Uh oh!
There was an error while loading. Please reload this page.
[improvement](recycler) Avoid single-point read/write during sequentially reading key - #62476
Conversation
hello-stephen
commented
Apr 14, 2026
Thank you for your contribution to Apache Doris. Please clearly describe your PR:
|
wyxxxcat
commented
Apr 14, 2026
/review |
There was a problem hiding this comment.
Blocking issue found.
Goal of this PR
Reduce per-rowset point reads/writes while scanning recycler keys. The batching direction is reasonable, but the current implementation changes recycler semantics on a correctness-critical path and is not safe as written.Critical checkpoint conclusions
- Goal accomplished: Partially. Point operations are reduced, but the deferred abort path introduces a race that can delete rowset data for a transaction/job that commits before
loop_done()runs. - Change size/focus: Focused to recycler batching, but it changes behavior in both
recycle_rowsets()andrecycle_tmp_rowsets(). - Concurrency: Involved and currently unsafe. The recycler scan thread now queues abort work and only executes it at batch end, while concurrent
commit_rowset/commit_txn/finish_tablet_jobRPCs can still succeed during that window whenenable_mark_delete_rowset_before_recycle=false. - Lifecycle/static init: No special lifecycle or static initialization concerns found in this PR.
- Config changes: No new configs added, but an existing supported config combination regresses:
enable_abort_txn_and_job_for_delete_rowset_before_recycle=truewithenable_mark_delete_rowset_before_recycle=false. - Compatibility: No protocol/storage compatibility change observed.
- Parallel paths: The same regression exists in both formal recycle-rowset and tmp-rowset paths.
- Special conditions: The existing
end_version() != 1gate remains; no new explanation issues beyond the race above. - Test coverage: Existing tests mainly exercise the mark-before-delete flow and do not cover the interleaving where a commit/job-finish wins the race before deferred abort execution. I did not find new coverage for this regression.
- Observability: Logging is adequate for tracing the new path.
- Transaction/persistence: No new persistence format issue, but transaction/job state handling is where the correctness regression is introduced.
- Data writes/modifications: Not safe on the affected path because object deletion can proceed after a successful commit that happened before the deferred abort ran.
- FE/BE variable passing: Not applicable.
- Performance: The batching optimization is valid in principle.
- Other issues: No second independent blocker found beyond the deferred-abort race.
- Recycler-specific checkpoints
- Mark-before-delete two-phase flow: Preserved when mark-delete is enabled.
- Abort-before-delete aligned with origin: The mapping is still correct (load -> txn, compaction/schema-change -> job), but timing is no longer safe because abort is deferred.
- Packed files: Not affected by this PR.
- Conflict/retry/idempotency: The new deferred abort flow is not restart-safe enough on the affected config path because a concurrent commit can invalidate the recycler's stale decision before deletion submission.
Because this is a data-correctness issue on a supported code path, this should not be approved yet.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
23e5acb to
2b1057eCompare2b1057e to
1600c44Comparewyxxxcat
commented
Apr 22, 2026
/review |
There was a problem hiding this comment.
Requesting changes.
The batching direction is reasonable, but the current implementation does not preserve the recycler's safety contract.
Critical Checkpoints
- Goal: Batch point reads/writes while preserving two-phase recycle safety. Partially achieved. The point operations are reduced, but the new async batching delays
is_recycled/abort and changes error propagation, so the safety goal is not preserved. No new tests cover the new race/failure modes. - Scope/Minimality: The patch is focused on recycler rowset/tmp-rowset paths, but the helper refactor changes semantics, not just batching mechanics.
- Concurrency: Applicable. Mark/abort now run asynchronously in worker threads. I did not find a lock-order problem, but the new ordering opens a race between recycler scans and
commit_rowset/finish_job/commit_txn. - Lifecycle/Static Init: No lifecycle or static-initialization issue found in this PR.
- Configs: No new configuration items.
- Compatibility: No FE/BE protocol or storage-format compatibility issue found.
- Parallel Paths: Applicable. The same deferred logic was added to both
recycle_rowsets()andrecycle_tmp_rowsets(), and both inherit the same problems. - Special Checks: The new deferred checks do not establish why delaying the mark step is safe; in the current code it is not.
- Test Coverage: Insufficient for this change. No test covers page-sized scans, concurrent commit/finish while a rowset is only queued for marking, or batched mark failure such as
TXN_BYTES_TOO_LARGE/ conflict. - Observability: Logging exists, but the new worker-side failures are only logged and the outer recycle pass can still report success, so observability is not sufficient.
- Transaction/Persistence: Applicable. The new batched metadata rewrite path does not respect TxnKv's byte-limit assumptions, and mark/abort failures no longer propagate to the caller.
- Data Writes/Modifications: Applicable. Mark-before-delete is weakened because marking is deferred until the worker runs after
loop_done(). - FE/BE Variable Passing: Not applicable.
- Performance: It likely reduces point reads/writes, but the fixed-size batch can hit TxnKv byte/conflict limits and the async error handling can hide failed work.
- Other Issues: See inline comments for the concrete blocking cases.
Because of the above correctness regressions, I am requesting changes.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
1600c44 to
f024980Comparewyxxxcat
commented
Apr 23, 2026
run buildall |
f024980 to
019e763Comparewyxxxcat
commented
May 9, 2026
run buildall |
PR approved by at least one committer and no changes requested. |
hello-stephen
commented
May 9, 2026
Cloud UT Coverage ReportIncrement line coverage Increment coverage report
|
wyxxxcat
commented
May 11, 2026
/review |
There was a problem hiding this comment.
I am requesting changes based on the existing unresolved inline findings in this PR. I did not add duplicate inline comments because the current open threads already cover the blocking issues I found: deferred mark-before-delete reopens the commit race, deferred abort-before-delete can delete data that becomes committed, fixed-size batch marking can exceed TxnKv byte limits, and worker-side mark/abort failures are logged but no longer fail the recycle pass.
Critical checkpoint conclusions:
- Goal and tests: the goal is to reduce point read/write overhead during cloud recycler rowset cleanup while preserving recycle safety semantics. The current code does not fully preserve those semantics, and the PR does not add an automated regression/unit test proving the race-sensitive mark/abort paths.
- Scope: the change is focused on recycler batching, but it changes ordering and failure propagation in safety-critical paths rather than being a local performance-only optimization.
- Concurrency: this code is inherently concurrent with commit_rowset, commit_txn, finish_job/tablet_job, and recycler worker tasks. The existing threads identify windows where the scanner observes stale rowset state, defers mark/abort until loop_done/worker execution, and then deletes using stale eligibility. That violates the recycler's two-phase and abort-before-delete contracts.
- Lifecycle/static initialization: no new static/global lifecycle issue found.
- Configuration: existing config paths enable_mark_delete_rowset_before_recycle and enable_abort_txn_and_job_for_delete_rowset_before_recycle are affected. The supported disabled/enabled combinations are not preserved safely by the deferred flow.
- Compatibility/storage format: no new wire/storage format incompatibility found.
- Parallel paths: both recycle_rowsets and recycle_tmp_rowsets have the same class of deferred mark/abort failure; existing comments cover both paths.
- Conditional checks: the special base-version and PREPARE checks were reviewed; the main issue remains operation ordering and propagation rather than a missing local condition.
- Test coverage: no new automated test coverage was added for the changed recycler concurrency/failure behavior.
- Test results: no generated/modified test result files are part of this PR.
- Observability: added logs are not sufficient because worker failures are converted to log-and-return while the recycle pass can still report success.
- Transaction/persistence: the PR changes TxnKv transaction grouping and failure behavior; existing comments cover byte-limit risk and success reporting despite failed mark/abort transactions.
- Data writes/modifications: physical deletion can be scheduled after stale eligibility checks when mark/abort are delayed, which risks deleting data that concurrently becomes live.
- FE/BE variable passing: not applicable.
- Performance: batching may reduce per-rowset KV calls, but the optimization is not safe until ordering, byte-limit handling, and error propagation are fixed.
User focus: no additional user-provided review focus was specified.
Uh oh!
There was an error while loading. Please reload this page.
…ally reading key (#62476) fix: #58459 This PR reduces point-read overhead in Cloud recycler rowset cleanup. Previously, each scanned rowset could immediately trigger metadata reads/writes to mark it as recycled or abort its related transaction/job. The new flow records rowset keys during scanning, then batch-processes recycled marks and deferred abort tasks in worker batches. Prepare rowset deletion is also deferred so the recycler re-reads the latest metadata before deleting data. This keeps the existing recycle safety semantics while reducing per-rowset KV operations during large recycle scans. **Release mode test** Recycling 10,000 rowsets, after enabling `enable_mark_delete_rowset_before_recycle` and `enable_abort_txn_and_job_for_delete_rowset_before_recycle`, the processing time increased by approximately 10%. **3514 ms -> 175 ms(mark) + 3811(abort and recycle)**
…ally reading key (#62476) fix: #58459 This PR reduces point-read overhead in Cloud recycler rowset cleanup. Previously, each scanned rowset could immediately trigger metadata reads/writes to mark it as recycled or abort its related transaction/job. The new flow records rowset keys during scanning, then batch-processes recycled marks and deferred abort tasks in worker batches. Prepare rowset deletion is also deferred so the recycler re-reads the latest metadata before deleting data. This keeps the existing recycle safety semantics while reducing per-rowset KV operations during large recycle scans. **Release mode test** Recycling 10,000 rowsets, after enabling `enable_mark_delete_rowset_before_recycle` and `enable_abort_txn_and_job_for_delete_rowset_before_recycle`, the processing time increased by approximately 10%. **3514 ms -> 175 ms(mark) + 3811(abort and recycle)**
…ally reading key (apache#62476) fix: apache#58459 This PR reduces point-read overhead in Cloud recycler rowset cleanup. Previously, each scanned rowset could immediately trigger metadata reads/writes to mark it as recycled or abort its related transaction/job. The new flow records rowset keys during scanning, then batch-processes recycled marks and deferred abort tasks in worker batches. Prepare rowset deletion is also deferred so the recycler re-reads the latest metadata before deleting data. This keeps the existing recycle safety semantics while reducing per-rowset KV operations during large recycle scans. **Release mode test** Recycling 10,000 rowsets, after enabling `enable_mark_delete_rowset_before_recycle` and `enable_abort_txn_and_job_for_delete_rowset_before_recycle`, the processing time increased by approximately 10%. **3514 ms -> 175 ms(mark) + 3811(abort and recycle)**
What problem does this PR solve?
fix: #58459
This PR reduces point-read overhead in Cloud recycler rowset cleanup. Previously, each scanned rowset could immediately trigger metadata reads/writes to mark it as recycled or abort its related transaction/job. The new flow records rowset keys during scanning, then batch-processes recycled marks and deferred abort tasks in worker batches. Prepare rowset deletion is also deferred so the recycler re-reads the latest metadata before deleting data.
This keeps the existing recycle safety semantics while reducing per-rowset KV operations during large recycle scans.
Release mode test
Recycling 10,000 rowsets, after enabling
enable_mark_delete_rowset_before_recycleandenable_abort_txn_and_job_for_delete_rowset_before_recycle, the processing time increased by approximately 10%.3514 ms -> 175 ms(mark) + 3811(abort and recycle)
Issue Number: close #xxx
Related PR: #xxx
Problem Summary:
Release note
None
Check List (For Author)
Test
Behavior changed:
Does this need documentation?
Check List (For Reviewer who merge this PR)