Reduces the CI cache growth and disk pressure behind the flaky `No space
left on device` failures in
https://github.com/go-gitea/gitea/issues/37974.
**`go-cache`** — the cache-seeder saved with a `restore-keys` prefix
fallback, so every `go.sum` change restored the previous cache and
re-saved the union; old module versions and stale build objects
accumulated (~3 GB → ~7 GB) and overflowed disk on smaller runners. Drop
`restore-keys` from the seeder **save** branches so each `go.sum` seeds
a clean, size-bounded cache. PR runs keep `restore-keys` for warm-start
fallback.
**`free-disk-space`** — delete the unused preinstalled toolchains in
parallel (~86 s → ~54 s) and log `df -h /` before/after.
Measured during review: the hosted `ubuntu-latest` fleet is
heterogeneous — most runners have ~89 GB free on `/` (a full pgsql
integration shard peaks at ~17 GB used), but a minority arrive nearly
full and fail mid cache-restore. The toolchain deletion is the headroom
that keeps those runners green, so it stays; the cache bound shrinks the
footprint for every runner.
Authored with assistance from Claude (Opus 4.8).
---------
Signed-off-by: silverwind <me@silverwind.io>
Co-authored-by: bircni <bircni@icloud.com>
Critical path ~25:42 → ~19:56 (−22%), ~0% CI minutes.
- `test-pgsql` shards 2-way. Branch protection: replace `test-pgsql`
with `test-pgsql-shards (1)` + `test-pgsql-shards (2)`; `test-unit`,
sqlite/mysql/mssql unchanged — pgsql dominates the critical path.
- `test-unit` runs `bindata` then `bindata gogit` sequentially.
cache-seeder pre-warms the race-instrumented test compile cache and the
integration test binary so PR jobs warm-start.
- Cache writes restricted to cache-seeder; PR jobs use
`actions/cache/restore`. Defends against PR cache poisoning and frees
the 10 GB cap from PR churn.
- `go-cache` action: dropped the `cache-name` input. One gobuild cache,
one golangci-lint cache. Seeder lint job restores but doesn't save
gobuild, so only one writer populates it.
- `tools/test-integration.sh` shards the integration binary via
`-test.list`; `TestMain` short-circuits DB init in list mode.
`TestAPILFSNotStarted` / `TestAPILFSLocksNotStarted` switched to
`test.MockVariableValue` — latent `setting.LFS.StartServer` global-state
leak uncovered by sharding.
---
This PR was written with the help of Claude Opus 4.7
---------
Signed-off-by: silverwind <me@silverwind.io>
Co-authored-by: Claude (Opus 4.7) <noreply@anthropic.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: Nicolas <bircni@icloud.com>
Cache includes go, lint and unittests. Integration tests with their
standalone binaries are uncacheable with their current architecture.
Every Go job uses a new composite action (`.github/actions/go-cache`)
that restores and saves the Go module cache, a shared build cache, and
the golangci-lint cache. A `cache-seeder` workflow runs on `push: main`
to pre-populate those slots; PRs read them via GitHub's default-branch
fallback, so the common case is warm from the first commit.
Also dropped `-coverprofile` from `test-unit` (it silently disabled Go's
test result cache), and `-race` from `test-pgsql` and `test-mysql` (kept
on `test-unit` and `test-sqlite`).
Signed-off-by: silverwind <me@silverwind.io>
Co-authored-by: Claude (Opus 4.7) <noreply@anthropic.com>
Co-authored-by: Nicolas <bircni@icloud.com>
