Depot

We reimplemented LVM for our microVM hypervisors, with worse guarantees on purpose. LVM has been rock-solid for two decades, but its volume-group-wide lock is held ~100ms per operation. That's invisible at one op a week and brutal when you launch up to 200 microVMs a second. So we kept the kernel's device-mapper, wrote our own control plane, dropped crash safety our ephemeral workloads never needed, and made allocation roughly 100x faster. https://lnkd.in/gTydRxmJ

Depot Metal is live: our next-generation compute and storage platform, already running Depot CI and Sandboxes up to 50% faster with nothing to configure and no added cost. July also brought Depot Code, our own diskless git hosting, into private beta, plus general availability for test results with org-wide flaky and slow-test analytics. Depot CI picked up Tailscale support, Datadog CI Visibility (private beta), three new triggers, GitLab OIDC trust relationships, and always-on Sherlock analysis on any workflow or job.

Noisy test failures burn hours of engineering time. People reread the same logs, retry jobs to see if a failure disappears, and slowly stop trusting failures that may be pointing at something worth fixing. The way out is a stable way to identify each failure and compare it against its own history, so a brand-new problem stops looking like the same flaky test you've seen a dozen times. https://lnkd.in/gXZhMdbJ

An agent can hand you a fix that works and passes its tests but could still be the wrong call for your system. You only catch that if you've kept enough context of your own to challenge it. Iris on where she stays involved so she keeps that context: planning decisions, close review, and a cap on parallel work. https://lnkd.in/gWagKNve

We're now the only container build service to offer SOCI 🚀 You can build a huge image in seconds and still sit there while every node pulls and decompresses gigabytes before anything starts. Most of that startup time is spent reading files your container never even touches on boot. SOCI fixes that by making the image seekable. The runtime gets an index of where every file lives so it can fetch just the bytes it needs to start, then lazy-load the rest in the background. The awkward bit used to be where that index came from. You had to bolt on a second pipeline to pull the image back down, index it, and push the metadata. So we added SOCI support to our container build engine and walked through how we did it: - Build the normal gzip layers you already expect - Generate the SOCI v2 index (zTOC) from those compressed layers - Push the image and SOCI metadata together so the moment it hits the registry it is ready to lazy-load Turn on `soci=true` in your Depot build and the fast startup ships with the image. No post-push indexer. No extra pipeline to keep warm. Full post here: https://lnkd.in/ev9drE3w

When AI-written code breaks in production, you're the one on call, not the agent. Taking that seriously means making AI-generated code maintainable: linter guardrails, reference tests, tight AGENTS.md files, and small, reviewable PRs. A practical look at staying in control of what gets merged. https://lnkd.in/gNjMFKgN

It's been a pretty wild day at Depot. I spent the day talking with teams affected by the GitHub Actions outage, helping them get back up and running while answering questions along the way. If you're curious about what we've been seeing or want to chat about your CI setup, reach out!

Cloud-native worked because we built on a few composable primitives: compute became VMs, storage became object stores, networking became load balancers. Software delivery still hasn't had that shift. Kyle names six infrastructure primitives to build on instead of bending a human-centric collaboration model to machine-scale output. https://lnkd.in/ggtEvjtX

Test results are now generally available in Depot! Point it at your JUnit XML, and pass/fail/error/skip counts will show up on the CI job that produced them. Use test results from Depot CI or Depot GitHub Actions runners to get reporting, analytics, timing-based splitting, and a single command that runs your suite and uploads the report. https://lnkd.in/g-4bJYVE

CI logs are a good starting point, but they're noisy and don't give you data you can filter or join across runs. Structured test results give you data you can reason about over time: which tests fail most often, which are slow enough to drag the whole suite, and what's getting worse across the org. That's now generally available in Depot, from JUnit XML on any Depot CI or GitHub Actions job. https://lnkd.in/dMtdKz25