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.
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.
Agents can investigate, plan, and implement faster than I can, but reading a final diff is not the same as understanding why the system works the way it does. I own how the system behaves in production, so I want to keep building my intuition about it.
Report JUnit from your Depot CI or Depot GitHub Actions jobs to see failures in context. A single command runs your suite and uploads the report. Split suites by historical timing and they'll automatically rebalance over time.
To make recurring test failures easier to reason about we needed to include historical context. Here's how we derive a stable identity for each test and a fingerprint for each failure, so you can tell a brand-new problem from the same flaky test you've seen a dozen times.
Most CI jobs run as one long serial script even when half the work doesn't depend on the other half. You can run independent steps in parallel for faster CI. We'll compare two ways to do it: shell backgrounding and Depot CI's parallel steps.
Depot CI's orchestrator runs on AWS Lambda durable functions. The orchestrator checkpoints state, suspends between events, and never burns compute while waiting. Here's how we designed the run-workflow-job hierarchy and the callback-driven orchestration loop that powers it.
AWS Lambda durable functions let you write long-running processes that survive crashes and suspend compute entirely while waiting for external events. Getting them right requires understanding how replay works and the distinct failure modes of durable functions. Here's what we've learned from running durable functions for real.
We built a fun Spotify Wrapped-style recap for Depot customers. Handling a year of data and making big numbers relatable was easy compared to teaching AI that "0.1 of a marathon" isn't exactly celebratory.
When upstream systems fail, retries pile up fast and your database becomes the bottleneck. Here's how we used PlanetScale's Insights to identify and fix the queries that were slowing us down during thundering herd events.
Experiments with using Claude in Depot remote sandboxes to identify documentation gaps, generate drafts, and collaborate with your team on keeping docs up to date.
AI has become incredibly powerful for engineering work, but treating it as an autopilot without providing a ton of context can lead to messy, unmaintainable code. Here's how I use AI as a copilot instead to accelerate my work while keeping control over architecture, patterns, and code quality.
