| A collection is down or replicas are stuck recovering after a restart | The cluster state in the Collections API and Admin UI, then the ZooKeeper connection and the leader election in the node logs. A shard with no leader-eligible replica up cannot elect one, and PULL replicas are never eligible. | Bring an eligible replica back, or fix the ZooKeeper quorum |
|---|
| Long garbage-collection pauses or OutOfMemoryError on a node | Heap use against the filter and query result caches, and which fields facets and sorts run on. Fields without docValues can cost far more heap to facet or sort on, so one high-cardinality facet can dominate memory. | docValues, cache sizing, then heap and node count |
|---|
| Queries slowed after a reindex or an upgrade | debugQuery output for the same request before and after, then segment count and merge activity. A changed analysis chain, a new default or a cold cache each leave a different signature in the timing breakdown. | Analysis or request-handler change, or warming and merge settings |
|---|
| Indexing stalls or falls behind on a busy collection | autoCommit and autoSoftCommit settings against the update rate, merge pressure and disk I/O. Commits that fire too often create segments faster than merges can combine them. | Commit and merge settings, then batching in the client |
|---|
| Replicas serve different results or lag behind the leader | Which replica type each copy is, then replication and update-log state against the leader. NRT replicas index locally; TLOG and PULL replicas copy the index, so a lagging copy points at the leader, the network or the copy's disk. | Replica type and placement, or a rebuild of the stale replica |
|---|
| ZooKeeper session expired or connection loss in the Solr logs | ZooKeeper latency and the ensemble's member health, then garbage-collection pauses on the Solr node that outlast the client timeout. A long pause looks identical to a lost network link from ZooKeeper's side. | ZooKeeper capacity, or the Solr heap that caused the pause |
|---|
| Exports or crawlers get slower with every page | The request's start offset. Deep paging with a large start value makes every shard return and sort that many documents, so the cost grows with each page. | Cursor-based paging with cursorMark |
|---|
| A node refuses to start during a rolling upgrade | The node's version against the highest version already in the cluster. From Solr 9.10, SolrCloud nodes refuse to start if their major.minor version is lower than the cluster's highest, which blocks a rolling downgrade. | Finish the upgrade forward, or restore from backup on a planned window |
|---|