Sriharsha Chintalapani 5620121e50 SearchIndex: tunable index settings + per-stage latency metrics (#27865) ... * SearchIndex: configurable index settings + per-stage latency metrics Adds two diagnostic and operational improvements to the distributed search indexing pipeline so operators can both tune cluster behavior per installation and pinpoint where reindex latency is being spent. Configurable index settings (per-installation, no code changes needed) - New SearchIndexing app config fields: liveIndexSettings (post-promote), bulkIndexSettings (during reindex), and per-entity overrides. - DefaultRecreateHandler applies bulk overrides on staged-index creation (e.g. refresh=-1, replicas=0, async translog) and reverts to live values before alias swap. Optional force-merge before swap. - Safety revert ensures the promoted index never inherits a disabled refresh interval, even if the admin only configured bulk overrides. - Live UX is preserved: refresh defaults to 1s so users and agents that read-after-write see near-real-time results. - New IndexManagementClient methods (updateIndexSettings, forceMerge) with implementations for OpenSearch and Elasticsearch. Per-stage latency metrics (consumer-vs-producer attribution) - StageStatsTracker accumulates per-stage wall-clock time alongside existing counters; added timing-only addStageTime() so per-record callbacks and per-batch wall-clock don't double-count. - DB migration 1.13.0 adds readerTimeMs / processTimeMs / sinkTimeMs / vectorTimeMs columns to search_index_server_stats. Existing rows get DEFAULT 0; aggregation queries SUM the new columns. - Reader timing wraps PartitionWorker.readEntitiesKeyset (DB latency). Process timing wraps the doc-build join in OpenSearch and Elasticsearch bulk sinks (CPU/serialization). Sink timing wraps client.indices().bulk (pure search-cluster latency), attributed per participating tracker. - DistributedJobStatsAggregator surfaces totalTimeMs on each StepStats so the UI can compute avg latency = totalTimeMs / successRecords and throughput = successRecords / (totalTimeMs / 1000) on every WebSocket push without server-side derivation. - New per-server aggregation query (getStatsByServer) for distributed visibility, fed into SearchIndexJob.ServerStats with timing fields. UI: each of the four stage cards (Reader / Process / Sink / Vector) shows "Latency: X ms · Y r/s" when timing is available; per-entity table gains Sink avg + Sink throughput columns. Docs panel updated. New SearchIndexing config section added with sane defaults that preserve current behavior. Tests: 6 new StageStatsTracker timing tests, new aggregator test that asserts StepStats.totalTimeMs is populated at job and per-entity level. All existing tests updated for new arg shapes; 60 unit tests pass. The pattern operators see: Reader avg climbing means DB-side issue (cache/autovacuum); Sink avg climbing means OS-side issue (segments/ back-pressure); only one entity's row climbing identifies the offender.		2026-05-02 20:11:06 -07:00
..
migrations	SearchIndex: tunable index settings + per-stage latency metrics (#27865)	2026-05-02 20:11:06 -07:00
schema	Fix Metrics collection; reduce no.of metrics; improve slow request lo… (#25751)	2026-03-13 13:38:31 -07:00