<?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 Jamie Thirlwell on Medium]]></title>
        <description><![CDATA[Stories by Jamie Thirlwell on Medium]]></description>
        <link>https://medium.com/@jamie.thirlwell_22015?source=rss-eabdb7be36c0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*1dkncVdTt0vbC63r</url>
            <title>Stories by Jamie Thirlwell on Medium</title>
            <link>https://medium.com/@jamie.thirlwell_22015?source=rss-eabdb7be36c0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 03:23:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jamie.thirlwell_22015/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[Build, Baby, Build.]]></title>
            <link>https://tech.loveholidays.com/build-baby-build-276e445f539e?source=rss-eabdb7be36c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/276e445f539e</guid>
            <category><![CDATA[continuous-integration]]></category>
            <category><![CDATA[github-actions]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[google-cloud-build]]></category>
            <dc:creator><![CDATA[Jamie Thirlwell]]></dc:creator>
            <pubDate>Mon, 02 Mar 2026 15:14:31 GMT</pubDate>
            <atom:updated>2026-03-02T15:25:48.374Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Blue building blocks illustration, used to represent a CI build" src="https://cdn-images-1.medium.com/max/895/1*SJ7hcugqugmnu1EIFNndjQ.png" /></figure><h3>Waste in Continuous Integration</h3><p>There’s a surprising amount of waste hidden in a typical software engineering process.</p><p>Continuous integration (CI) is where this waste really shows up.</p><p>Builds are constantly waiting, running, testing, failing… then waiting and running again. Sometimes a single build takes more than 20 minutes. That’s 20 minutes where a developer is blocked, context is lost, and flow is broken. Multiply that across a team, and it becomes very expensive for the business and devastating for engineering productivity.</p><p>And it’s not just humans who suffer from slow feedback. As AI takes on a bigger role in how we write code, with tools like Claude Code writing and iterating autonomously, build speed matters even more. These tools thrive on tight feedback loops: generate code, validate, iterate. A 20-minute build doesn’t just break a developer’s concentration; it stalls the entire human-AI collaboration cycle. Fast builds mean faster feedback for both the developer and the tools working alongside them.</p><p>It gets worse. We repeat this on feature branches and then do it all over again on merge to master, often with additional checks that may or may not pass. We want to iterate quickly. We deploy around 200 times a day. In November 2025 alone, that amounted to roughly 284 hours per day of CI compute time spent just running builds at loveholidays.</p><p>This isn’t just an annoyance; it directly affects how fast we can deliver value. In <em>Accelerate</em> (2018), Forsgren, Humble, and Kim identified Lead Time for Changes, the time from code committed to code running in production, as one of four key metrics that distinguish high-performing engineering teams. Build time is a significant contributor. The Extreme Programming community recognised this early on, advocating for a “ten-minute build” on the basis that anything longer discourages frequent integration and breaks the feedback loop.<a href="https://martinfowler.com/articles/continuousIntegration.html#build-fast"> Martin Fowler’s guidance on Continuous Integration</a> reinforces the point: every minute shaved off a build is a minute saved for every developer, on every commit.</p><p>At loveholidays, we care a lot about speed. We’re proud to have the<a href="https://tech.loveholidays.com/staying-ahead-of-the-curve-on-web-performance-7ace5663ff12"> fastest website in travel</a>, and we want our delivery process to match. To keep shipping features quickly and stay ahead of our competitors, we knew our CI had to get dramatically faster.</p><p>So over the last year, we set out to transform our CI performance. This is the story of what we changed, what worked, what didn’t, and what we learned along the way.</p><h3><strong>Our CI History with Google Cloud Build</strong></h3><p>Historically at loveholidays we’ve used Google Cloud Build for CI. It’s been a reliable workhorse, running millions of build minutes for us with barely a blip in uptime. But reliability alone wasn’t enough, and over time its limitations became increasingly painful.</p><p>The hardware was slow and dated. Start-up times were sluggish. You were capped at 30 concurrent builds per project. Machine sizes came in awkward increments thanks to a 1:1 CPU to RAM ratio, which meant paying for CPUs we didn’t need just to get enough RAM for browser tests. Everything was possible in theory, but even simple things like Slack notifications required DIY effort. The documentation felt thin, debugging was hard, and the UI was clunky compared to more modern tools.</p><p>These constraints had a very measurable cost. At their worst, our slow builds were costing around $7 per commit on our core customer-facing application. Multiply that across a team deploying 200 times a day and the numbers add up fast.</p><p>Some teams were already feeling this friction and had started to experiment with GitHub Actions.</p><h3><strong>Measure All the Things</strong></h3><p>Before committing to a large-scale migration, we needed to be confident that any changes we made were actually making things better. We had plenty of CI data from Cloud Build, but as some repositories moved to GitHub Actions, we lost a single source of truth. Comparing performance across systems became difficult.</p><p>So we set out to build a unified CI dataset (built on our<a href="https://tech.loveholidays.com/a-data-mesh-for-the-rest-of-us-12e2c10ac128"> Data Mesh</a>) that combines Cloud Build and GitHub Actions into a single source of truth. It took several iterations, but we eventually had reliable, normalised data flowing into one place, which we used to build dashboards that gave us a clear picture of our CI performance end to end.</p><p>Once we could actually see what was happening, problems that had previously been invisible started jumping out:</p><ul><li>Why is this job sitting in a queue for five minutes before it even starts?</li><li>Why are retries causing massive spikes in total build time?</li><li>Why are our end-to-end tests both slow and frequently failing?</li></ul><p>Armed with this visibility, we began systematically investigating issues in three areas: our data, our workflows, and even specific commits.</p><h3><strong>Targeting the Worst Offenders</strong></h3><p>We started with the slowest and highest-volume pipelines. These were typically front-end repositories, where we run a lot of tests: unit, integration, end-to-end, performance, accessibility, visual regression… the list goes on.</p><p>One of the first things we focused on was flakiness. CI should be reliable and only fail when there’s a genuine problem. Every flaky failure forces an engineer to retry, adds noise to the signal, and wastes time. Measuring flakiness is harder than it sounds, but once we had visibility into it, our visual regression tests stood out as the primary culprit. We made a series of targeted improvements to make them more deterministic and resilient, which significantly reduced false failures across the board.</p><h3><strong>Migrating to GitHub Actions</strong></h3><p>The writing was on the wall. With our data now telling us exactly where the pain was, we committed to a full migration to GitHub Actions.</p><p>We started with GitHub-hosted runners, with a plan to move to self-hosted runners once we had baseline data for comparison. When we did move, we chose self-hosted runners on Google’s C4A Arm-based machines, which reduced build times by an impressive 50%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IK-5-fbHCkmuDpGx8G271w.png" /><figcaption>Reduction of build times by over 50% from mid-July</figcaption></figure><h3><strong>Caching All the Things</strong></h3><p>With GitHub Actions, we had access to a much richer ecosystem of caching options than Cloud Build offered. We leaned into this heavily: caching dependencies, build artifacts, and anything else that could safely be reused between runs.</p><p>One tip worth sharing: don’t overlook your <a href="https://github.blog/changelog/2025-11-20-github-actions-cache-size-can-now-exceed-10-gb-per-repository/">cache size limits.</a> We recently increased cache allocations for some of our repositories and saw a meaningful improvement in build times. The cost of larger caches is negligible and quickly pays for itself in faster builds. It’s a small configuration change that’s easy to miss but well worth checking.</p><h3><strong>Challenges</strong></h3><p>Not everything was straightforward. Along the way, we hit a few interesting problems.</p><p>That move to self-hosted ARM runners wasn’t without friction. ARC version mismatches caused unexpected build queues, and moving to ARM meant multi-architecture builds for some applications, which risked doubling build times. In practice, the impact varied: Java builds handled multi-arch nearly for free, while Go required compiling separate binaries per architecture rather than relying on slower in-docker emulation.</p><p>A common assumption is that more parallelism is always better, but GitHub Actions doesn’t support running steps simultaneously within a job. We worked around this with multiple jobs, but at one point we were parallelising so aggressively that it introduced flakiness: too many jobs running at once would max out CPUs, trigger throttling, and cause intermittent test failures. We actually saw better reliability by pacing the work and being more deliberate about what truly needed to run in parallel.</p><p>One of our repositories uses Turborepo, an incremental build tool for monorepos. In GitHub Actions this shows up as a single build step, but under the hood Turbo is orchestrating many smaller tasks. That made debugging performance tricky. We’ve started uploading Turbo build log summaries and using a<a href="https://github.com/NullVoxPopuli/turborepo-summary-analyzer"> visualisation tool</a> to inspect what’s really happening inside that step, but it’s still an area we’re improving.</p><p>Finally, watch your build logs. When builds output thousands of lines of noise, it becomes harder to spot what actually went wrong. It also slows down your browser when inspecting failures and adds a small but cumulative performance cost across thousands of builds. Log what matters. Log when processing starts, when it finishes, and when something fails.</p><h3><strong>The Results</strong></h3><p>The cumulative impact was significant. Our core front-end repositories went from 20-minute builds down to around 8 minutes. Across the board, the migration to GitHub Actions with self-hosted ARM runners cut build times by over 50%.</p><p>The full picture shows up in our<a href="https://download.loveholidays.com/engineering/wrapped/lh-engineering-wrapped-2025-public.html#builds"> 2025 Engineering Wrapped:</a> across 567,891 builds over the year, average build time dropped from 8.16 minutes to 2.74 minutes, median build time went from 7.12 minutes to 21 seconds, and our success rate climbed from 80.2% to 90.6%. We ran 4.4x more builds than the previous year, and they were 3x faster.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lQSEW37HD9iukAbyDCx4qA.png" /><figcaption>Average build times during the migration from Cloud Build to GitHub Actions</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qaWaqdXbAuiLafqVIFGerg.png" /><figcaption>Build time improvements on core front-end repositories</figcaption></figure><h3><strong>What’s Next</strong></h3><p>There are still plenty of repositories we haven’t optimised yet, but we now have patterns we can reuse: shared caching strategies, sensible parallelisation defaults, and much better visibility into what our pipelines are actually doing. Some builds will always be harder to make fast, things like real-device testing of our app or running a full end-to-end suite. We’re also thinking about how much of the feedback loop can shift even further left. We already run linting, code health checks, and other validations locally before they hit CI on many of our projects. The more we can move into the developer’s environment and into AI tooling and make it frictionless, the tighter the loop becomes.</p><p>We’re starting to explore how AI affects CI quality. By combining our CI data with AI adoption data, we’re tracking AI vs human CI failure rates over time. Early signs are that LLM-generated changes are getting it right the first time more often as we improve prompts, guardrails, and review practices. But faster builds also mean faster iteration, and we’re seeing AI-assisted development drive more experimentation, which naturally comes with more failures too. We started this post by arguing that fast builds matter more in an AI-assisted world. They give AI the tight feedback loop it needs to fail quickly, learn, and try again.</p><p><em>Thanks to David Kasakaitis, Dmitri Lerko and Stuart Caborn for their contributions to this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=276e445f539e" width="1" height="1" alt=""><hr><p><a href="https://tech.loveholidays.com/build-baby-build-276e445f539e">Build, Baby, Build.</a> was originally published in <a href="https://tech.loveholidays.com/">loveholidays tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Staying ahead of the curve on web performance]]></title>
            <link>https://tech.loveholidays.com/staying-ahead-of-the-curve-on-web-performance-7ace5663ff12?source=rss-eabdb7be36c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ace5663ff12</guid>
            <category><![CDATA[web-performance]]></category>
            <category><![CDATA[core-web-vitals]]></category>
            <dc:creator><![CDATA[Jamie Thirlwell]]></dc:creator>
            <pubDate>Wed, 16 Apr 2025 14:11:07 GMT</pubDate>
            <atom:updated>2025-04-22T08:12:53.026Z</atom:updated>
            <content:encoded><![CDATA[<p><em>By </em><a href="https://www.linkedin.com/in/jamie-thirlwell-52726b153/"><em>Jamie Thirlwell</em></a><em>, </em><a href="https://uk.linkedin.com/in/ben-girvan-1a16831b3"><em>Ben Girvan</em></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hHDIaa8qPbEhLl7X" /></figure><p>At loveholidays, we know that delivering a fast, seamless customer experience is crucial — not just for increasing conversion, but for building long-term customer loyalty. <a href="https://tech.loveholidays.com/does-website-speed-contribute-to-customers-booking-holidays-online-f17bd696bc30">Website speed is a key factor in driving customer bookings</a>, so we continually benchmark our performance against other leading holiday providers — and even Google. This gives us insights into industry best practices and highlights areas for improvement — ultimately helping us boost customer retention, optimise conversion rates, and grow revenue.</p><p>In this blog post, we share how we monitor competitor website speed trends to stay ahead of the curve.</p><h3>How we’re doing it</h3><h4>Core Web Vitals via CrUX API</h4><p>Core Web Vitals are a set of metrics that measure user experience on the web, including loading performance, interactivity, and visual stability. These metrics are crucial for SEO and user satisfaction.</p><p>Google provides a <a href="https://developer.chrome.com/docs/crux/api">public dataset</a> of these metrics via the Chrome User Experience Report (CrUX), collected from millions of real Chrome users. Using this, we can query Core Web Vitals for competitor websites. You can specify a particular URL or aggregate data across an entire domain, returning key metrics such as:</p><ul><li><strong>Largest Contentful Paint (LCP)</strong>: How quickly the main content appears</li><li><strong>Cumulative Layout Shift (CLS)</strong>: How stable the page is as it loads</li><li><strong>Interaction to Next Paint (INP)</strong>: How responsive the page feels when users interact with it</li></ul><p>These metrics aren’t just technical vanity metrics — they directly impact how users perceive our site and whether they complete their booking or bounce to a competitor.</p><p><strong>Top tip:</strong> As an alternative to the API, you can explore <a href="https://developer.chrome.com/docs/crux/dashboard">Crux dashboards in Looker Studio</a> or the <a href="https://cruxvis.withgoogle.com/#/">Crux Vis</a> tool.</p><h3>Creating an automated data pipeline</h3><p>Instead of manually checking performance each month, we built an automated system to collect, process, and visualise this data:</p><ol><li><strong>Data collection</strong>: We use Airbyte to fetch weekly data from the CrUX API for our site and key competitors</li><li><strong>Data processing</strong>: dbt transforms this raw data into usable formats within our Data Mesh — refined subsets of raw data that serve as the foundation for dashboards and deeper analysis in BigQuery (<a href="https://tech.loveholidays.com/a-data-mesh-for-the-rest-of-us-12e2c10ac128">something we’ve previously blogged about</a>)</li><li><strong>Visualisation</strong>: Dashboards in Looker make the performance trends accessible across the team</li></ol><p>This automation ensures we have consistent, reliable data that we can trust when making decisions about performance optimisations.</p><p>End to end, our pipeline looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YDdRXZRFRjWPgGlt" /></figure><h3>What the data shows</h3><p>Our dashboards highlight monthly performance across each key metric:</p><h4>Average Largest Contentful Paint (LCP)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fEmui_lpbFIuJC6k" /><figcaption><em>Lower is better: loveholidays consistently outperforms all competitors and Google’s “good” threshold of 2.5 seconds</em></figcaption></figure><p>While competitors like TUI, Jet2holidays, and EasyJet struggle to keep their LCP times under Google’s recommended 2.5 seconds, we’ve consistently stayed around 1.5 seconds — giving us a crucial competitive advantage during the booking process.</p><h4>Average Cumulative Layout Shift (CLS)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3LxJhpBpgAYR8YeH" /><figcaption><em>Lower is better: Our pages stay stable during loading, avoiding shifts that lead to misclicks. We’re beating Google!</em></figcaption></figure><h4>Average Interaction to Next Paint (INP)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yafsAESlpkOCOOe8" /><figcaption><em>Lower is better: Our site responds to user interactions in about half the time compared to competitors</em></figcaption></figure><h3>Challenges &amp; next steps</h3><p>Our holiday search experience is one of our biggest strengths. However, we’ve noticed that Largest Contentful Paint (LCP) data can be misleading on competitor sites that use loading screens — these often trigger the LCP measurement before the actual results appear.</p><p>To address this, we developed a visual comparison tool that records real search journeys across all major competitors:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*eP9QXgZCW_5ef2YueyWUGw.gif" /><figcaption>Searching for holidays to Italy on a simulated 4G mobile connection</figcaption></figure><h3>How this data drives our decisions</h3><p>Being the fastest site in travel isn’t just a technical milestone — it’s a business advantage. It shapes decisions across our organisation.</p><ol><li><strong>Weekly performance reviews</strong>: Our <a href="https://tech.loveholidays.com/how-loveholidays-is-solving-the-technical-excellence-paradox-933987956758">BOOST team</a> meets weekly to analyse trends and identify any unexpected regressions</li><li><strong>Investment prioritisation</strong>: When deciding where to invest development resources, performance impact is a key consideration</li><li><strong>Replatforming strategy</strong>: Our architectural decisions are guided by clear performance targets derived from our competitive analysis</li><li><strong>Celebrating success</strong>: When we make significant improvements, we share the results company-wide to reinforce the importance of performance</li></ol><h3>What you can do today</h3><p>Whether you’re in travel or another industry, here are three steps you can take to start building your own performance intelligence system:</p><ol><li><strong>Set up basic monitoring</strong>: Track Core Web Vitals for your website</li><li><strong>Benchmark against competitors</strong>: Choose 3–5 key competitors and compare your performance regularly</li><li><strong>Connect metrics to business outcomes</strong>: Track how improvements in site speed correlate with conversion rates, bounce rates, and revenue</li></ol><p>By making performance a strategic focus with clear visibility, you can unlock the same competitive advantage we’ve built at loveholidays.</p><p><em>Thanks to </em><strong><em>Ben Girvan</em></strong><em>, </em><strong><em>Igor Naumov</em></strong><em>, and our </em><strong><em>Web Infrastructure Team</em></strong><em> for their brilliant work developing the competitors data product. And to </em><strong><em>Luke Williams</em></strong><em>, </em><strong><em>Dmitri Lerko</em></strong><em>, </em><strong><em>George Malamidis</em></strong><em>, </em><strong><em>Misha Griffiths</em></strong><em>, </em><strong><em>Mike Jones</em></strong><em>, and others for their helpful feedback and review of this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ace5663ff12" width="1" height="1" alt=""><hr><p><a href="https://tech.loveholidays.com/staying-ahead-of-the-curve-on-web-performance-7ace5663ff12">Staying ahead of the curve on web performance</a> was originally published in <a href="https://tech.loveholidays.com/">loveholidays tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>