Repository navigation
perf(trace): post-commit starts three git processes for each committed file #3756
Description
Problem
The post-commit hook of Chainloop Trace starts 3 git processes for each committed file that has pending AI lines. A large commit makes the hook slow.
settleCommittedLines (app/cli/pkg/action/trace_hook_handler.go) calls client.FileAt(repoRoot, sha, path) one time for each pending path in each session. ExecClient.FileAt (app/cli/internal/trace/git/lines.go) runs these commands for each call:
git rev-parse --verify --quiet <sha>^{commit}git cat-file -e <sha>:<path>git cat-file blob <sha>:<path>
The commit SHA is the same for all the files. Thus step 1 runs again for each file without a reason.
Measurement
Local clone of this repository (3.3k files). One commit changes 301 files, and all the changes are AI lines that are pending.
| Item | Value |
|---|---|
Git processes that post-commit starts |
More than 900 |
Time of post-commit |
Approximately 5.2 s |
This is on main. It is not caused by #3631.
Suggested fix
Read all the files with one git cat-file --batch process. Send one <sha>:<path> line for each path on stdin. For a path that is not in the commit, Git prints <sha>:<path> missing. Use this result as "no content", as FileAt does now. Resolve the commit one time before the loop.
For example, add a Client method that takes a revision and a list of paths and returns the content by path. The GoGitClient version can read the commit tree one time.
recordRootChanges (app/cli/pkg/action/trace_agent_hook.go) calls FileAt for each changed path in the same way, during PostToolUse for Bash calls. The same method can fix that caller too.
Activity