When we started measuring deploy time, the median was 73 seconds. After three weeks of work: 24 seconds. Here's what changed.
The first bottleneck was our build scheduler. We were using Nomad, which is great for long-running services, but the cold-start overhead for short builds was ~4 seconds. We wrote a bespoke Go scheduler that reuses warm workers and keeps a hot pool per runtime.
The second was dependency installation. npm install was taking 18 seconds on median for Node projects with ~800 packages. We now hash the lockfile and cache node_modules between deploys of the same project. Cache hits take 1.1 seconds.
The third was build output reuse. If your next.js build output hash matches the previous deploy's, we skip the build entirely and just re-upload. Rare edge case but huge win when it lands — zero-second builds.
Together: 73s → 24s, with p95 down from 140s to 41s. That 12-second difference matters — it's the difference between deploying six times before lunch and deploying twelve.