fix: ES index upgrade never completes when several connectors share an alias - EXO-88232

[boubaker]

Problem

While upgrading notes_v1 → notes_v2 (Meeds-io/notes#1756), the upgrade failed with:

ERROR | An error occurred while upgrading index notes_v1
java.lang.NullPointerException: Cannot invoke "java.util.Set.size()" because the return value of "java.util.Map.get(Object)" is null
    at ...ElasticIndexingOperationProcessor$ReindexESType.run(ElasticIndexingOperationProcessor.java:792)

Notes declares two connectors on the same index_alias (WikiPageIndexingServiceConnector and NoteVersionLanguageIndexingServiceConnector). One ReindexESType task is submitted per connector. The first task to finish saw size() > 1, did nothing, then its finally removed the whole alias entry from indexUpgrading → the second task NPE'd. Net effect: the alias was never switched to the new index and the old index was never deleted, on every alias shared by ≥ 2 connectors.

Fix

• Track completion per connector (markConnectorUpgraded); switch the alias and delete the old index once, after the last connector of that alias finishes.
• On failure, release only the finished connector's slot; drop the alias entry when nothing is pending anymore.
• Send the ES-side _reindex once per alias/pipeline instead of once per connector (it was duplicated for notes).
• Logs name the connector; the pipeline-failure WARN now carries the exception.
• setExecutors(ExecutorService) / isIndexUpgrading(alias) added for tests.

Tests

Two new tests in ElasticOperationProcessorTest with two connectors on notes_alias, run on a same-thread executor:

• alias switch / old-index delete / _reindex each happen exactly once, after both connectors;
• after only the first connector finishes, the alias is untouched and the upgrade is still pending.

Mutation-verified: with the previous size() > 1 + whole-alias remove logic put back, both tests fail; with the fix, the full commons-search suite passes.

Classification

N2 — shared indexing engine (cross-domain seam), no ACL/schema/REST surface touched.

Knowledge: none — behaviour fix in the indexing engine, no domain-doc claim changes.

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[boubaker]

Second commit: a further test run on notes showed the main-thread side of the same problem — sendInitRequests ran the "upgrade interrupted" detection per connector, so the second connector of notes_alias found the notes_v2 just created by the first one, deleted it while the _reindex was running (404 index_not_found_exception) and recreated it empty; the alias then got switched to an empty index.

The new index is now (re)created once per alias (upgradeInitializedAliases), each connector still submits its own reindex task. New test start_twoConnectorsOnSameAlias_newIndexCreatedOnceAndNeverDeleted (mutation-verified: fails with the guard removed).