Project

Canary CI cache

A content-addressed build cache that cut CI time from 14 to 3 minutes without changing a single Dockerfile.

  • Go
  • CI/CD
  • S3

Source ← All projects

The problem

Fourteen-minute CI runs, 80% of them spent re-compiling unchanged inputs.
Classic solution is Docker layer caching, but our monorepo touched Go source
on every commit, blowing the cache from layer 2 onward.

Design

Content-addressed cache keyed on a merkle hash of inputs:

  • Input hashing: go.mod + go.sum + source tree fingerprint (git
    tree hash) → cache key. A commit that only edits docs reuses everything.
  • Storage: S3-compatible blob store; entries are gzipped tarballs of
    $GOCACHE and $GOMODCACHE.
  • Fallback: on miss, build as before and publish. Cache-aside, never
    cache-through, so a poisoned entry can be deleted without ceremony.

Results

Median pipeline time: 14 min → 3 min. Cache hit rate stabilized at 87%.
The interesting failure mode was flaky test pollution — cached test results
were never trusted; tests always run fresh, only compilation is cached.