almostintuitive opened a new pull request, #4437:
URL: https://github.com/apache/iggy/pull/4437
## Which issue does this PR address?
Relates to #4436
## Rationale
New bugs should come with a failing test, so this PR adds only that test. The
fix belongs in the Rust SDK's send path, and is left for discussion.
## What changed?
When two partitions of a topic have their primaries on different nodes, a
producer that alternates between them is refused (`TransientNotAccepted`) on
every send. It waits out two 2s transient windows, then moves its only
session to
another node. Each send takes about 4s. Polls avoid this through
`PollRouter`.
Sends have no equivalent.
`given_partitions_on_different_primaries_when_sending_round_robin_should_keep_the_session`
builds on the existing `seed_split_primaries` fixture. Partition 0 keeps its
primary through a metadata-only election. Partition 1, created afterwards
with
`create_partitions`, is seeded onto the metadata leader. A precondition
checks
via `GET_POLL_ROUTING` that the two primaries are on different nodes. After
one
warm-up send to each partition, the test alternates four sends and asserts
that
none of them moves the session to another node. It is
`#[ignore = "#4436: ..."]`, so default CI stays green, and it runs with
`--ignored`.
## Local Execution
- Not passed: the new test fails as intended, 3 of 3 runs on `master` @
`43fb44364`, every time with
`4 of 4 sends moved it to another node ... [(0, 2, 0, 4.020694671s), (1,
0, 2, 4.018801233s), (0, 2, 0, 4.019829788s), (1, 0, 2, 4.018055753s)]`.
On tag `iggy-0.11.0` it fails the same way (about 4.07s per send).
`cargo fmt --all --check` and
`cargo clippy -p integration --all-features --tests -- -D warnings` are
clean.
- Pre-commit hooks: `prek run --files <changed>` ran. Everything that applies
passed (typos, whitespace, trailing newline, binary artifacts, skills)
except
`license-headers`, which could not run because `jq`/`hawkeye` are not
installed. The file already carries the ASF header and the diff does not
touch it. cargo fmt and clippy ran separately (see above).
- The last change before the push only filled in the issue number in the
`#[ignore]` reason. The file was checked with `rustfmt --check` after that
change, but the test was not rerun.
## AI Usage
1. Claude (Claude Code).
2. Investigation (reading the TCP client's refusal, routing-lock and
roster-walk
path, and comparing it with `PollRouter`), the test and the write-up.
3. The test was run against the 3-node harness cluster and fails as shown
above. The client log confirms the mechanism: about 4s of refusals, then
`trying the next cluster node` or `redirecting to leader` on every send.
The
fixture asserts its precondition, that the partition primaries are on
different nodes, before any send, so it cannot fail for an unrelated
reason.
4. Yes.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
--
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]
