10%. This adjustment aims to dial down the rate of false positives, even if it slightly compromises the model’s accuracy.</li><li><strong>Relevant Alerts: </strong>In case a metric is out of range for three consecutive days, we added an alert on a specific Slack channel that directly tags the responsible team in order to ensure a rapid and effective response.</li></ol><h3>Navigating the Future with Artemis</h3><p>Artemis embodies the principles of the <a href="https://medium.com/adevinta-tech-blog/building-a-data-mesh-to-support-an-ecosystem-of-data-products-at-adevinta-4c057d06824d"><strong>data mesh architecture</strong></a>. Data mesh is a conceptual framework that focuses on four key pillars, which Artemis not only adheres to but brings to life within our operations.</p><ol><li><strong>Domain-Oriented Distributed Architecture</strong>: Artemis is built with a clear understanding of the distinct domains it serves. It decentralises data ownership and architecture, ensuring that data is treated as a product. This means that each domain within Adevinta Benelux can manage and govern its own data landscape.</li><li><strong>Product Thinking </strong><a href="https://adevinta.com/techblog/evolving-from-a-data-platform-to-a-data-products-platform/"><strong>(Data as a Product)</strong></a>: With Artemis, data is not a byproduct of our operations; it is the product. This mindset shift is crucial, as it focuses on creating data products that deliver value to both internal stakeholders and external customers. Artemis supports data that is accessible, reliable and effectively used to drive decision-making processes across all levels of Adevinta.</li><li><strong>Self-Service Infrastructure as Platforms</strong>: Artemis champions the principle of self-service by enabling teams to interact with data without the need for specialised intermediaries. By offering a platform that teams can use independently, Artemis facilitates a more dynamic and agile approach to data quality.</li><li><strong>Federated Computational Governance: </strong>Artemis supports the principle of federated governance. While the autonomy of domains is respected, Artemis provides a governance model that ensures data quality across the entire organisation.</li></ol><p>The journey of Artemis is a testament to the power of the data mesh approach. It sets the stage for a future where every strategic move is informed by high-quality data.</p><h3>In Conclusion: A New Era for Adevinta</h3><p>As we continue to expand the reach of Artemis, we are deeply grateful for the unwavering support provided by the product teams throughout this process.<em> </em>We look forward to continued collaboration and development.</p><p>Artemis marks the dawn of a new era in data analytics at Adevinta. It’s a testament to our commitment to innovation, our dedication to data health and our vision for a future where every piece of data empowers us to create better experiences for our users across the globe.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df59c0d86ef7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/artemis-fostering-autonomy-in-monitoring-data-quality-df59c0d86ef7">Artemis: Fostering Autonomy in Monitoring Data Quality</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How we moved from local scripts and spreadsheets shared by email to Data Products]]></title>
            <link>https://medium.com/adevinta-tech-blog/how-we-moved-from-local-scripts-and-spreadsheets-shared-by-email-to-data-products-edaec9228753?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/edaec9228753</guid>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Thu, 04 Apr 2024 07:31:27 GMT</pubDate>
            <atom:updated>2024-05-14T13:41:43.176Z</atom:updated>
            <content:encoded><![CDATA[<h4>Part I</h4><p>By: <a href="/service/mailto:marta.diaz@adevinta.com">Marta Diaz Lopez</a> &amp; <a href="/service/mailto:sergio.couto@adevinta.com">Sergio Couto</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7asB7Lq2ug32nMQR" /></figure><h3>Introduction</h3><p><a href="https://adevinta.com/">Adevinta</a> is a global online classifieds platform with a presence in over 10 countries worldwide, facilitating a secure and substantial volume of transactions to our users (2.5 billion monthly visits worldwide). <a href="https://adevinta.com/brands/?_gl=1*tfwm87*_up*MQ..*_ga*ODUzMjc4MTA0LjE3MTE0NDY3ODE.*_ga_F3G37WMXH8*MTcxMTQ0Njc4MS4xLjAuMTcxMTQ0Njc4MS4wLjAuMA..*_ga_JMYMMK7V90*MTcxMTQ0Njc4MS4xLjAuMTcxMTQ0Njc4MS4wLjAuNzY2MzUxMDE5">In Spain</a>, we proudly stand as the largest online classifieds group, where 1 out of 2 internet users connects to our platforms daily. Consequently, we are entrusted with managing a vast amount of data (with an input of more than 2Tb of data and 43 million events per day), bearing a significant responsibility to handle this data in a secure, responsible and sustainable manner.</p><p>We view <a href="https://adevinta.com/techblog/evolving-from-a-data-platform-to-a-data-products-platform/?_gl=1*1lkjd5t*_up*MQ..*_ga*MjMyNDc4MDUzLjE3MTE1Mjk0MzE.*_ga_F3G37WMXH8*MTcxMTUzODQ1Ni4yLjEuMTcxMTUzODQ2NC4wLjAuMA..*_ga_JMYMMK7V90*MTcxMTUzODQ1Ni4yLjEuMTcxMTUzODQ2NC4wLjAuMTY4NzQ2NjUwMA..">our data platform</a> not merely as a service but also as a product that acts as the crucial bridge between business strategy and technological execution.</p><p><em>“A product mindset is a strategic and customer-centric approach to problem-solving and innovation that prioritises continuous learning, iteration and delivering value to users to achieve long-term success.”</em></p><p>Our product vision for CompaaS (the Data Platform team in Adevinta Spain) is centred on empowering every Adevintan in Spain with scalable access to high-quality data through a Common Platform and Data as a Product. It aims to maximise the value generated from data throughout the entire value chain, promoting autonomous value extraction and expediting the generation of data products through robust quality processes and governance automation.</p><p>To realise this vision, we concentrate on maximising the value of data across the entire data value chain, with data products as the minimal unit. We prioritise reusability, security and scalability as the primary drivers for maximising data value.</p><p>Supporting this product vision and goals is one of our foundational pillars which we term “self-serve data product builders.” These are frameworks comprising infrastructure, computational governance and services. Our frameworks provide users with a unified, scalable, understandable and trusted means of generating valuable data products. By offering tools and processes to our Data Analysts, Data Scientists and data enthusiasts, we enable them to focus on extracting value while leaving operational aspects to our data product builders.</p><p>Our journey in discovering these data product builders with users begins with quick wins, guided by a clear product mindset. One notable quick win involved identifying scripts that were running locally and also replacing manual processes which had low reusability potential and trust. This approach ensures that our users can concentrate on value extraction and leave behind tasks that can be more efficiently handled through automation.</p><p>Building our new data product always begins with hands-on collaboration with end users, learning from them at every stage from conception through planning, implementation and subsequent iterations. Rather than viewing problem-solving as a checklist of user requirements, we advocate for a collaborative approach where we work together towards a common goal of addressing identified outcomes. We believe in co-creating solutions that align with the company’s strategy, maximising the value of data throughout the entire data value chain.</p><h3>Where we came from</h3><p>As stated before, we spent a lot of time collaborating with users so we can build our data products to match their needs and help the company extract value from data. We found that some Analysts and Data Scientists were using a workflow involving local environments and we (both the data platform and the users themselves) started planning a way to improve that.</p><p>In the existing workflow, the user acquires data from various sources or even generates it locally to serve as input for their script. Because it can be processed in a local environment, this input data is generally not more than a hundred Mb in size.</p><p>Users would manually execute the scripts, and the resulting output data, typically in CSV or Excel format, is either analysed by the users themselves or utilised as input for subsequent processes. In some cases, users may also share the output data via email with others to perform similar analyses.</p><p>This approach exhibits several disadvantages, some more apparent than others, here is a list of the ones we found:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*nGl57zKUmhJMUtVV" /></figure><p><strong>Versioning</strong></p><p>Without a centralised repository such as GitHub, tracking changes and accessing previous script versions becomes challenging. Identifying errors and reverting to the last working version may also be difficult in this setup, requiring much manual work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*z5JUIQYTSxGECgIK" /></figure><p><strong>Scheduling</strong></p><p>Since the script is executed in a local environment upon manual request, the entire responsibility for its timely execution falls on the owner. If the owner is dealing with an urgent matter, arrives late to work, or encounters any other issues, no matter if private or work-related, the process won’t run at the intended time, potentially leading to delays in critical business decisions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*wSZaROirSWMzb2n6" /></figure><p><strong>Knowledge silos</strong></p><p>Storing the “production” version of the script on a local machine creates a challenge in terms of knowledge sharing. Additionally, sharing the script by mail will create an alternative “production” version that may divert from the original, potentially resulting in two distinct output sets of data for the same input data and script.</p><p>Furthermore, If a similar process is developed, it will be done independently, starting from scratch rather than leveraging shared knowledge. It may be done from scratch instead of starting from the shared knowledge within the company, leading the company to redundancy in efforts, rework and lack of consistency in processes.</p><p>Linked to the previous point regarding the difficulty to develop similar processes, the same applies to reusability. Since the script is stored on a local computer, it is impossible to reuse. Reusability should be one of the main goals in a data product platform</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*J6IKYAXtJrR_oe-t" /></figure><p><strong>Security</strong></p><p>Sending data via mail possesses inherent risks, raising concerns about confidentiality. There is no way to ensure the email is addressed to the correct person for instance. Moreover, emails are susceptible to interception, potentially leading to data leakage, breaking the trust and our duty as custodians of the customers’ data or even violation of data privacy regulations. This reduces the traceability needed for our data protection commitment with clients and users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*KX8t0zzozcD2vOZg" /></figure><p><strong>Scalability and maintenance</strong></p><p>Adopting this workflow as the standard approach for Adevinta would see us managing and maintaining a growing number of locally executed scripts and Excel files. We would also have to delegate manual process data governance, which besides being dangerously prone to errors, is increasingly complex and time-consuming.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*BA6uz1BR4h5EFop3" /></figure><p><strong>Governance and compliance</strong></p><p>Relying on local processes and data makes it impossible to implement computational governance. You simply cannot define policies and processes in an automated way to ensure the company’s regulations are being fulfilled without a manual process involved. Which also means that you cannot liberate the user from heavy and repetitive tasks such as asking for permissions to read input data, or to write data in the proper output path.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*rTQLLuB1xhKHL0IO" /></figure><p><strong>Accuracy</strong></p><p>As stated in the security section, sending a file by mail is prone to errors. There is no way, other than manually checking, to ensure the data sent is the most recent, or to ensure the addressee is reading the proper version, especially considering the typical look of our Downloads folder.</p><p>Accidental or even intentional modifications to the data can occur, undermining accuracy. In scenarios where multiple individuals or teams need to review the same data, this can result in varying interpretations, potentially leading to users working with slightly different versions of the data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*WKC2L4Vim6-CReQ-" /></figure><p><strong>Data reliability</strong></p><p>Similar to the susceptibility of the output data, input data presents risks to its consistency. When working in a local environment, data can be modified in error, rendering the whole process invalid from the very beginning. Even if the data is not manipulated at all, there are no guarantees that the input is clean enough for downstream processes and models to run smoothly or to give a trustworthy result.</p><h3>What we wanted to achieve</h3><p>Our vision for the Data Analyst workflow revolves around <strong>providing tools and defining processes</strong> so Data Analysts can work in an improved way. We wanted not only to avoid errors or make the daily work easy for the users but also to embed the <strong>company’s culture</strong>, introducing some <strong>engineering mindset</strong> into the flow such as <strong>automation</strong>, testing and scheduling. Ultimately, we wanted to build a workflow that requires as little manual intervention as possible.</p><p>We also wanted to focus on <strong>reliability</strong>, building repeatable, scalable and safe processes through <strong>continuous integration and continuous deployment</strong>. The process should also be governed by introducing<a href="https://martinfowler.com/articles/data-mesh-principles.html#FederatedComputationalGovernance"> computational governance</a> practices, i.e. <strong>governance by design</strong>. These are automated processes and flows that allow governance without human interaction, and are the only way of addressing many issues such as security, privacy or data cataloguing before any breach, problem or even redundancy can happen.</p><p>We need to establish a process so the users can work seamlessly in an efficient manner, reducing risks of <strong>redundancy</strong> and <strong>data inconsistency</strong>.</p><p>So, to summarise, we wanted to go from using scripts and Excel to a <strong>Data Product</strong>. This is why we called this part of our platform, The Data Product Builder.</p><p>To achieve this goal, we focused on the following specific points:</p><ul><li>Script versioned in a GitHub repository managed by the Data Analysts team.</li><li>Scheduling through an Airflow DAG.</li><li>Input data must be stored in Adevinta Spain’s data lake to ensure we maintain our data contract guarantees.</li><li>Output data also needs to be stored in Adevinta Spain’s data lake.</li><li>Both input and output data (like all data in the data lake) must be accessed through governed processes and tools.</li><li>Integrate CI/CD practices, creating an automated and testable flow that can be evaluated in a test environment before being promoted to production.</li><li>Provide a test environment as easy to work with as the local environment, so both script and input data can be tested quickly.</li><li>Provide a way so the process can be promoted to production seamlessly.</li></ul><p>If we want to succeed<strong>, we should provide tools with computational governance embedded</strong>. Ensuring governance practices are introduced from the very beginning in an automated way, without requiring the user to perform manual or costly processes, requests, jiras etc.</p><h3>Challenges</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/200/0*5_jVllzx87-I60aj" /></figure><p><strong>Culture change</strong></p><p>The primary challenge to change lies in navigating and even pushing a cultural shift. Currently, the existing processes are effective for both users and the company. Users can work and fulfil their responsibilities within the company, which is operating successfully. In light of this, a question arises: Why change?</p><p>Presenting the need for change could potentially be interpreted as a critique, leading users to resist the intended change. So first of all, we need to work hand-in-hand with our users.</p><p><strong>Product mindset: it’s not about technology, it’s about a common goal</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*KjTP3r693r0bUFrh" /></figure><p><strong>Code migration</strong></p><p>As part of the transition from the local environment to an automated setup in the Adevinta Spain Data Platform, which uses Databricks, certain code adjustments will be necessary. In the current scenario, users read and write data directly using Python. However, in a Databricks environment, this process is more efficient if executed through Spark.</p><p>While the changes required are not big and affect only the input/output sections, they must be addressed. This implies that the migration will not be as seamless as we wanted.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*QImeoX0EuESq5S6F" /></figure><p><strong>Data migration</strong></p><p>Considering that the entire process is set to operate a production environment with automatic scheduling and minimal human interaction, it’s imperative that both input and output datasets reside in our data lake. That means that the processes that create input data need to be planned, built and scheduled.</p><p>This is the only way to ensure data integrity from the very beginning. Unlike a straightforward script migration, the whole process needs to be migrated, thus Data Engineers will play a crucial role in assisting analysts/scientists to architect and implement pipelines that feed input data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*UpShdAH0Iuy4WAr-" /></figure><p><strong>Distributed vs non-distributed</strong></p><p>Another issue has emerged as we address the challenges and complexities of such a cultural change and process enhancements. While using Spark to read data enables users to handle vast amounts of data, the specific process targeted in this project typically involves working with smaller input datasets.</p><p>On one hand, users possess the capability to read a substantial table as input, applying filters to load only the smaller subset relevant to their processing requirements. This flexibility proves advantageous, especially given the data storage context in the data lake as discussed earlier. On the other hand, the challenge lies in the misalignment between the typical user-profiles and the general usage of Spark. The average user profile does not align with the use of Spark, therefore, as a team, we need to shift our focus towards training and instilling confidence in users, enabling them to use Spark effectively, and overcoming any apprehensions or challenges they may encounter.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/128/0*Kdc_L0_zOPv4Pvio" /></figure><p><strong>Developing</strong></p><p>The users of this workflow are accustomed to swiftly developing and testing in their local environments, both in the sense of the script code itself and in cleaning or fixing input data. To successfully integrate them into our Data Product Builder Platform it is imperative that we provide users with equivalent capabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*rp8hcHEcbp7gwGLo" /></figure><p><strong>People involved</strong></p><p>Finally, it’s crucial to consider the diverse needs of all individuals and teams involved. When working to build a tool or framework, we tend to focus on the developer, in this case, the Analysts that develop the script, who are the main users. We also talked about Data Engineers, who will play a pivotal role in constructing data pipelines to schedule and produce the input data needed.</p><p>We also need to consider a broader spectrum of users, including those responsible for analysing and processing the output data. Their workflows are also bound to undergo significant changes, moving from receiving data via mail to reading it directly from the Adevinta Spain Data Platform. Recognising and accommodating the varied requirements of all user roles is essential for successfully adapting the entire workflow.</p><p>Learn how we tackle these challenges in part two — coming soon!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=edaec9228753" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/how-we-moved-from-local-scripts-and-spreadsheets-shared-by-email-to-data-products-edaec9228753">How we moved from local scripts and spreadsheets shared by email to Data Products</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Analyst Enablers Community: 
A co-leadership journey]]></title>
            <link>https://medium.com/adevinta-tech-blog/data-analyst-enablers-community-a-co-leadership-journey-91186a584e1b?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/91186a584e1b</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[collaboration]]></category>
            <category><![CDATA[organizational-culture]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 27 Feb 2024 09:21:48 GMT</pubDate>
            <atom:updated>2024-02-27T09:21:48.932Z</atom:updated>
            <content:encoded><![CDATA[<h3>Data Analyst Enablers Community: <br>A co-leadership journey</h3><h4>How we broke silos to maximise data analytics potential and drive organisational success</h4><p>By: <a href="https://www.linkedin.com/in/aingerudf/">Aingeru Duarte</a>, <a href="https://www.linkedin.com/in/javierroldan/">Javier Roldan</a>, <a href="https://www.linkedin.com/in/oriol-vila-masip/">Oriol Vila</a> and <a href="https://www.linkedin.com/in/robertamanuelacoda/">Roberta Manuela Coda</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QSrGnCOhYPZAdXpfF6cToQ.png" /></figure><p>In 2023, <a href="https://adevinta.es/">Adevinta Spain</a> embarked on a transformative journey to enhance collaboration and standardise practices across our diverse team of over 40 Data Analysts. This initiative was driven by the growing size and complexity of the data area, leading to the formation of the Data Analytics (DA) Enablers Community.</p><p>As co-leads we would like to share all the learnings, as well as the value and impact on<a href="https://adevinta.com/about-adevinta"> Adevinta</a>, we achieved thanks to the work carried out by the DA Enablers Community during 2023.</p><h3>How it all started</h3><p>At Adevinta Spain we had a complex personnel structure with Data Analysts spread across different teams and marketplaces.</p><p>The day-to-day life of a Data Analyst in Adevinta Spain is based on several fields of expertise. Among these, we identified four that needed to be prioritised:</p><ul><li>Digital analytics: tagging, behavioural analytics etc.</li><li>Experimentation: assessing new features through controlled online experiments, often called A/B tests</li><li>Data Visualisation: foster a common culture, through collaboration, best practices and agreements around Tableau Server governance, for improved performance, security and end-user experience</li><li>Global analytics: here we have worked on initiatives across the analytics domain, for example, Data Products development for Analysts, definition of technical data debt etc.</li></ul><p>For each of these fields, there is a lead whose mission is to help the analysts maximise delivered value.</p><p>Previously, each lead organised weekly meetings with representatives of the different Data Analytics teams to boost initiatives meant to reduce analysts’ pains and explore new opportunities. However, without an official time allocation, the representative found it almost impossible to work on those initiatives. As the analyst community grew in number, field leads realised it was increasingly difficult to scale to drive good practices, standardise ways of working and take advantage of our scale.</p><p>Nonetheless, the entire organisation recognised there was value in the initiatives that emerged from those spaces.</p><p>Working with the Data Directors, we looked at how other areas within the Company addressed similar situations. We found that areas such as Data Engineering, SW Development and others, relied on profiles called “Enablers.” These profiles have an official nature, with dedicated time allocation to promote initiatives that improve their discipline — precisely what we were looking for. This was when the seed of the DA Enablers Community was planted.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mRWKWZrC8Dw0egUa" /><figcaption>Community meeting</figcaption></figure><h3>Finding our way through the challenges</h3><p>As Discipline Leads, we identified three crucial areas requiring attention to ensure the DA Enablers Community’s success:</p><ol><li><strong>Effective Communication: </strong>The diverse backgrounds and work practices of the enablers initially created communication barriers. To overcome this, we implemented a feedback-driven approach, utilising insights from quarterly surveys and monthly retrospectives. This iterative process enabled us to develop a customised Agile methodology that resonated with our unique community, ultimately promoting more efficient communication. Our refinement went beyond just session structuring; it included fine-tuning the tone and questions used in interactions with enablers, streamlining follow-up processes, and clarifying ownership responsibilities for each task.</li><li><strong>Sponsorship and Commitment:</strong> Ensuring the sustained engagement and commitment of the enablers was paramount. To achieve this, we established a clear agreement of sponsorship. Becoming an enabler meant a commitment to dedicate up to 20% of their work time, typically one full day per week, to community-related activities. This agreement, endorsed by managers, provided the enablers with the necessary time and resources to contribute effectively, fostering a culture of recognition and value for their efforts in strengthening the community.</li><li><strong>Prioritisation and Time Allocation: </strong>While the DA Enablers Community’s goal was to prioritise tasks from each discipline, we quickly realised that the community should have a single backlog. Maintaining a backlog for each field did not solve the issue of where to allocate our efforts for the next sprint and proved to be less effective. Additionally, we transitioned towards a more enabler-centric approach, giving more weight to the enablers and the co-leads by adopting a Servant Leadership approach. This shift meant that enablers were now actively involved in deciding which tasks took precedence, significantly enhancing their autonomy and sense of responsibility. Our role as co-leads evolved to focus more on facilitating these prioritisation sessions, guiding the enablers while respecting their insights and decisions. This approach not only streamlined our processes but also instilled a deeper sense of ownership and commitment among the enablers, leading to more engaged participation and higher-quality outcomes.</li></ol><h3>Learnings as co-leads</h3><p>Building a community from scratch has been a challenging, yet enormously enriching, journey for us as co-leads:</p><p>First, we learnt how to leverage our different backgrounds to get to know each other and build trust. This process was then repeated within the entire group as a necessary prerequisite to create a foundation for a unified approach to our Community’s goals. Our shared understanding fostered collaboration, highlighting the importance of interpersonal connections as a catalyst for effective leadership.</p><p>Second, we needed to navigate organisational challenges. We understood the need to seek guidance from the DA Distributed Enablers who had a greater Agile methodology experience.. We also realised the importance of Servant Leadership to leverage the strengths of the entire team.</p><p>Last but not least, despite moments of doubt, our manager offered unwavering support, as the owner and sponsor of the initiative. She kept us on course, strengthened our trust in the Community and encouraged us to keep iterating for continuous improvement.</p><p>As co-leads, we’ve experienced the transformative power of trust, Servant Leadership and the importance of a strong sponsorship.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*llzkKGM-dgp7thBk" /><figcaption>Community members having a chat on our office terrace in Barcelona</figcaption></figure><h3>Delivering value to the organisation</h3><p>In just one year, thanks to this Community, we have achieved notable achievements that have strengthened our Data Strategy and accelerated its implementation across all verticals and areas. It has helped break the silos and build a common ground for the entire Adevinta Analytics Community.</p><p>The initiatives carried out have raised the bar of the Data Analytics discipline within the company, in particular:</p><p><strong>Data Debt Management:</strong> We have managed to elevate Data Debt to the category of Technical Debt, thus guaranteeing that data-related issues have their assigned space in the teams’ backlog, similar to Technical Debt.</p><p><strong>Best Practices and Agreements</strong>: In the area of best practices, specific guidelines have been implemented and disseminated in several key areas. Agreements have also been stipulated for their implementation, in particular, for Product Development Experimentation as well as Tableau Server enablement and governance.</p><p><strong>Improved Communication and Shared Culture:</strong> We have designed a structure that takes advantage of the scale of our organisation, establishing a channel that effectively conveys the Data Analysts’ feedback regarding emerging challenges. Thanks to the DA Distributed Enablers bridging their teams’ needs with all leads and vendors responsible for the Data Strategy implementation, we ensure communication flows in both directions.</p><ul><li>Communicating and making shared decisions about crucial common issues, such as including changes to cookie consent and violations in the use of events in our tagging plan.</li><li>Coordinating common actions, like the Tableau Server pre and post-upgrade testing and related upskilling training.</li><li>Identifying and resolving gaps in the analytics area through bottom-up prioritisation. For example, the lack of alignment between the Data Strategy and its implementation, how to ensure data analysts get upskilled to gain increased autonomy in building data pipelines or how data analysts get involved in the strategic decisions that affect their work.</li></ul><h3>Final thoughts</h3><p>Looking back, the journey to build the community has been beyond a doubt worthwhile so far. It has been a treasure trove of learning about leadership, collaboration and resilience. And it has also opened the path to a more efficient and scalable approach for Data Analytics in a fast-paced and constantly evolving environment. As Adevinta is headed towards a global verticalisation, we are looking forward to the new opportunities to come in 2024 for the DA Enablers Community to keep growing and bringing value to the company, possibly on a larger global scale.</p><p><em>Acknowledgements:</em></p><p>Thanks to our manager and Data Director, <a href="https://www.linkedin.com/in/lauralara/">Laura Lara</a> for the continued support on this journey, to <a href="https://www.linkedin.com/in/enric-martinez-ibarra/">Enric Martínez Ibarra</a>, Data Engineering Enabler Lead, for sharing his experience with us and for the invaluable help as representative of the CompaaS team, to the ex-members of the community <a href="https://www.linkedin.com/in/alexfrauca/">Alex Frauca</a>, <a href="https://www.linkedin.com/in/olha-vons/">Olga Vons</a>, <a href="https://www.linkedin.com/in/polhernandezmolin/">Pol Hernandez</a>, for being there in the early stages. Last but not least, to all DA Distributed enablers, <a href="https://www.linkedin.com/in/bego%C3%B1a-menchon-01365853/">Begoña Menchon</a>, <a href="https://www.linkedin.com/in/mari-sanchez-garcia-7468138/">Mari Sanchez García</a>,<a href="https://www.linkedin.com/in/idoiamartialuja/">Idoia Martí i Aluja</a>, <a href="https://www.linkedin.com/in/mireiagarciasorribes/">Mireia Garcia Sorribes</a>, <a href="https://www.linkedin.com/in/olayafernandezmartin/">Olaya Fernández Martín</a>, <a href="https://www.linkedin.com/in/roger-ten-valls/">Roger Ten</a>, <a href="https://www.linkedin.com/in/seraf%C3%ADn-cruz-garrido-7a831917a/">Serafín Cruz Garrido</a> for believing in this community and their amazing work to make it a reality!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=91186a584e1b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/data-analyst-enablers-community-a-co-leadership-journey-91186a584e1b">Data Analyst Enablers Community: 
A co-leadership journey</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Glitches to Grins: The Support Superhero Squad’s Epic Journey to SLO Success!]]></title>
            <link>https://medium.com/adevinta-tech-blog/from-glitches-to-grins-the-support-superhero-squads-epic-journey-to-slo-success-df1ccce9dda1?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/df1ccce9dda1</guid>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Wed, 13 Dec 2023 14:48:51 GMT</pubDate>
            <atom:updated>2024-01-16T09:39:19.705Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Overcoming Hurdles: The Support Superhero Squad’s Quest for SLO Triumph</em></p><p>By Tulio Cruz and Berta Amat, Adevinta</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DLXsBbiSgN8PQuLugeQXRQ.png" /></figure><p>This article delves into the critical significance of Service Level Objectives (SLOs) within <a href="https://adevinta.com/">Adevinta’s</a> operational framework. As organisations increasingly rely on seamless digital services, SLOs emerge as a pivotal component, ensuring optimal performance and reliability. Berta Amat, Engineering Manager in Adevinta’s Data Catalogue team, along with <a href="https://medium.com/@tuliolucascruz">Tulio Cruz</a>, Tech Principal at <a href="https://www.thoughtworks.com/">Thoughtworks</a>, working closely with Adevinta’s Spanish marketplace — Infojobs, offer insights into why SLOs hold substantial relevance for Adevinta’s operational efficiency and technical advancement.</p><p>Welcome! Have you ever wondered about the buzz surrounding SLOs, error budgets, and SLIs? It’s not just another temporary tech trend, and we’re here to shed some light on why they matter — and why they are here to stay.</p><p>The concept of Site Reliability Engineering (SRE) and Service Level Objectives (SLOs) has its roots in Google’s engineering teams back in the early 2000s. They were faced with the monumental task of ensuring the reliability of their large-scale, distributed system (Google Site Reliability Engineering Team, n.d.).</p><p>Now that we’ve set the scene, feel free to get comfortable, perhaps with a cup of tea, as we share this story with you…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pX08Q7wHR3PpSADU" /></figure><p><em>Credits: Malu. “Portfolio.” Malu.dev, </em><a href="https://malu.dev./"><em>https://malu.dev.</em></a></p><h3><strong>Once Upon a Time: When Heroes Meet Chaos</strong></h3><p>In the fast-paced world of tech, ACME, a company building user-facing services, embarked on a tumultuous journey. At the crux of their evolution, they transitioned to a distributed system while riding the waves of exponential user growth.</p><p>The teams were diligently crafting fresh features to both retain and entice new users. Each team expedited the implementation of these features. However, the rapid pace of development became a problem, as some of these changes often led to broken functionalities or misbehaving areas of the system that affected services, creating a disconcerting user experience.</p><p>ACME encountered a significant hurdle in the form of a lack of consensus regarding the definitions of availability and reliability within the organisation. Each team determined the availability and reliability of their services autonomously. There was no consistency between the different services, and with no prior deliberation or consensus at either the team or company level.</p><p>This situation led to negative consequences.</p><p>The sheer volume of metrics being collected became overwhelming, with many alerts proving irrelevant or non-actionable. Swamped by the constant barrage of alerts, team members began to ignore them, resulting in chaos when real issues arose. Sometimes team members had to wake up in the middle of the night and assess whether the alert was actionable or not. If it was actionable, the stressful path to find out the problem started…</p><p>Team morale plummeted, and relationships between team members and other teams deteriorated. Trust levels hit an all-time low.</p><p>Internal conflicts within the organisation had a significant impact on the platform’s users. They experienced a noticeable decline in the quality of the platform’s functionalities, with increased downtime and latency, ultimately resulting in user dissatisfaction.</p><p>Unfortunately, the Customer Satisfaction team struggled to identify the root causes of the issues being reported. Consequently, the platform witnessed a decline in its active user base.</p><p>Following thorough deliberations, ACME’s team members resolved to address the predicament through the lean inception of the ‘SLO initiative’.ACME initially aimed for 100% reliability. However, they soon realised the complexity and cost of this approach outweighed its benefits for users. It became evident that better alternatives existed.</p><p>After exhausting various strategies on their own, including creating a multitude of monitors and alarms, they came to the realisation that external assistance was needed. That’s when they discovered ‘The Support Superhero Squad’. This dedicated team swoops in to aid their clients whenever challenges arise. With each member possessing unique strengths and abilities, they stand ready to tackle any situation that comes their way.</p><h3><strong>Let’s meet the Superheroes: Exploring The Support Superhero Squad</strong></h3><h3><strong>Metric Maven (SLI — The Service Level Indicator)</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IBVpFReKffrThgam" /></figure><p><em>Credits: Malu. “Portfolio.” Malu.dev, </em><a href="https://malu.dev./"><em>https://malu.dev.</em></a></p><p>Meet Metric Maven, the embodiment of dynamism armed with an unparalleled ability to intricately measure every dimension of service performance. This persona remarkably parallels the criticality of Service Level Indicators (SLIs) within the landscape of SLO methodology. Much like guiding beacons, SLIs play an illuminating role, shedding light on the effectiveness and dependability of a service. These indicators meticulously assess pivotal metrics such as response times, uptime percentages, and error rates, thereby providing invaluable insights indispensable for refining and optimising performance</p><h3><strong>Sentinel Speedster (SLO — The Service Level Objective)</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*766XcBJXJa0lxEk5" /></figure><p><em>Credits: Malu. “Portfolio.” Malu.dev, </em><a href="https://malu.dev./"><em>https://malu.dev.</em></a></p><p>In the bustling metropolis of ACME, where the pursuit of excellence is akin to a finely tuned symphony, the Sentinel Speedster emerges as the living, breathing embodiment of Service Level Objectives (SLOs). Picture this swift and impeccably precise superhero as the guiding North Star, illuminating the path towards unrivaled service excellence that your company aspires to achieve. SLOs, much like the conductor of an orchestra, set the pace and precision at which your services should harmonise, ensuring customers revel in nothing short of a truly extraordinary experience. With the Sentinel Speedster leading the charge, ACME is not merely raising the bar — it’s crafting a magnum opus of exceptional service!</p><h3><strong>Error Buster</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*v8Ry0YNeMDfgeN-3" /></figure><p><em>Credits: Malu. “Portfolio.” Malu.dev, </em><a href="https://malu.dev./"><em>https://malu.dev.</em></a></p><p>As the self-styled guardian of the error budget, Error Buster stands steadfast, ready to shield your services from unforeseen challenges. Much like this vigilant defender, the error budget provides a safety net. It allows for a margin of imperfection, ensuring that your services can operate within defined parameters while still meeting the prescribed SLO.</p><h3><strong>Alarm Oracle (the Alarms)</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*X_InFx4eW_QxGo_V" /></figure><p><em>Credits: Malu. “Portfolio.” Malu.dev, </em><a href="https://malu.dev./"><em>https://malu.dev.</em></a></p><p>In the ever-evolving landscape of SLOs, Alarm Oracle emerges as a seer, sensing trouble before it manifests. This astute hero operates much like alarms in the world of SLOs. Picture them as vigilant sentinels, alerting your team at the first sign of deviation in service performance. These early warnings enable timely intervention, preempting potential issues before they escalate.</p><h3><strong>Collaboration Crusader (Cross-Team Collaboration)</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kAn_xHQGKc1paJT9" /></figure><p><em>Credits: Malu. “Portfolio.” Malu.dev, </em><a href="https://malu.dev./"><em>https://malu.dev.</em></a></p><p>In the tapestry of SLO methodology, Collaboration Crusader takes centre stage, emphasising the significance of cross-team collaboration. Like a unifying force, Collaboration Crusader rallies teams together. SLOs thrive on seamless cooperation between diverse departments — support, engineering, product teams, all working in harmonious synchrony. This collective effort ensures that everyone marches in unison towards achieving the defined SLO.</p><h3><strong>The Support Superhero Squad in action</strong></h3><p>Initially, the Collaboration Crusader took the stage, employing their exceptional interpersonal skills to unite key stakeholders and capture the attention of all involved teams. Together, they worked to define the availability within the context of their service and determine the necessary level of reliability.</p><p>Following this, Metric Maven, armed with their extraordinary ability to measure a wide range of metrics, contributed to defining key performance indicators. These included metrics like request latency, batch throughput, and failures per request, all essential for achieving the previously established goals.</p><p>Subsequently, with their analytical expertise and holistic approach, Sentinel Speedster guided ACME’s stakeholders in defining SLOs based on the previously established definitions. This involved agreeing on a service’s minimum reliability before users would notice any issues. They also discussed the margin for errors and the room for introducing new functionalities without hindering the overall process, thereby defining the acceptable risk of the system.</p><p>The Sentinel Speedster also shared an important concept during these discussions. ACME initially aimed for 100% availability, but the Sentinel Speedster highlighted the drawbacks of this approach, including its high cost and technical complexity. Ultimately, they worked together to find a more suitable solution for the agreed objective.</p><p>Next in line was Error Buster, armed with rational, pragmatic, and confident capabilities. Working in tandem with other superheroes, Error Buster tackled the task of determining the error budget — essentially, the permissible level of unavailability tolerated for applications before intervention became necessary, based on the defined SLOs.</p><p>Error Buster collaborated with various stakeholders, and the Product team played a significant role in these discussions. The discussions primarily focused on determining the error budget, taking into consideration the importance of the previously established acceptable risk and SLOs in this context. For example, if an ACME team wanted to deliver many risky features, they had to configure a loser SLO compared to a team that wanted to release fewer risky features.</p><p>Furthermore, Error Buster, in tandem with the Collaboration Crusader, communicated to ACME the potential hurdles teams might encounter when surpassing the error budget. The paramount repercussion is that, should the Error Budget be exceeded, we’ll be compelled to postpone a release until we bolster the system’s reliability. Therefore, it’s imperative to bear in mind that when we’re inclined to take more substantial risks, we may require more adaptable SLOs to avert swift overages.</p><p>The predefined metrics and monitoring will assist the company in tracking our proximity to exceeding the error budget. This warns us if our SLIs are failing a lot and may exceed that target.</p><p>However, a pressing question arises: can we afford to overlook an exceeded error budget? The heroes proposed the idea of establishing an exception system that permits teams to release under exceptional circumstances a limited number of times per year. These two heroes helped ACME establish a contingency plan should this error budget be exceeded.</p><p>With all our superheroes operating at peak efficiency, having meticulously defined the SLOs, SLIs, and Error budget, it was Alarm Oracle’s turn to step in. Armed with their unique ability to anticipate trouble before it arises, Alarm Oracle set up a comprehensive alarm system, ensuring that every critical aspect was monitored vigilantly, all while respecting the error budget parameters defined by Error Buster.</p><p>The Alarm Oracle concentrated on creating metric dashboards and alarms in collaboration with stakeholders, with a specific focus on actionable alerts to prevent unnecessary noise or alert fatigue.</p><h3><strong>Conclusion</strong></h3><p>Ultimately, the combined efforts of these extraordinary heroes, each embodying a crucial facet of SLO methodology, paved the way for ACME’s success. The Collaboration Crusader fostered unity, Sentinel Speedster provided clarity, Metric Maven ensured precision, Error Buster fortified resilience, and Alarm Oracle stood vigilant. Together, they defined SLOs, SLIs, and error budgets, fostering a culture of reliability and innovation within ACME.</p><p>ACME team members quickly witnessed the profound impact of these changes and the effectiveness of SLOs. The availability and reliability of the services they provided showed a remarkable exponential improvement. When alerts were triggered, they were all actionable, enabling our teams to readily pinpoint and address issues. All these changes lead to a boost in team morale. These transformations didn’t go unnoticed by the Customer Support team either; they observed a reduction in user issues, resulting in an upsurge in user satisfaction and an increase in the user base.</p><p>With all the efforts and all the SLO methodology applied, the ACME could see that they had greater control of their entire ecosystem. They could pay attention to user responses, always knowing what to expect from their applications. And because they could focus much more on improving their business, they saw revenues increase in a more precise and consistent way.</p><p>As the journey continues, the lessons learned from the ‘The Support Superhero Squad’ will continue guiding ACME toward a future where excellence is not just a goal but a way of life. It’s a testament to the power of SLOs in transforming not just services but entire organisations into beacons of reliability in the ever-evolving landscape of technology.</p><p>Zero to Hero, Error to Zero! Support Squad, Let’s Roll!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NsvYgjL9j0nEeKK3" /></figure><p><em>Credits: Malu. “Portfolio.” Malu.dev, </em><a href="https://malu.dev./"><em>https://malu.dev.</em></a></p><p><strong>References</strong></p><p>Google Site Reliability Engineering Team. (n.d.). “Service Level Objectives.” In the SRE Book. Retrieved from <a href="https://sre.google/sre-book/service-level-objectives">https://sre.google/sre-book/service-level-objectives</a> (Accessed on October 25, 2023).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df1ccce9dda1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/from-glitches-to-grins-the-support-superhero-squads-epic-journey-to-slo-success-df1ccce9dda1">From Glitches to Grins: The Support Superhero Squad’s Epic Journey to SLO Success!</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lessons learned from organising our first global AI hackathon]]></title>
            <link>https://medium.com/adevinta-tech-blog/lessons-learned-from-organising-our-first-global-ai-hackathon-c19cbc86e449?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/c19cbc86e449</guid>
            <category><![CDATA[collaboration]]></category>
            <category><![CDATA[hackathon-organizing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[hackathons]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 19 Sep 2023 08:32:21 GMT</pubDate>
            <atom:updated>2023-09-19T08:32:21.323Z</atom:updated>
            <content:encoded><![CDATA[<p>Learn from our experience and best practices when organising your own hackathon</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/984/1*4KGM9rGE563j6BiUeoKHJg.png" /></figure><p>In June 2023, a first global hackathon was organised <a href="https://www.adevinta.com/">at Adevinta</a>, bringing together people from <a href="https://www.adevinta.com/our-brands">our marketplaces</a> <a href="https://www.leboncoin.fr/">Leboncoin</a>, <a href="https://www.kleinanzeigen.de/">Kleinanzeigen</a>, <a href="https://www.milanuncios.com/">Milanuncios</a> and Global teams. The goal was twofold: To bring together people from different sites and countries and have fun building something useful for the company using generative artificial intelligence (AI).</p><p>So how did the hackathon turn out? In a nutshell, pretty well! Overall, participants’ feedback was positive, both during the event and in the satisfaction survey sent afterwards. In addition to meeting colleagues from other <a href="https://www.adevinta.com/careers/our-locations">Hubs</a>, they were able to learn about generative AI tooling and its stack.</p><p>But now that some time has passed, we’d like to take a step back to identify what went well and what could have been done differently. There is always room for improvement in every project, especially when we only had a month and a half to prepare for it, and none of us were experts in this kind of event planning.</p><p>But before we examine the lessons learned from this experience, let’s give you some more context.</p><h3>Setting the scene</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D_jXqk_1wNs2u4ctLl15UA.jpeg" /><figcaption>The setup of the hackathon in our Paris office</figcaption></figure><h4>A few stats</h4><p>This Adevinta AI hackathon was held over two days in Paris, with physical attendance required. There were 110 participants, plus 30 organisers, external partners, jurors and stakeholders. In total four countries were represented: France, Germany, Spain and the Netherlands.</p><p>Note: The hackathon was only open to those Adevintans working on a specific internal transformation project, not everyone employed by the company.</p><h4>Teams and topics</h4><p>The participants were free to propose any AI topic they wanted. After the kick-off meeting one month before the event, a <a href="https://miro.com/">Miro</a> board was created for people to post ideas. Then individuals organised themselves into teams based on the topic they found the most interesting.</p><p>There were some criteria set for team formation:</p><ul><li>It was recommended that teams not exceed six members.</li><li>A minimum of two countries/marketplaces needed to be represented in each team.</li><li>The aim was for each team to have a mix of Product Management, Engineering, Data/machine learning (ML) and UX/UI people.</li></ul><p>The recommendation regarding the maximum number of people per team was not always followed, resulting in only 12 being created out of the 110 participants.</p><h4>AI platforms</h4><p>The generative AI platforms used were <a href="https://huggingface.co/">Hugging Face</a>, <a href="https://aws.amazon.com/?nc1=h_ls">AWS</a> and <a href="https://openai.com/">OpenAI</a>. Hugging Face and AWS representatives were present during the whole hackathon to assist teams. A dedicated table was set up, but they also circulated and directly checked in with teams at their tables.</p><h4>Pitches</h4><p>We chose to make the pitch sessions really short: four minutes and no more. Each team had to use the same slide deck. They were also offered help with practicing their pitches.</p><h4>Votes</h4><p>There were two votes to reward the best projects: a jury vote and a popular vote.</p><p>A total of six people were involved with the jury vote, using four criteria for their decision-making:</p><ul><li>Impact (30%): Financial impact or indirect/strategic impact.</li><li>Innovation (30%): Is the idea unique?</li><li>Feasibility (30%): Does the hack have a realistic chance of succeeding with the time, skills and technology we have?</li><li>Collaboration (10%): Best group picture, with bonus points for teams representing more than two marketplaces.</li></ul><p>As with the teams, the jury was composed of members from multiple marketplaces and countries, with a variety of roles represented (CTO, CPO, Data/ML, and Product Management). Grades were given for each category using a scale of 1 to 5. The results for each criterion were then added up.</p><p>The popular vote was conducted by the participants themselves.</p><h4>Prizes</h4><p>The winners of the jury vote received smart watches, invitations to AI conferences and the opportunity to pitch to high-level management.</p><p>Snack boxes were delivered at home or the office for the team that won the popular vote.</p><p>With all the details about the hackathon now covered, let’s take a look at the lessons learned.</p><h3>Key lessons learned: What worked well</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uYO7F137euU1qVIRSLearw.jpeg" /><figcaption>Participants working together</figcaption></figure><h4>Building a team you can trust</h4><p>Ensuring you surround yourself with motivated and aligned team members makes a huge difference! Our project manager was assisted by one person on site who was responsible for logistics, one for communication and design and one for AI expertise. Two of us had already organised hackathons, but none were experts. We communicated weekly in quick synchronisation meetings, which became twice-weekly as the deadline approached.</p><p>We received the full confidence of senior management to organise this event successfully. They did not require regular updates. Even though the initial request came from the top, we took this project on and made it our own, and we could organise things however we wanted.</p><h4>Picking a quick-to-build area</h4><p>Hackathons usually last a few days, so participants should have access to easy-to-use technologies to be able to build their demos as quickly as possible.</p><p>AI is a good example of an area that has become more accessible. Generative AI is much easier to use than the previous ML generation as models don’t need to be trained (which took ages). Using foundation models makes AI more accessible for participants, since the main concern is integrating it, not building models.</p><p>We were happily surprised to see how many teams were able to have a demo working by the end of the 2 days.</p><h4>Providing the same experience for everyone</h4><p>We were convinced from the start that every participant should have the same experience. This meant that no remote work was allowed: Everyone had to be physically present at the Leboncoin offices in Paris. Moreover, all the participants worked in the same room rather than being dispatched throughout the building.</p><p>The external partners followed the same logic. Hugging Face and AWS representatives were present in person because we felt the participants would not be comfortable/confident enough to contact them by Slack or email. As more than 50% of respondents said they were “very satisfied” with the help of the external partners, we know now it was the right decision.</p><p>In addition to providing a sense of unity, having everyone in one location allowed us to provide massages, showers, nap locations, catering services and even a marching band. This also made it easier to organise an icebreaker before hacking sessions: The participants were split into groups of two teams, with each being given a card with pictures of different items on and instructions to act out the items, charades-style, to the other team. Also, in the evening of the first day of the hackathon, a dinner was arranged so that the participants could enjoy a meal together.</p><p>We’re not saying that online hackathons don’t work, but from our experience we recommend saving them for groups of people who already know each other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*09T4DOyTg0Sb9UkWwwHfuQ.jpeg" /><figcaption>Participants discussing their project</figcaption></figure><h4>Putting learning first</h4><p>The more people can learn from the experience, the more willing they are to give their time. That’s why we provided assistance for participants to learn as much as possible about generative AI.</p><p>Special sessions were held during working hours two weeks before the event to allow people to ask questions. Additionally, we helped them set up their AWS, Hugging Face and OpenAI accounts.</p><p>An introductory talk about AI was given the first morning of the hackathon, and throughout the event a total of about 15 people from AWS and Hugging Face attended at various times to provide assistance.</p><p>We recommend choosing tools that participants already know or are used within the company to avoid having to spend too much time on the learning curve.</p><h4>Giving participants the freedom they need</h4><p>By definition, hackathons are about breaking the rules and doing things differently. So even if you define some ground rules, don’t be too strict about them. Also, don’t worry too much if things don’t go as planned and let people self-organise as much as possible.</p><p>During this hackathon, people were free to use any AI tool they wanted, even if AWS, Hugging Face and OpenAI accounts had been set up.</p><p>We recommend applying the same approach to choosing topics. It’s important to let people express their opinions even if the topic isn’t 100% what you had envisaged being part of the session. Because hackathons are untimed momentum sessions meant for exploration, you also shouldn’t be too concerned about implementing a particular hack. In the same way, we don’t recommend defining specific use cases for people to follow. Overall, don’t expect hackathons to answer management’s questions.</p><p>In addition, be flexible about the hours: Don’t force participants to start at 9AM if they prefer working late.</p><h4>Backing up logistics as much as possible</h4><p>Let’s be honest, our biggest fear was that the Wi-Fi would go down on D-Day and ruin the event. In order to prevent this from happening, we worked with an external company, who provided an ethernet connection for every table. But they ended up doubling Wi-Fi because not many people used the RJ45 cables.</p><h4>Defining the objective criteria for the votes</h4><p>The best way to avoid bias is to grade the pitches based on criteria that have been defined in advance. Our four criteria (impact, innovation, feasibility and collaboration) worked pretty well. This system prevented good grades being awarded to a team just because they knew a jury member or two. In the event, the majority of the jurors arrived at the same conclusion regarding the winning team.</p><p>So let’s review what could have been done better now that we’ve reviewed the best practices for our hackathon.</p><h3>Things we can improve upon for the next iteration</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rK5XdySC9uwYmpdFAmp5Ag.jpeg" /><figcaption>Surprise band performance with a member of the leboncoin team</figcaption></figure><h4>Giving innovation more room</h4><p>The hackathon was definitely a first step in attempting to generate more innovation around AI within Adevinta. Having the Adevinta CPTO give the introductory speech demonstrated that this was an important event for the company, and soon after it was held, the winning team (based on jury votes) pitched its idea to management. However, no commitments were made that the hackathon ideas would be incorporated into the existing roadmap.</p><p>The success of this first global hackathon has led us to wonder how to integrate a culture of innovation in the company. Right now, Adevinta doesn’t have an R&amp;D department with people working outside the roadmap. It does not mean that teams are not doing R&amp;D: some engineers are using new tools to develop new things, and we are putting together two teams to introduce recent AI innovations to the rest of the teams. However, we still need to find a way to spread AI within teams with senior management working on two-year roadmaps that don’t leave much room for innovation.</p><h4>Focusing even more on the learning experience</h4><p>Some feedback we received from participants was that the hackathon was a bit short on time and so they weren’t able to learn everything they wanted to about AI. Extending the event to three days and adding more breaks could be the solution for a potential next edition. Organising more AI-related initiatives before the hackathon would certainly have been beneficial as well.</p><h4>Welcoming non-tech and non-product people</h4><p>Since our hackathon focused on generative AI, we only invited people with tech and product backgrounds, which included Engineers, Data/ML experts, Product Managers, and UX/UI Designers.</p><p>But we are now wondering whether inviting people from sales and marketing teams would bring more creativity. Innovative ideas sometimes come from people outside the area of expertise.</p><h4>Providing better catering services</h4><p>Food was the subject of most of the negative feedback we received. Participants complained about the lack of inclusivity, such as vegan food not being clearly labeled as such, and the location where food was served, as well as the amount of food served.</p><h4>Working on a better ending</h4><p>Some people expected a better ending after the pitching sessions. Perhaps we could have saved some surprises for the end.</p><h3>Going further</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZSL6CMBxV1RAkX5zTeHIZQ.jpeg" /><figcaption>Picture of the all participants</figcaption></figure><p>In conclusion, Adevinta’s first global AI hackathon was a huge success, bringing together participants from different countries and marketplaces to explore generative AI and build innovative solutions.</p><p>The hackathon demonstrated that AI is one of the biggest technological breakthroughs, which will change how we think about user experience and coding. It is our responsibility as a company to prepare our employees for the future by connecting them with AI experts and training them.</p><p>The biggest challenge we have for the future is to find a way to incorporate innovation into our processes. A hackathon like this proved that disruptive events can speed up innovation within an organisation. The next step is to figure out how to inject an enduring culture of innovation within the company.</p><p>Want to see what it looked like? You can watch the recap video <a href="https://youtu.be/yd5QtlwP0uA?si=givsLfuS0zJgo7xY">here</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c19cbc86e449" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/lessons-learned-from-organising-our-first-global-ai-hackathon-c19cbc86e449">Lessons learned from organising our first global AI hackathon</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How we matured Fisher, our A/B testing Package]]></title>
            <link>https://medium.com/adevinta-tech-blog/how-we-matured-fisher-our-a-b-testing-package-110c99b993fc?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/110c99b993fc</guid>
            <category><![CDATA[experimentation]]></category>
            <category><![CDATA[a-b-testing]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ab-testing]]></category>
            <category><![CDATA[data-analytics]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 22 Aug 2023 08:26:12 GMT</pubDate>
            <atom:updated>2023-09-01T08:23:03.862Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The story of collaboration between teams that led to an advanced mode feature on our global experimentation platform</em></p><p>By <a href="https://www.linkedin.com/in/alejandro-alvarez-p%C3%A9rez-a21912b0/">Alejandro Alvarez Perez</a> and <a href="https://www.linkedin.com/in/aleksandar-jovanovic-ai/">Aleksandar Jovanovic</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rDKxjfKTQoP1CttfwN0uWw.png" /></figure><p>Like an organisation, product experimentation can be seen through the lens of process, culture and tooling. Let’s call these an organisation’s <em>cogwheels of innovation</em>. These are the things that, when working well together, multiply the value from each experiment. In this scenario, if you break one cog you’ll hear screechy, grinding sounds in the machinery.</p><p>In <a href="https://www.adevinta.com/">Adevinta’s</a> Dutch online marketplace, <a href="https://www.adevinta.com/our-brands/marktplaats">Marktplaats</a>, we posed ourselves two questions<em>:</em> <em>In which of these areas are we excelling? And in which less so? </em>We realised that none of the three areas — process, culture, tooling — could be neglected. We really needed to create agile <em>processes</em>, foster an <em>experimentation culture</em> <em>and mindset</em>, while empowering operations with <em>fit-for-purpose tooling</em>.</p><p>We saw a clear starting point, though. Today’s article will take us through the journey of how we matured <em>Fisher</em>, our A/B testing package. In the end, we were able to free up our Data Scientists’ resources, and eventually scale up to a full fledged statistical engine for our global experimentation platform <a href="https://medium.com/adevinta-tech-blog/a-b-testing-to-improve-recommender-products-8b18c77af1c3"><em>Houston</em></a><em>.</em></p><h3>The problem</h3><p>If our Data Scientists were craftsmen, then A/B testing used to be our craft. A/B tests were a gateway to creative exploratory data analysis that we rigorously deployed, aiming to provide our product teams with as much learning as possible. Quite literally, we spent days analysing experiments and crafting insights and reports. Certainly, we were <em>winning,</em> or <em>losing</em>, but we were not doing so very <em>early</em>, nor quickly. The analysis step was long and the Data Scientists were often the bottleneck in the innovation process. Especially, when the number of experiments at any given point was high. All in all, we had a scalability issue, and the main question to be answered was: <em>How can we reduce the time a Data Scientist spends on an experiment?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Tim7uV7ttayWdDDH" /><figcaption>Adevintans collaborating in the Amsterdam office</figcaption></figure><h3>The solution</h3><p>The <a href="https://www.adevinta.com/our-brands/marktplaats">Marktplaats</a> team built <em>Fisher: </em>a python package that enables our Data Scientists to run straightforward <em>hypothesis testing</em> and <em>to produce comprehensive reports</em> with very few lines of code. Fisher supports the entire analysis workflow: from fetching experiment meta(data), to applying standard data cleansing procedures and transformation, to generating the final report. It integrates neatly into our teams’ ways of working by connecting to Jira, Github and different databases, rendering the entire analysis workflow close to fully automated.</p><p>Written in the object-oriented programming paradigm, the package has grown to be an ecosystem and vessel for advanced methods that give our experiments some edge. Over time, we’ve implemented <em>variance reduction</em> methods like CUPED, and corrections to multiple-testing scenarios. Respectively bringing down the time-to-learning by cutting the runtime of our experiments substantially and upholding the quality of the inference. Fisher being the common tool, makes introducing such advances smoother across the board. Enabling a new feature for all Data Scientists is done as fast as pull requests can be made.</p><p><em>A sample code chunk of the analysis step with Fisher:</em></p><pre>from fisher import Analysis, Report, TTest<br><br><br>data = …<br>tests = …<br>analyser = Analysis()<br>results = analyser(tests=tests, method=TTest(), data=data)<br>report = Report.from_jira(&#39;bnl1234&#39;, results=[results])<br>report.save()</pre><p>The look and feel of the main analysis actions in Fisher. The analysis class orchestrates the hypothesis test with a simple t-test on a dataframe characterised by entity, variant and metrics columns. The Tests class represents hypothesis tests that are characterised by contrasts, dependent variables and segments that map the content/columns of the dataframe, as well as other input parameters for the error-types alpha and beta. Validity assessments and diagnostics are recorded under the hood and can be accessed as part of the analysis metadata. The Report generates a report and renders it as an html page. This is done based on the experiment design fields on the Jira tickets maintained by the experimenting team and the current analysis results.</p><h3>The Outcomes</h3><p>We’ve seen the adoption of Fisher ramping up ever since its early days as an internal product. Over 90% of all experiments in Marktplaats were conducted with Fisher within the first year of its introduction. We have seen Data Scientists’ hands-on time go down to an average of 3 hours per experiment. All in all, we estimate freeing over 9 full weeks per year, at the current rate of experimentation. And most importantly, we have achieved our goal of enabling faster experimentation by removing the bottleneck we faced!</p><p>So, we sat back and enjoyed our moment of glory..</p><p>Or not.</p><p>The scale we had just achieved in Marktplaats was great, but Adevinta is a <em>group</em> of <a href="https://www.adevinta.com/our-brands">online marketplaces</a> across Europe. Other brands are <a href="https://www.adevinta.com/our-brands/kleinanzeigen">Kleinanzeigen</a> in Germany, <a href="https://www.adevinta.com/our-brands/leboncoin">leboncoin</a> in France and <a href="https://www.adevinta.com/our-brands/subito">Subito</a> in Italy. There was a much <a href="https://medium.com/adevinta-tech-blog/working-as-a-global-team-in-a-distributed-tech-company-2df4088f2645">larger scale</a> to unlock when it came to enabling experimentation through tooling.</p><h3>Fisher at scale</h3><p>While the Marktplaats team was developing and using tools like Fisher, one of Adevinta’s <a href="https://www.adevinta.com/careers">central teams</a> developed a relatively mature, marketplace agnostic A/B testing platform called <a href="https://medium.com/adevinta-tech-blog/a-b-testing-to-improve-recommender-products-8b18c77af1c3">Houston</a>. Houston has already established a robust infrastructure with a comprehensive platform, user interface, and automation capabilities for data storage and tracking, becoming a de facto experimentation tool in many of our marketplaces. And around the time when the Marktplaats team decided to investigate how they could leverage Houston, the Houston team was thinking about how they could extend their analytics capabilities to support state-of-the-art methods at scale.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/998/0*qm1C3cXFrtRZTZNL" /></figure><p>The two teams realised that they complement each other well, and joining forces seemed like an obvious next step. It didn’t take long before Fisher was adopted into the Houston ecosystem, and became a central analytics toolkit to be shared across all Adevinta marketplaces.</p><p>By adopting Fisher, Houston became the central hub for all statistical analysis and experimentation within the Adevinta group. The platform seamlessly incorporates Fisher’s features, creating a powerful combination that elevated the organisation’s data-driven decision-making processes. With it, doors to many new possibilities were opened.</p><p>The three biggest improvements were:</p><ol><li>Support for Advanced Analysis Mode within the Houston platform</li><li>A centralised Stats-Engine Service for statistical computations</li><li>Enabling a Unified Experimentation Research process</li></ol><h4><em>Advanced Analysis mode</em></h4><p>First to come was the advanced analysis option directly within the Houston platform. Leveraging the capabilities of Fisher in Jupyter notebooks, Houston now offers an extensive range of statistical analysis tools, making complex hypothesis testing, data cleaning and report generation more accessible. This allows analysts to perform custom analyses and go beyond what is available in the Houston UI.</p><h4><em>Stats Engine for Houston</em></h4><p>Our next improvements followed naturally. The need to enhance analytical capabilities independently of the rest of the already complex system of tracking, assignment and Houston UI, gave birth to Adevinta’s experimentation stats-engine — Nightingale. Nightingale, leveraging Fisher, could now become the engine powering all complex statistical calculations in Houston. This helped ensure consistent and reliable results across Adevinta, and enabling new methods in the Houston UI became as simple as calling a different endpoint in the Nightingale service.</p><h4><em>Enabling Unified Experimentation Research</em></h4><p>Last, but not least, having a central place for implementing analysis methods, and a well defined interface around it, opened the opportunity to propel company-wide experimentation research like never before. With the creation of Experimentation Labs, and leveraging Fisher for implementation of statistical analyses, the Houston team created a simulation and validation toolkit that allows us to seamlessly test new methodologies on synthetic data, modelling many complex processes that could be encountered in reality. As a result, we can quickly make sure that whatever we release provides value even before being applied to real-world use cases.</p><p>In addition, every successful research project that yields benefits is then smoothly integrated with Nightingale thanks to both of them using Fisher. This allows every analysis to eventually end up in the Houston UI, enabling everyone in the company to reap the benefits of experimentation research. And thus the circle from research to production is closed seamlessly, enabling continuous improvement and development of the whole platform.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ge3mhC6C2CdBDR83" /></figure><p><em>A diagram showing high-level steps in the research process — from idea conception and discovery, to deployment in production.</em></p><h3>Conclusion</h3><p>Fisher is a story of <a href="https://medium.com/adevinta-tech-blog/working-as-a-global-team-in-a-distributed-tech-company-2df4088f2645">collaboration</a>, teamwork and a shared drive for excellence. It shows how well planted seeds can grow deep roots and become something much bigger — a platform and a community that is significantly more than the sum of its parts. With Fisher as a key component of our experimentation platform and our analytics toolkit, we have made experimentation faster and better, and opened the door to further incremental improvements. We are looking forward to seeing where it takes us in the future!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=110c99b993fc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/how-we-matured-fisher-our-a-b-testing-package-110c99b993fc">How we matured Fisher, our A/B testing Package</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How we joined forces to enable GPU serving for our AI solutions]]></title>
            <link>https://medium.com/adevinta-tech-blog/how-we-joined-forces-to-enable-gpu-serving-for-our-ai-solutions-ed3587b8e0d5?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/ed3587b8e0d5</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[paas]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 08 Aug 2023 13:10:51 GMT</pubDate>
            <atom:updated>2023-09-01T08:23:14.374Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Cognition team needed better autoscaling, availability and features for our GPU workloads. The Common Platform Team was there to help.</em></p><p>By <a href="https://es.linkedin.com/in/joaquincabezas">Joaquin Cabezas</a>, Machine Learning Engineer and <a href="https://www.linkedin.com/in/danielricart/">Daniel Ricart</a>, DevOps Engineer</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MuG8BaEYZ4n7svEXN0GaAg.png" /></figure><h3>Turn on the GPU again…</h3><p><em>And when everything works just fine, someone goes into the meeting room and says: “would it be possible to have a GPU for serving deep learning models directly in Kubernetes? Asking for a friend…”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fdDt6zIDnVgE-LR9" /><figcaption>A quiet and calm room in the Miyagi prefecture, Japan. Source: <a href="https://unsplash.com/es/fotos/c9zqE0QvKIU">Marisa from Unsplash</a>.</figcaption></figure><p>The light bulb in the image above is switched on. You know that pressing the power switch will turn it off, even if you don’t really understand <em>how</em> it works. In the same way, I know I can put any of my services online because I can trust that <a href="https://www.adevinta.es/stories/articles/how-to-build-and-scale-an-internal-paas-for-1500-engineers-galo-navarro">someone else will keep the lights on</a>.</p><p>But this superficial understanding also creates awkward moments: someone will ask for a feature they see as <a href="https://justsimply.dev/">“simple”, “easy” or even “trivial”</a>. Imagine someone asks you to “simply” change the location of the wall switch; this may require you to cut a hole in the wall to re-route the electrical wires.</p><p>Now imagine applying the same thinking to artificial intelligence computing. Instead of a light bulb, you’re asked to “simply” adjust GPU (Graphical Processing Unit) availability for Deep Learning Inference. Currently, there are two technical teams within Adevinta (Cognition and the Common Platform team) who are already thinking about very different ways to fulfil this “simple” request. Cognition focuses on making AI solutions available to all Adevinta while the Common Platform team (CPR) works on providing Adevinta’s multi-tenant Kubernetes distribution in the form of a Platform-as-a-Service. This is where Software Engineers across <a href="https://www.adevinta.com/our-brands">our marketplaces</a> can deploy and manage their microservices.</p><p>And now, as in Kurosawa’s classic movie <a href="https://en.m.wikipedia.org/wiki/Rashomon">Rashomon</a>, let each part tell its own story.</p><h3>Cognition’s tale</h3><p>It’s the beginning of 2023, and someone is asking to change the light bulb. My name is <a href="https://es.linkedin.com/in/joaquincabezas">Joaquin Cabezas</a> and I am a Machine Learning (ML) Engineer within the Cognition team, focused on Deep Learning for images and text.</p><p>For any image going through Adevinta’s marketplaces there is a Deep Learning model operating on it. It can be a text detector, a car classifier, a background remover etc. As these models are better served using GPUs, we usually rely on the GPUs available through Amazon Web Services (AWS) Sagemaker.</p><p>This time was different, we wanted to build a project where we could benefit from a bit more control and coupling between parts of the service. This project is Cognition’s next generation of image embeddings.</p><p>(Quick tech note: an image embedding is the representation of an image in a way that is convenient for lots of tasks like classification or computing similarity between any pair of images. It is a critical component of many ML platforms).</p><p>So, very excited about the project our team went and asked for GPUs in our cluster for the production environment. As we already have GPUs in <a href="https://medium.com/adevinta-tech-blog/introducing-unicron-our-big-data-and-machine-learning-platform-5c69fb648a41">our training platform</a>, this should be a no-brainer, right?</p><p>Graphical Processing Unit, or GPU, is just a general term for a type of computing technology that has taken the world by storm because of the AI revolution. In this context, it is a way to serve our ML models faster and cheaper, so it is a sound business case to present to upper management.</p><p>But GPUs alone are not enough to improve the service. If we are moving our serving endpoints there are also a few more things to consider. We already know this is not going to be a simple feature request…</p><h3>It’s a long way to Tipperary</h3><p>It is common to start with a list of requirements, but on this occasion it would be better described as “a wishlist”. There was no rush and nothing was broken, so we took the opportunity to outline our (long) wishlist:</p><ul><li>Use GPUs for serving</li><li>Connect our services using GRPC</li><li>A faster autoscaling than the one available on Sagemaker</li><li>Custom metrics for triggering autoscaling (not just CPU, memory…)</li><li>Full observability in our Grafana dashboards</li><li>Choose different instance types of GPU</li><li>Continue using our <a href="https://www.youtube.com/watch?v=szVzIFhzFD4">Golden Path</a> with <a href="https://fiaas.github.io/">FIAAS</a>, an abstraction to reduce cognitive load when writing Kubernetes manifests.</li></ul><p>Wait! Why overload our colleagues with so many items? Well, the list grew organically while discussing the details and defining priorities together. After all, if only one team prioritises the list before discussing it together, it’s not a joint decision.</p><h3>GRPC, a non-negotiable feature?</h3><p>So, as we already said, we are looking to boost the performance of the service. We already know that <a href="https://hackernoon.com/how-we-improved-tensorflow-serving-performance-by-over-70-f21b5dad2d98">GRPC is helping us increase throughput</a> and we didn’t want to lose that benefit, so this was a top priority. And like in many good stories, with the first clear objective comes the first clear letdown.</p><p>Turns out our Kubernetes ingress controller (a reverse proxy) is only working with HTTP/1.1, and our beloved <a href="https://dev.to/techschoolguru/http-2-the-secret-weapon-of-grpc-32dk">GRPC is working on HTTP/2</a>. Before giving up, we continued experimenting with port-forwarding on a single pod. But at a certain point we needed to connect to the service from outside the cluster (to use the load balancer).</p><p>We partnered with the CPR team, to explore alternatives — and we learned that they were already aware of this limitation which they had encountered as part of their work on the <a href="https://medium.com/adevinta-tech-blog/achieve-a-smaller-blast-radius-with-highly-available-kubernetes-clusters-96f9a475544b">SCHIP product</a>. After they shared some workable solutions, we decided to move the consumer (a JAVA application) into the cluster. This would enable us to call the service (through the load balancer) just by using the name of the deployment as hostname. And for testing, we could deploy our Locust workers with Kustomize in the same cluster, so… problem solved!</p><h3>Autoscaling for the masses</h3><p>One of the critical components in our serving endpoints is the elasticity, or the ability to acquire resources when needed (and release them as soon as they are no longer required). Applying a simple rule of autoscaling when a CPU threshold is passed over a specified period of time will probably suffice. But if we also consider other conditions we can improve the overall performance and save costs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/855/0*sMsUhzVurZQ9jxUb" /><figcaption>A simple ramp load test to trigger autoscaling. Source: Adevinta</figcaption></figure><p>“If we could create the auto scaling triggers just like we create Grafana alerts, that would be awesome”, someone commented in our Slack channel, without much hope. We were used to the AWS Sagemaker autoscaling policies, which are limited in comparison with EC2, and we thought that it would be super hard to build our desired functionality.</p><p>But we were mistaken! The Common Platform team had covered this in an experimental way, <a href="https://keda.sh/">using Keda</a>, and there was a slot for us in the next “Loko Friday” (similar to Google’s 20% side-projects). While not yet a self-serve feature (we don’t have the permissions yet), we were able to experiment with a few rules — and the feeling was positive.</p><p>This was also the litmus test that we had already full observability using our usual Prometheus and Grafana configuration. Everything we needed was there. In particular, for the <a href="https://itnext.io/monitor-deployed-tensorflow-models-with-prometheus-and-grafana-28d6135a2666">metrics coming directly from TensorFlow Serving</a>, it was really important to have the <em>processed batch size distribution</em> along with the rest of the metrics. This is a key indicator of how we are using GPU memory resources.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ylruHQ1bpbKuUFlZ" /><figcaption>Playing with different traffic patterns to watch how the processed batch size is impacted. Source: Adevinta</figcaption></figure><h3>It’s not all a bed of roses</h3><p>From the list of priorities we presented, not all were fulfilled:</p><ul><li>We still don’t have the GRPC ingress so we are constrained to a single cluster</li><li>We were accustomed to selecting different instance types per application, because it is not the same to serve a Resnet-101 image classification service than a LLM like Llama2 for text generation. Deciding on a single instance type for all of them is a hard compromise between direct serving costs and engineering effort costs.</li><li>Also, giving up our cozy FIAAS for an approach using “raw” Kubernetes manifests was a hard tradeoff. A first direct effect was that I had to take the CKA course from the Linux Foundation, which was helpful. A less appealing consequence would have been modifying our production pipeline, but luckily we didn’t have to do that.</li></ul><p>We tend to emphasise victories when writing articles like this one, but it is important to note that we experienced a combination of success and failure. As one of our Adevinta <a href="https://www.adevinta.com/careers/why-join-us">Key Behaviours</a> says: “Win and lose together”.</p><h3>Common Platform team’s perspective</h3><p>The Common Platform team (CPR) is responsible for Adevinta’s <a href="https://medium.com/adevinta-tech-blog/achieve-a-smaller-blast-radius-with-highly-available-kubernetes-clusters-96f9a475544b">SCHIP product</a>. In our feature backlog we had the ability to offer specialised hardware in our clusters, which could be <a href="https://medium.com/adevinta-tech-blog/transparently-providing-arm-nodes-to-4-000-engineers-c09c92314f2f">ARM instances</a> or instances with GPUs. When we started exploring how the instances with GPU worked, and we had developed an initial approach that worked, we reached out to our friends in the Cognition Team. They had the workloads and the expertise in using GPUs that we lacked.</p><p>I am <a href="https://www.linkedin.com/in/danielricart/">Daniel Ricart</a>, an engineer in the Common Platform team at Adevinta. As part of the team’s experimentation time, that we call ‘Loko Fridays’, I worked on providing GPU instances in our clusters. And as introduced in this post, I am currently working on offering new ways to horizontally scale the workloads running in the cluster.</p><p>As part of our feature set, we offer our users some pre-made dashboards to check the health of running applications. Each user can create their own dashboards to fit their use case, but we offer some defaults that highlight the most basic signals of any pod, ingress or deployment. Once we added the GPU-powered instances to the fleet, we extended the dashboards to show some metrics about GPU usage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oVNZ94kHmFUm6qSR" /></figure><p>Another default feature of SCHIP is scaling based on CPU and memory usage. With this feature, any user can make sure their applications respond automatically to increased demand or reduce the number of idle resources when not in use.</p><p>By leveraging the newly added GPU-powered instances, the extended application dashboards and GPU metrics, and existing basic autoscaling features, Cognition, and any other ML team could start their tests. And they jumped into it.</p><p>We offered the same experience they already knew, to deploy CPU-based applications with a new resource:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/516/0*1e9_IEMgL7cB0Flp" /></figure><p>Using this resource, Cognition can deploy GPU-based applications in SCHIP using the interfaces they are familiar with. But there was a previously unknown caveat: they can only scale their applications based on the CPU load of a GPU-based application. There’s a correlation between the usage of both resources, but it’s not similar enough to be reliable.</p><h3>Scaling</h3><p>At Adevinta, we allocate one day every two weeks to run technical explorations. In CPR this day is called “Loko Friday” and it has fostered several initiatives that are now part of the portfolio. If an idea that comes up is approved by the rest of the team, and it adds value to the users, it is usually further developed and added to our regular portfolio as Alpha stage. Later it will graduate as a supported feature. As part of this allocated time, I started exploring a long-forgotten user request: how feasible would it be to offer horizontal scaling based on any metric instead of just allowing the usual CPU and memory metrics?</p><p>I had some breakthroughs, and leveraging components already available in the cluster for internal use, so I came up with a very neat solution in a couple of “Loko Fridays”. The proposed solution still required polishing and it was primarily aimed at advanced users of SCHIP, because it lacked integration with FIAAS. But according to my tests, we <em>should</em> be able to test the solution with real workloads.</p><p>And while that was happening from Friday to Friday, as part of our conversations with Cognition engineers, they sent us their wishlist for the next steps of serving GPU-based applications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/398/0*AQheeHJ-zuXCOUE2" /><figcaption>Everything started from a Slack conversation on 11 July 2023. Source: Adevinta</figcaption></figure><p>Here we had the second collaboration with the Cognition Team. In CPR, we were ready to test and validate our technical exploration with some users, and this was the perfect opportunity. Also, Cognition wanted to explore this new feature we were testing.</p><p>On the following Friday I shared with Joaquin that I was working on this feature, and guided him through the required setup on his side. That same day we generated some satisfactory results and a short list of improvements to the feature.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*REvGW7YZATeD27yr" /><figcaption>Number of pods serving a tensorflow application. Source: Adevinta</figcaption></figure><p>What’s the magic behind it?</p><p>In SCHIP we use <a href="https://keda.sh/">keda.sh</a> (Kubernetes Event-driven Autoscaler) to scale ingress based on the load we are serving at any given time. It periodically runs a Prometheus query to fetch average requests and perform the required calculations to maintain a pool of ingress pods. To setup a KEDA autoscaler, you provision a ScaledObject to your namespace with some specific settings about the scaling behaviour and datasource. It then controls an horizontalPodAutoscaler object to add or remove application pods according to the values of your ScaledObject.</p><p>During my “Loko Friday” work, I crafted a helm chart that encapsulates a ScaledObject.keda.sh object with all the settings that fit SCHIP. This is then connected to the appropriate Prometheus that hosts the tenant metrics. There are some customisation options like the Prometheus query to run, minimum and maximum size of the pool and a few other parameters. Most of the parameters of the object are tailored to a sane approach we have already tested in other cluster components.</p><h3>Wow!</h3><p>So far we have fulfilled four-and-a-half items of the Cognition wishlist.</p><ul><li>✅ Use GPUs for serving</li><li>⚠️ Connect our services using GRPC</li><li>✅ Faster autoscaling than the one available on Sagemaker (a new instance in less than 3 minutes)</li><li>✅ Custom metrics for triggering the autoscaling (not just CPU, memory etc.)</li><li>✅ Full observability in our Grafana dashboards</li><li>❌ Choose different instance types of GPU (not possible for now)</li><li>⁉️ Continue using our Golden Path with <a href="https://fiaas.github.io/">FIAAS for </a>Kubernetes</li></ul><p>Why “and a half”? Well, deploying workloads using GPU instances is already part of the FIAAS for Kubernetes as an extension we added previously. Internally, FIAAS will add the required resource blocks to make sure that the pods are scheduled in a node with GPU support.</p><p>The ability to scale based on custom metrics is still an alpha feature that is not yet exposed to FIAAS in any way.</p><h3>A few final words</h3><p>It was funny that in both teams we thought we were asking for help, when really, it was a hot topic that happened to appear for us all at the same time. This just shows how valuable it is to share notes and observations beyond your team.</p><p>It also highlights how allocating some time for innovation (“Loko Fridays”, Discovery Days etc.) fosters collaboration, and drives spontaneous innovation that adds value to <em>all</em> users.</p><p>We have described how the need of using a feature and the willingness to serve this feature is a powerful combination that needs to be leveraged when it (rarely) appears at the same time. This pushes teams to work together in the best way possible. We hope this article (which is really about collaboration between two technical teams) will inspire you to find these “matches” in your organisation to join forces during “long ways” (to Tipperary…), not just high intensity meetups.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed3587b8e0d5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/how-we-joined-forces-to-enable-gpu-serving-for-our-ai-solutions-ed3587b8e0d5">How we joined forces to enable GPU serving for our AI solutions</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Our microsite architecture]]></title>
            <link>https://medium.com/adevinta-tech-blog/our-microsite-architecture-df656527a29a?source=rss-b8cb87b84913------2</link>
            <guid isPermaLink="false">https://medium.com/p/df656527a29a</guid>
            <category><![CDATA[frontend]]></category>
            <category><![CDATA[microsite]]></category>
            <category><![CDATA[backend]]></category>
            <category><![CDATA[nextjs]]></category>
            <category><![CDATA[microservices]]></category>
            <dc:creator><![CDATA[Adevinta]]></dc:creator>
            <pubDate>Tue, 11 Jul 2023 08:32:16 GMT</pubDate>
            <atom:updated>2023-09-01T08:27:41.415Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How we accelerated the frequency of releases and the independence of the teams for the site </em><a href="http://subito.it/"><em>Subito.it</em></a><em> by applying a microsite approach</em></p><p>By <a href="https://www.linkedin.com/in/alessandro-grosselle-28965b18/">Alessandro Grosselle</a>, Senior Frontend Engineer</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Re7bJfr0GNRK1sLol_RGEg.png" /></figure><p><a href="https://www.adevinta.com/our-brands/subito">Subito</a> is a re-commerce generalist marketplace and part of <a href="https://www.adevinta.com/about-adevinta">Adevinta</a>. At <a href="http://subito.it/">Subito</a>, we are organised into different teams, each with well-defined objectives. For example, there is the Content Team responsible for rendering ads and listing announcements, and the User Team responsible for the account authentication process.</p><p>Each team is autonomous and independent, with specific result-oriented goals and responsibilities.</p><p>From a technical point of view, following this type of organisation as closely as possible is a very stimulating challenge:</p><ul><li>A team should be able to release a new feature without waiting for other teams.</li><li>In case of bugs or malfunctions, responsibility should be clearly assigned.</li><li>To minimise dependencies, we must build components that are as decoupled as possible; or better yet there should be no dependencies between components from different teams.</li></ul><p>Subito’s backend has brilliantly tackled this challenge by implementing a microservice architecture. Like many of you, we started from a monolith website and wanted to implement an architecture that would allow the various teams to work as independently as possible. How did we achieve this?</p><h3>The idea</h3><p>We were looking for a microservices-style solution that could be easily integrated and, above all, implemented without interrupting the product’s evolution. We also wanted it to align well with our technological choice of adopting “<a href="https://nextjs.org/"><em>Next.js</em></a>” as the web framework.</p><p>During this study phase, one of the most important activities was defining the various responsibilities. The output was very interesting because it allowed us to understand that ownership could be at the “page level” rather than at the level of individual components. For example, <a href="https://areariservata.subito.it/ricerche?from=dropdown"><em>the saved searches page</em></a> belonged to the User Engagement Team, the<a href="https://areariservata.subito.it/login_form"> <em>login page</em></a> to the “User quality” team, and so on.</p><p>Here’s the main insight: <em>our monolithic architecture could become an architecture composed of many small independent sites, where each site has, at most, a few pages. Each microsite then belongs to a specific team:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YkfdMBzlIzudiHm7" /></figure><ul><li>Each microsite has its own git repository.</li><li>Each microsite has its own exclusive release pipeline.</li><li>Each microsite has its defined routes.</li><li>Each microsite has a defined team (obviously).</li></ul><p>We unimaginatively called this a “<strong>microsite architecture.</strong>”</p><h3>Development</h3><p>An architecture like this seemed immediately actionable and implementable without heavily blocking product activities.</p><p>We gradually migrated all the pages, step by step. Starting with <a href="https://areariservata.subito.it/login_form"><em>the login page</em></a>, we created its repository, <a href="https://nextjs.org/">Next.js</a> app and defined the route on Akamai.</p><p>The work to take the “first” microsite into production was the toughest because as Frontend Developers, we didn’t have much experience in CDN configuration and deploying individual applications in our cluster. But with the support of our Platform Team, we completed the migration successfully.</p><p>During our second migration, the <a href="https://areariservata.subito.it/preferiti?from=dropdown"><em>favourites page</em></a>, we realised that the process was always the same:</p><ul><li>Create the repository.</li><li>Create a <a href="https://nextjs.org/">Next.js</a> app, applying the same layouts, header and footer.</li><li>Define the pipeline.</li><li>Define the route.</li></ul><p>From this, we understood that to speed up the migration we could implement a scaffold based on <a href="https://nextjs.org/">Next.js</a>, with everything needed to develop and deploy a new microsite. Where automation was not possible, we created guides to help developers bring the microsite into production. For example, route definition is not automated yet, but is easily configurable from the Akamai dashboard. For this step, we created a guide which we’ll discuss more in the next section.</p><h3>What if Team A wants to include a component on Team B’s page?</h3><p>This is clearly a legitimate requirement. For this case we defined a process:</p><ol><li><strong>Team A</strong> implements the new component in our shared monorepo, wrapping it in an ErrorBoundary component to “catch” the error and prevent the entire React tree from breaking:</li></ol><p>&lt;ErrorBoundary fallback={&lt;p&gt;Something went wrong&lt;/p&gt;}&gt;</p><p>&lt;Profile /&gt;</p><p>&lt;/ErrorBoundary&gt;</p><p>2. The component is released with its own version in our artifactory.</p><p>3. The component is then included at build time in the project of <strong>Team B.</strong></p><p>In this way, Team B’s project contains only a few lines of code from Team A: importing the component and using it. And <em>the page always belongs to Team B, hosting a component from Team A.</em></p><h3>Conclusions</h3><p>Today, the Subito website is mostly composed of microsites.</p><p>The monolith is still online but only serves a few pages (less than five) with very little traffic. Ironically, we can say that the monolith has also become a microsite itself.</p><p>This approach works well for our context: we are agile, able to deploy quickly and have a clear technological division. Of course this is not the only architecture approach; the right one for you will depend on how your company is structured.</p><p>From our experience, we have learned that before diving into an architectural restructuring, it is always good to pause and identify the solution that is the most feasible. This may not be “the best” solution, but it will be achievable — and it will work.</p><p>To conclude, years after adopting this architecture, we can evaluate the pros and cons of the approach chosen over time:</p><h3>Pros</h3><ul><li>Maintenance of a microsite is much simpler. For example, updating to React 18 for a project with one or two pages is much easier than updating an entire application that may have involved many developers and teams.</li><li>Each team can experiment more quickly. The project belongs to the team, and the app belongs to the team. They are free to integrate new libraries, which will positively or negatively impact their team.</li><li>Due to this page-level separation, the production deployment of pull requests has increased, requiring less time to obtain approvals.</li></ul><h3>Cons</h3><ul><li>In Subito, we have a monorepo where we share common components among all teams, such as the header and footer. Updating these components in all microsites (e.g., adding a new item to the footer) manually is very tedious. Currently, it is still a manual process, but fortunately, the frequency of updating the header and footer is very low. (Spoiler: we are working on automating these types of updates).</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df656527a29a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/adevinta-tech-blog/our-microsite-architecture-df656527a29a">Our microsite architecture</a> was originally published in <a href="https://medium.com/adevinta-tech-blog">Adevinta Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>