0AyanamiRei opened a new pull request, #67137:
URL: https://github.com/apache/doris/pull/67137
### What problem does this PR solve?
Issue Number: None
Related PR: None
Problem Summary:
Nereids INSERT INTO SELECT and CTAS expose two query cancellation races:
1. The insert coordinator is created before transaction initialization, but
is published to StmtExecutor only after beginTransaction and sink finalization.
A KILL QUERY, JDBC Statement.cancel(), or REST kill request received before
publication finds the connection but no coordinator. The old path reports
success and loses the cancellation, so the coordinator can later dispatch
fragments and commit.
2. After coordinator.join() reports success, there is a window before
onComplete() starts transaction commit. A cancellation in this window only
targets an already completed coordinator, reports success, and does not prevent
commit.
The same false-success behavior exists in local REST cancellation and
follower-to-master forwarding because those layers do not propagate whether the
statement accepted cancellation.
This PR:
- remembers cancellation in StmtExecutor and replays it when an INSERT or
CTAS coordinator is published;
- checks coordinator cancellation while holding the dispatch lock before
sending fragments to BEs;
- establishes a linearization point between cancellation and transaction
commit: cancellation wins and aborts the transaction, or commit wins and the
cancellation request is rejected;
- propagates the accepted or rejected result through ConnectContext, local
schedulers, SQL KILL, JDBC cancellation, REST kill, Arrow Flight SQL, and
forwarded FE cancellation;
- adds deterministic unit and regression coverage for cancellation before
coordinator publication, cancellation before commit, and CTAS cancellation.
Scope: CANCEL LOAD is not changed by this PR. Running InsertLoadJob
transaction/query binding is an independent load-job lifecycle issue.
### Release note
KILL QUERY, JDBC Statement.cancel(), and the REST query kill API now
reliably cancel Nereids INSERT INTO SELECT and CTAS before transaction commit,
including coordinator-publication races. Cancellation attempted after commit
begins returns an error instead of a false success.
### Check List (For Author)
- Test
- [x] Regression test
- ./run-regression-test.sh --run -d insert_p0 -s
test_cancel_nereids_insert_query
- [x] Unit Test
- ./run-fe-ut.sh --run org.apache.doris.qe.StmtExecutorTest
- [x] Manual test
- Built FE with ./build.sh --fe -j48.
- Verified the regression suite against FE version
doris-0.0.0-1731787677f on query port 24030 and HTTP port 23030.
- Behavior changed:
- [x] Yes. Query cancellation is sticky until coordinator publication,
cancellation is serialized against INSERT commit, and cancellation that arrives
after commit begins returns failure.
- Does this need documentation?
- [x] No.
### Check List (For Reviewer who merge this PR)
- [ ] Confirm the release note
- [ ] Confirm test cases
- [ ] Confirm document
- [ ] Add branch pick label
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
