Most companies rent CI time from cloud providers and pay for idle capacity. We decided to own our infrastructure. The result: a 200-core bare metal machine running our full test suite in under 30 seconds, at roughly $1,000 per month — a fraction of what equivalent cloud capacity would cost.
The core insight: ZFS snapshot caching
The bottleneck in most CI systems isn't compute — it's setup time. Installing dependencies, seeding databases, warming caches. We solved this with ZFS filesystem snapshots. A known-good state is captured once, and each test shard restores from that snapshot in under 2 seconds. We get infinite test sharding without infinite setup overhead.
# Restore CI snapshot and run shard
zfs rollback ci-pool/test-base@ready
bun test --shard=${SHARD}/${TOTAL_SHARDS}Numbers that matter
- Test suites that took 10–15 minutes now run in 20–30 seconds across 12–15 shards
- ZFS snapshot restore: under 2 seconds from cold
- Infrastructure cost: ~$1,000/month for 200 dedicated cores
- Build time for the CI system itself: 1 day, written entirely by Autobuild
The entire system — infrastructure provisioning, shard orchestration, snapshot lifecycle management, and the monitoring dashboard — was built by Obvious's own AI in a single day. The CI engine is now the substrate that everything else in Autobuild runs on.