Blog

Better guesses about available capacity meant fewer failed requests. Our jobs spent less time waiting, and we avoided costly moves to other regions.

Depot Metal builds runtime estimates from job fingerprints and history. They helped cut p99 host drain time from just over three hours to about one hour.

Depot built an S3-backed Ubuntu mirror and tuned apt and dpkg to cut median package installation time from 19 seconds to 3 seconds. Common setup work now takes less time.

Depot Metal's autoscaler grew from a simple pressure loop into a topology-aware system for compute and storage. Next, it must predict demand and converge on desired capacity.

We built a production-shaped load test to find control-plane limits before moving more GitHub Actions traffic to Depot Metal. Each run made our scaling changes measurable.

Depot CI runs GitHub Actions workflows without GitHub’s control plane, keeping CI online during outages and making workflows portable to new forges like Origin.

Taking the Sandbox TTI benchmarks one step further for fun.

Our microVMs launch in under a second, and assembling their block devices in parallel put LVM's roughly 100ms volume group-wide lock directly in the critical path. We kept the kernel's device-mapper, wrote our own control plane and gave up crash safety our ephemeral workloads never needed, making allocation roughly 100x faster.

Most of a large container's startup goes to pulling and decompressing files it never reads to boot. SOCI makes the image seekable so the runtime loads only the bytes it needs, and Depot writes that index during the build, so the image can lazy-load the instant it's pushed.

Formal verification never seemed worth it: writing a faithful TLA+ spec took too long. But now agents can write the spec from your code. We tried it on the rebuilt Depot Registry garbage collector, and the model checker found something our tests and reviews didn't: a race that could delete live data.