SOLR-5754: StreamingSolrClients.getErrors() returns a snapshot copy instead of the live list

[nick-boss-tech]

🤖 AI text below 🤖 (posted on behalf of Nick Shanin)

https://issues.apache.org/jira/browse/SOLR-5754

What happens today

The retry pass reads the errors, then clears the list in a second call. An error added in between can be cleared without a retry.

Before this change, the retry pass copied the live error list and cleared it later in the same method. An error added between those two steps was cleared unseen. On the paths read, that window does not open, so this change is hardening, not a fix for a reproduced bug.

What this change does

Errors are read as a snapshot, and the retry pass takes them with one atomic drain.

• getErrors() returns a copy taken under the list's lock (StreamingSolrClients.java). This is a public behavior change. A caller now gets a snapshot, not the live list.
• drainErrors() copies and clears the list in one step under the same lock (StreamingSolrClients.java).
• The retry pass calls drainErrors() (SolrCmdDistributor.java). The later clearErrors() call is removed.
• The errors field is package-private and final, so the tests can seed it (StreamingSolrClients.java).

Proof

The combined run passed on a merged tree. There is no separate run without the fix for this branch; it was tested on the combined tree.

• Combined run, 2026-10-08, at head 7fbe0128d8b0: 26 of 26 tests pass, on a local merge commit of SOLR-5939 and this change, 1ddbf36202d, whose tree is f2e33340f0ea.
• Focused: TestTolerantUpdateProcessorCloud 19/19 and TestTolerantUpdateProcessorRandomCloud 2/2. The module checks pass.
• The current head 4bdf281d3cb3 removes only the root SOLR-5754-TESTING.md file, a note that is not part of the change; that removal was checked directly.
• StreamingSolrClientsTest has testGetErrorsIsASnapshot and testDrainErrorsReturnsAndClears. It passed 2 of 2 in the same combined run. It skips itself when Mockito cannot mock (L29-L32).

Limits

This is hardening. The race is not reproduced, and the tests are single-threaded.

• On the paths read, no runner thread adds an error during the retry window. This was checked by reading, not by a run.
• The tests check snapshot and drain behavior. They do not exercise the concurrent window. A follow-up submission is planned to test that window.
• clearErrors() stays public and has no caller in main code. This change does not remove it.
• SOLR-5939 changes the same two files. The textual merge is clean, and the proof above covers the merged tree.

Changelog: changelog/unreleased/SOLR-5754-streaming-clients-error-snapshot.yml (type: changed).

AI assistance

AI agents assisted with research, implementation, review, and drafting. Nick Shanin directed the work and takes responsibility for this contribution.