<?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 Dhruvik Shah on Medium]]></title>
        <description><![CDATA[Stories by Dhruvik Shah on Medium]]></description>
        <link>https://medium.com/@dhruvik04_44048?source=rss-c750aaa78a86------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*zy9peXlc9WlSCoxM</url>
            <title>Stories by Dhruvik Shah on Medium</title>
            <link>https://medium.com/@dhruvik04_44048?source=rss-c750aaa78a86------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:03:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dhruvik04_44048/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[Showing Delivery SLA at 10x Scale with 1/10th Latency]]></title>
            <link>https://blog.flipkart.tech/showing-delivery-sla-at-10x-scale-with-1-10th-latency-28b17b198cc8?source=rss-c750aaa78a86------2</link>
            <guid isPermaLink="false">https://medium.com/p/28b17b198cc8</guid>
            <category><![CDATA[flipkart]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[binary-encoding]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[aerospike]]></category>
            <dc:creator><![CDATA[Dhruvik Shah]]></dc:creator>
            <pubDate>Thu, 22 May 2025 05:37:02 GMT</pubDate>
            <atom:updated>2025-05-22T15:45:23.105Z</atom:updated>
            <content:encoded><![CDATA[<h3>Showing Delivery Date at 10x Scale with 1/10th Latency</h3><h3>Context</h3><p>Let’s begin with a brief overview of what a supply chain entails and the process involved in providing a Delivery Date to users on Flipkart.</p><p>Similar to other e-commerce platforms, an order placed on Flipkart undergoes several stages before reaching a customer. The journey starts at the seller’s location or warehouse, where the requested item is stored. Once the order is placed, the item is picked; packed into a shipment, and handed over to eligible delivery vendors. These vendors operate multiple (Mother) hubs where shipments are sorted and bagged according to their destination addresses. From there, the shipment is sent to the Delivery Hub, the hub closest to the customer’s address. Finally, a Delivery Agent is assigned to deliver the shipment directly to the customer’s doorstep.</p><figure><img alt="Supply Chain Network" src="https://cdn-images-1.medium.com/max/1024/0*Hkvuh0pnSuxK6nBg" /><figcaption>Supply Chain Network</figcaption></figure><p>We, the Promise team, are responsible for determining and providing the <strong>Delivery Date to customers</strong>, taking into account the daily capacities at various stages of the supply chain.<br>Across Flipkart, the Customer Promise Date (delivery date) is displayed at multiple points in the user journey, including the Product Page, Cart, Order Summary Page, and for certain products, on the Search &amp; Browse Page.</p><h3>Nomenclature</h3><ul><li><strong>Delivery Date (or Customer Promise Date):</strong> The specific date when an item will be delivered to the customer.</li><li><strong>Source:</strong> The warehouse or Seller location where the item is stored.</li><li><strong>Source Cutoffs:</strong> Specific times of the day when a source can dispatch items. For example, a warehouse might have cutoffs at 1:00 PM, 4:00 PM, and 10:00 PM.</li><li><strong>Source Capacity:</strong> The number of items a source can process at a given time. For instance, if a source capacity at the 12 PM cutoff is 10, it can only dispatch 10 items by that time. If the capacity is fully consumed, we move the order to the next available cutoff.</li><li><strong>Vendors: </strong>Delivery partners, such as Ekart, Bluedart, Delhivery, etc., which are responsible for shipping items from the source to customer.</li><li><strong>Vendor Capacity:</strong> The number of items a Vendor can process per day or timeslot.</li></ul><h3>Exact Delivery date calculation on Product and Order Summary Page</h3><p>When a product page is loaded, among other APIs, the Promise API is triggered to retrieve the delivery date for a specific combination of the Product Source and the Destination Pincode. Similar API calls are made on the Cart and Order Summary pages to compute the delivery date.</p><p>On a high level, the delivery date is calculated based on below parameters:</p><p><strong><em>Actual_Delivery Date = now + procurement_sla + source_handover_time + vendor_sla + vendor_capacity_days</em></strong></p><p><strong><em>Procurement_sla:</em></strong><em> Time required to pickup the item from seller/warehouse and reach flipkart hub<br></em><strong><em>Source_handover_time:</em></strong><em> Time to pack and dispatch the item to selected vendor<br></em><strong><em>Vendor_sla:</em></strong><em> Time taken by the vendor to take it to Delivery Hub (DH) closest to Customer<br></em><strong><em>Vendor_capacity_days:</em></strong><em> Number of days after which delivery would be attempted based on DH capacity</em></p><h4><strong>API Configuration</strong></h4><p>The avg batch size for API calls on the Product Page &amp; Order Summary page is <strong>7 items</strong> and <strong>5 items</strong> respectively. To calculate the delivery date, the following factors need to be considered:</p><ul><li><strong>Identifying Available Sources:</strong> Determining which sources can fulfil the requested item.</li><li><strong>Valid Delivery Routes:</strong> Identifying which routes from those sources can deliver the item to the customer’s address.</li><li><strong>Vendor Capabilities:</strong> Determining which vendors can pick up and deliver the items along those routes, while considering constraints such as category, item size, and payment mode. For instance, some vendors may only support prepaid options and not Cash on Delivery (COD), or certain vendors may not handle large category items (like furniture or washing machines).</li><li><strong>Optimized Source-Vendor Combination:</strong> Identifying the most optimized source-vendor pairing to ensure the best delivery experience for the customer.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1012/0*dukgahqj2RncvW7y" /></figure><p>Due to Flipkart’s high throughput and extensive network, the timeout for this API is set to 1000ms on the Product page and 1200ms on the Order Summary page to complete the computation.</p><h3>Estimated Delivery Date calculation on Search &amp; Browse (S&amp;B) Page</h3><p>When an item is searched on Flipkart, the user is given hundreds of options on what we call the Search and Browse(S&amp;B) page. The user query is sent to the Search ecosystem, which internally calls Promise’s API with a batch size of ~100 for the shortlisted hundreds of items to check the Availability and Serviceability to customer&#39;s pincode before presenting the results on the S&amp;B page.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/859/0*8ZSv2hH-T9xlYyP2" /></figure><h4>API Configuration</h4><p>Given that this is a search result, to ensure a smooth user experience, Search calls the Promise API in a batch of <strong>~100 items</strong> with a latency constraint of just <strong>100ms</strong>.</p><p>Currently, to meet this constraint of 100ms for 100 items, Promise <strong>doesn’t</strong> consider the source and vendor capacities while computing the delivery date, as doing so would be computationally expensive. It only checks whether the item is available &amp; if any of the vendor can serve the customer’s pincode. Therefore, the delivery date is determined using the fastest serviceable vendor’s delivery time, assuming capacity is available at the source and the Delivery Hub (DH).</p><p><strong><em>Estimated_delivery_date = current_time + vendor_sla</em></strong></p><p>This approach, while efficient, does not provide the correct delivery date for the user on the S&amp;B page in all scenarios. If we compute similar to other pages of the Flipkart, where 7 items take over 500ms (P99), it would be impossible to meet the 100ms constraint for ~100 items on the S&amp;B page. <br>This is done to ensure that availability and serviceability checks do not negatively impact the search page load times. As a result, except for a few items with offline delivery date calculations, the delivery date is not shown on the S&amp;B page.</p><h3>Problem Statement</h3><p>The goal is to <strong>display the delivery date in the Search and Browse (S&amp;B)</strong> section for all items, with a focus on items that can be delivered within 4 days. The calculations must be completed within the current S&amp;B API timeout of ~100ms for a batch size of 100 items. In addition to this latency requirement, the solution must also scale to handle 10x of current Product page QPS across all data centers (DCs) and availability regions.</p><blockquote>Compute Delivery date for 100 Items in 100ms at 10x QPS instead of 7 items in 500ms</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*S5dOmkRsB63AU87R" /><figcaption>Search and Browse Page with Delivery Date as seen on Flipkart</figcaption></figure><blockquote>Scope</blockquote><blockquote>The items are stored either in Flipkart’s warehouse or the seller’s location. Initially, we focused on items stored in Flipkart’s warehouse, as these tend to have faster delivery times due to a reduced procurement SLA. These are referred to as <strong>Fulfilled By Flipkart (FBF)</strong> items.</blockquote><p>Short on time? The TL;DR offers a high-level summary of the delivery date calculation optimisation. For a detailed breakdown of each approach and the technical challenges faced, continue reading the full blog.</p><h3>TL;DR</h3><p>Initial attempts to integrate the Product Page API flow for 100 items resulted in high latency (~1000ms) and CPU utilization. Caching ItemID x Pincode was impractical due to the massive data size (51TB). Source x Pincode caching reduced data cardinality but required considering item attributes (Pivots), leading to ~68 million combinations which proved challenging due to frequent capacity updates.</p><p>The final solution involved optimizing the existing flow by decoupling Source and Vendor Capacity stores. Source capacity was cached per cutoff in Aerospike, with capacity represented as bit-encoded integers to minimize data size and network utilization (90% reduction). Vendor capacity was cached at the Vendor-Pincode level, also using bit encoding to represent capacity for the next 25 days. These optimizations enabled real-time delivery date calculation within the 100ms latency constraint on the S&amp;B page with over 90% accuracy and supporting 10x QPS.</p><h3>Journey</h3><p>We began by brainstorming various approaches to calculate delivery dates within the 100ms limit for a batch size of ~100 items using the Promise API.</p><ol><li><strong>Integrating the Product Page Flow in the S&amp;B API</strong></li></ol><p>One potential approach was to incorporate the entire computation process from the Product Page API into the S&amp;B API. However, when we attempted this, the latency for 100 items was ~1000ms at P99 (for 7 items, it’s 500+ms on Product Page). Even with parallel computation, we observed that not only were we breaching the 100ms limit, but CPU utilization reached 100% on instances with 16 cores. To handle ~10x QPS, we would have to deploy numerous additional cores, making this approach impractical. Therefore, integrating the existing logic wasn’t feasible.</p><p>2. <strong>Exploring <em>Item_ID x Customer_Pincode -&gt; Delivery Date</em> Cache</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/0*CQ1MbwS_rc3X25R5" /><figcaption><em>Schema for </em><strong><em>Item_ID x Customer_Pincode -&gt; Delivery Date</em></strong><em> Cache</em></figcaption></figure><p>If we implement a cache where the delivery date for a given item and destination Pincode can be retrieved in O(1) time, the latency for our API would only see minimal increases. We can make a bulk call to the cache with 100 items + pincode from the request, and return the result without additional computation.</p><p>The main challenge, however, is the vast number of unique items being sold on Flipkart. Currently, there are approximately 270 million FBF items listed on Flipkart. If we factor in ~19,000 Pincodes that Flipkart delivers to, the total number of rows in the cache would be:</p><p>270,000,000 items * 19,000 Pincodes = 5,130,000,000,000 rows.</p><p>Assuming each row takes upto 10 bytes, this would amount to around 51 TB of data. Given the hardware costs and the need to update records with every capacity change, this solution becomes impractical.</p><p>3. <strong>Exploring <em>Source x Pincode -&gt; Delivery Date</em> Cache</strong></p><p>If we can reduce data cardinality, we might still be able to go with the above approach. What if, instead of each item to Pincode, we just store Source to Pincode mapping? For instance, a source (S1) to a Pincode (P1) would have a fixed SLA based on item attributes. So, can we consider the same SLA for all items stored at a particular source? Of course, we would also need to consider which vendor to select, as each vendor would have a different SLA. Considering only FBF sources, this would reduce the data cardinality to about:</p><p>40 warehouse locations * 19,000 Pincodes = 760000 rows</p><p>This is a significant reduction, making the approach more feasible.<br>However, selecting the right source for a given item becomes a challenge. The source and its corresponding delivery date depend on multiple attributes of the item. Some of the item-level attributes that can affect the SLA between a given source and destination include:</p><p><strong>ItemSize:</strong> Small, Medium, Large etc.<br><strong>Category:</strong> Book, apparel, electronics, mobiles, computers, etc.<br><strong>ShipmentType:</strong> Inflammable, Fragile, Dangerous, etc.</p><p>In total, there are about 18 different attributes/Pivots that could influence the source &amp; vendor selection and, consequently, the delivery date computation. On average, there are approximately 90 combinations of these Pivots. For example, a <em>single source</em> could have items that are <em>small, books, and fragile, or small, electronics, and dangerous</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R956JegkiuJf5fPG" /><figcaption>Schema of <strong><em>Source x Pincode -&gt; Delivery Date</em></strong><em> Cache. Here, x, and y may or may not be same depending on the various item level pivots.</em></figcaption></figure><p>If we can store and look up the delivery date for each combination of Source x Destination Pincode x Pivots across the 90 possible combinations, we would be able to retrieve the Delivery date for all potential combinations.</p><p>40 FBF sources x 19000 Pincodes x 90 Pivots combinations = ~68 million combinations</p><p>From a size perspective, 68 million combinations are manageable — approximately 5GB (considering Pivots storage size). This could easily be stored in an in-memory datastore, allowing for fast lookups.</p><p><strong>The challenge?</strong> — Keeping them updated.</p><p>The delivery date from a source to a pincode also depends on the capacities available at a particular source or vendor on any given day. If, on Day D0, all source or vendor capacities are full, we would need to schedule dispatch or delivery for the following day. Thus, for every order Flipkart receives, we would need to track the consumption of source/vendor capacities and update multiple Source x Pincode x Pivots combinations based on these updates.<br>With Flipkart processing millions of orders daily, this approach became infeasible due to the massive number of calculations and the millions of updates required every second.</p><p>4. <strong>Optimise and approximate current flow</strong></p><p>We began by analysing the Product/Order Summary Page flow and breaking it down into three main steps:</p><ol><li>Pivot Matching: For each item in the request, retrieve the item attributes from central service to match the Pivots of Source Capacity.</li><li>Source Capacity: Based on the matched Source Capacity, calculate the handover time to the vendor.</li><li>Vendor Capacity: From the handover time, calculate the vendor’s delivery SLA for eligible vendors based on Vendor Capacity.</li></ol><p>To optimize this process, we decided to split the cache store mentioned in Approach 3 into two separate stores:</p><ul><li>Source Capacity Store (Step #2)</li><li>Vendor Capacity Store (Step #3)</li></ul><p>By doing this, we wouldn’t need to recompute the delivery date with every change in source and vendor capacity. Instead of directly storing the delivery date, we could store simplified versions of both capacities and compute the delivery date in memory.</p><p>We collaborated with the planning team, who are responsible for defining and updating the future Source and Vendor Capacities. After discussions with them, we discovered that most Source Capacity plans rely primarily on two Pivots out of the 18 pivots:</p><ul><li><strong>Tier (Tier1, Tier2, Tier3, Tier4) </strong>— This defines the eligible supply chain mode for the order’s shipping. For example, Tier1 signifies faster shipping.</li><li><strong>Vendor Type (LT1, LT2, LT3, LT4, LT5) </strong>— These represent the various delivery partners associated with Flipkart.</li></ul><p>By approximating handover time using only the Tier and Vendor Type pivots, we could avoid matching against all the source capacity plans. While this won’t be 100% accurate, it should give us approximately 95% correct handover times for FBF items, as over 95% of the capacity plans don’t have additional pivots.</p><p><em>Additionally, reducing the number of pivots meant we no longer needed to call the central service to fetch item attributes (Step #A). This change saved around 50–100ms per request.</em></p><h4>Source Capacity Store:</h4><p>A source capacity is recorded at each cutoff time throughout the day. The capacity is stored for each cutoff at Source x Pivot combination. To optimise the process, we decided to cache the capacity for each cutoff. Start by matching Pivots from first cutoff and stop as soon as we are able to find the capacity. However, building this cache required creating a complete data flow pipeline to ensure that the source handover time is recalculated whenever the capacity for a cutoff is exhausted.<br>As illustrated in the diagram below, in addition to capacity updates triggered by orders and reservations, any new source capacity uploads would also trigger corresponding updates to the data pipeline. The ingestion cluster would retrieve the relevant data from the source capacity MySQL database, recalculate the Source Capacity, and update our new cache-store.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DeKym8-BlF2752Yc" /><figcaption>High level Design</figcaption></figure><p>Since we already use Redis as one of our data stores, where the data is updated with each order reservation, we built a pipeline to publish these updates to our message broker.<br>The pipeline consists of several key components:</p><ul><li><strong>Redis Replicator:</strong> Source capacity updates in Redis from order reservations are sent to a message broker via Redis replicator.</li><li><strong>Message Broker:</strong> We chose Apache Pulsar as our message broker to publish updates regarding source capacity consumption or new uploads.</li><li><strong>Ingestion Cluster:</strong> This component listens for updates and fetches data from the source capacity MySQL store. It then recalculates the capacity for each source cutoff, considering the Tier and Vendor Type pivots.</li><li><strong>Distributed Cache:</strong> The recalculated data is stored in a distributed cache, which is then used by the user request-serving application to calculate the delivery date.</li></ul><p><strong>NFR requirements for Source Capacity Cache store <br></strong>The cache store must be available across all the availability regions, ensuring that any update is propagated to both regions with minimal delays (a few milliseconds). Currently, we break a request for 100 items into 4 parallel compute batches. For each query, considering approximately 5 sources and 10 cutoffs per source across the 100 items, a request would result in <strong>5 * 10 * 4 = ~200</strong> key fetches from the cache.</p><p>Given the expected high QPS for this API on the Flipkart search platform, the cache would need to support <strong>few</strong> <strong>10s of millions of </strong>fetches per second.</p><p>After several evaluations, we chose Aerospike as the solution. It supports our high QPS requirement and provides Cross-DC Replication (XDR) capabilities. This means we can write updates to one availability region and use XDR to replicate those updates to the other region.</p><p><strong>Schema stored in aerospike:<br></strong>To reduce size, we created ordinals for Tier &amp; VendorType, so that it would be stored as integer rather than strings. For example, a value of <strong>907 at (0,1)</strong> represents a capacity of 907 in the Tier1 and LT2 VendorType for the source and cutoff <strong>s1.1690203601</strong>. Similarly, Tier2 and LT1 (1,0) has 0 capacity available in s1.1690203601<strong>.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pHwMMJTZtIOvUuvk" /><figcaption>Source Capacity Cache Schema</figcaption></figure><p>With this approach, we were able to calculate the handover time in under 10 ms under BAU load. However, during load testing for sale, Aerospike instances began experiencing <strong>network throttling</strong> due to the large volume of transmitted data. This issue arose at around 12K+ QPS.<br>To reduce the data size in the Aerospike response, we optimised the storage by retaining only non-zero values for each capacity and storing only the available capacity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0kP0pvsR1tLStC4o" /><figcaption>Source capacity Schema without 0 capacities</figcaption></figure><p>After this optimization, the network utilization was reduced by more than half.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6WXW7joxCkoRP6Rs" /><figcaption>Improvement in Transmit &amp; Receive Bandwidth</figcaption></figure><p>We were still encountering 1% timeouts due to network latency. To address this, we decided to store the data in an encoded format. Since our primary concern was not the actual remaining capacity but simply whether the capacity was available (i.e., greater than 0), we represented the data using bits. Each bit indicates whether a specific VendorType (LT) and Tier combination has capacity available for the given cutoff.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RB2RMb2qYL09HDFm" /><figcaption>Each bit representing availability of Source Capacity for Tier x VendorType combination</figcaption></figure><p>This change significantly reduced our network utilization. Previously, we stored a map of Tier/VendorType Ordinals to Capacity in Aerospike, which consumed over <strong>300 bytes</strong>. We switched to using an integer data type, <strong>which only requires 8 bytes, resulting in a 90% reduction in network bandwidth usage.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-5vm-RkHwzJwyj8k" /><figcaption>Bit encoded Source Capacity Cache</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YOT-t1AqFqvwsYLV" /><figcaption>Optimization achieved in Transmit &amp; Receive Bandwidth</figcaption></figure><p>We retrieve the <strong>Capacity integer</strong> from Aerospike and perform bit operations in memory within the application to check the availability of capacity for a specific <strong>Tier</strong> and <strong>VendorType</strong>. This optimisation allowed us to improve the Source capacity store, reducing the handover time computation to approximately <strong>10ms even at 10x QPS of Product Page</strong>.</p><h3>Vendor Capacity</h3><p><strong>Vendor spillage</strong> refers to the number of days the capacity at the Delivery Hub (DH) remains unavailable after an item arrives at the DH. So Vendor spillage of 2 means, a vendor cannot deliver any new order until next 2 days.</p><p>We found that the <strong>P95</strong> for vendor spillage is typically 0 on normal days, and remains constant during sale periods. This indicates that, on normal days, the item is attempted for delivery on the same day it reaches the DH.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uGwjGAwP4uAnYp-d" /><figcaption>Vendor Splliage as observed during non-sale days</figcaption></figure><p>We initially added a constant spillage of 0 days to the handover time retrieved from Aerospike. With this approach, we were able to calculate the delivery date with <strong>80% accuracy</strong>. However, during sale days, when orders increased by upto 5 times, the accuracy dropped to<strong> below 50%</strong>. To improve accuracy, we began analysing how capacity is looked up on the Product Page.<br>On the Product Page, we first determine the <strong>Delivery Hub</strong> for each eligible vendor based on the customer’s pincode. Then, we fetch the available capacity and compute the delivery date based on that capacity.<br>To streamline this process and avoid the first step (fetching the Delivery Hub for the customer’s pincode), we created a new cache in Aerospike. This cache directly provides the capacity at the <strong>Vendor-Pincode level</strong>. It stores capacity for the next 25 days in bit format, where each bit at position X represents whether the vendor has capacity available on the <strong>Xth day</strong> from today. For example, <strong>101101101101101011010101</strong> indicates that capacity is available today (represented by 1 in the most significant bit), not available tomorrow (0 in the second position), and so on. This allowed for quick lookups while minimizing memory usage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/0*FM_2onqawKH1PRev" /><figcaption>Vendor Capacity Cache Schema</figcaption></figure><p>We used the same aerospike cluster and data pipeline which we had created above and the accuracy during sale days reached above 80%.</p><h3>Result</h3><p>Previously, we had workarounds for caching the delivery date on the Search side by querying the Product Page Promise API offline, but this approach wasn’t scalable beyond a fixed set of categories and pincodes. With the optimisations we implemented, we were able to provide real-time delivery dates on the Search &amp; Browse (S&amp;B) page, all while maintaining the 100ms latency constraint.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/714/0*hONAKdOGVWosXnkB" /><figcaption>Search and Browse Page with Delivery Date as seen on Flipkart App</figcaption></figure><h3>Conclusion</h3><p>In the early stages, our goal was clear: enhance the user experience on the S&amp;B page by displaying delivery dates and helping customers make more informed decisions. We embarked on our journey with straightforward and intuitive solutions, gradually evolving towards more intricate processes, and tackling challenges one step at a time. The pivotal decision lay in striking a balance between integrating extensive hardware for 100% accuracy and opting for a more resource-efficient approach, achieving a commendable 90%+ accuracy.<br>As users move through the funnel — from Search to Product to Order Summary Page — the delivery date accuracy improves, though at the cost of increased latency. This was due to our strategic decision to bring computation closer to the data source rather than relying solely on the cache. This allowed us to take a more streamlined approach, which enabled a faster transition to production. With this, we were able scale the Order Summary/Product Page flows, to 10x throughput on Search &amp; Browse at just 100ms (1/10th) latency.<br>Our next steps involve increasing accuracy by adding more pivots to the current solution and enabling the ability to filter items based on the delivery date on the Search &amp; Browse page.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=28b17b198cc8" width="1" height="1" alt=""><hr><p><a href="https://blog.flipkart.tech/showing-delivery-sla-at-10x-scale-with-1-10th-latency-28b17b198cc8">Showing Delivery SLA at 10x Scale with 1/10th Latency</a> was originally published in <a href="https://blog.flipkart.tech/">Flipkart Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>