Loading
Draft: Extract ManualMergeDetector from RefreshService (reference for !264935)
Reference only - closed on creation, not for merge. It's !264935 (merged) with the manual merge detection moved out of RefreshService#post_merge_manually_merged into Gitlab::MergeRequests::ManualMergeDetector. Behaviour is unchanged, and with the flag off it is faster than master.
Built on !264935 (merged)'s head (bb5197fe). The last commit (c5f6bf1a) is the only new one.
Detailed context for AI agents
ManualMergeDetector#executereturns{merge_request => merge_commit_sha}. It owns the candidate query (moved fromRefreshService#post_merge_candidates), the empty-diff filter, the stored diff head check (stored_head_landed), the flagged source branch tip fallback (current_branch_tip_landed,detect_manual_merge_by_source_branch_tip) andBranchPushMergeCommitAnalyzer.RefreshService#post_merge_manually_mergedkeeps onlycommits_by, the merge source lookup and the per-MRPostMergeServiceloop, so it sits next to theMergeSourceResolverextraction in !260499 (merged). The two merge without conflicts.- Push commit ids are a
Set, and MRs with an empty or missing diff are dropped before their diff head is read, so there is noFindCommitcall for MRs without a diff. - No unit spec for the detector yet.
Verification
spec/services/merge_requests/refresh_service_spec.rb: 110 examples, 0 failures. Rubocop clean.- Break matrix, 25 scenarios x flag off/on (50 examples, 0 failures). Flag off is compared against master and flag on against
bb5197fe, each in a fresh project. MR states, merge SHAs, merged events, errors,ListRefspatterns and processing order all match, and it never makes more Gitaly calls. Scenarios: stored head vs branch tip, empty/no-diff/fork/orphan filters, non-ASCII branch, deleted source branch, mixed and batch pushes, closed/locked MRs, target create/delete/rewind pushes,PostMergeServiceandListRefsfailures.
Performance (test env, local Gitaly, 30k branches, PostMergeService stubbed, median ms of post_merge_manually_merged, push with no manual merge; relative numbers only)
| Open MRs, push size | master | !264935 (merged) flag off | !264935 (merged) flag on | this, flag off | this, flag on |
|---|---|---|---|---|---|
| 2,000 MRs, 29 commits | 312 | 314 | 477 | 258 | 457 |
| 2,000 MRs, 10k commits | 540 | 492 | 773 | 267 | 551 |
| 5,000 MRs, 29 commits | 1,102 | 1,058 | 1,927 | 616 | 1,383 |
| 5,000 MRs, 10k commits | 1,530 | 1,405 | 2,256 | 657 | 1,415 |
- Flag off is never slower than master. It is up to 2x faster on large pushes and makes 0
FindCommitcalls instead of 1 per MR with no diff. - Flag on adds one
ListRefscall with one pattern per unmatched open MR: about 40 ms at 267 patterns, 150-290 ms at 1,780 (gitlab-org/gitlab scale: 1,987 open MRs target master, 29.6k branches) and 600-900 ms at 4,450, on every push to the target branch.
Edited by Marc Shaw