Limitations

Limitations #

Known limitations of the current release, and where to read more.

256 logical partitions per rolling are fixed — the partition count is a design constant, not a creation-time setting. What you scale is the number of physical shards (via the sharding strategy and online reshard).
Object storage and Parquet are optional, not the default — persistence is Pizza’s FIRE columnar format on local disk. S3-backed and Parquet-mount backends are opt-in engine features.

Cluster state is local-region only — GET /_cluster/state returns the serving region’s state; querying a remote region by UUID answers 501 not implemented yet (same for remote region health). The blocks and version state filters are accepted but are no-ops.
No backup/restore API yet — metadata is covered by Raft snapshots, but data backup is a manual data-directory procedure. See Backup and restore.
Collection-level replica changes apply to new rollings only — resizing the replicas of an existing rolling uses Update rolling settings; a rolling being resharded has its replica count frozen until the reshard commits.

SQL/PRQL are subset-shaped — POST /_sql and POST /_prql serve a documented subset (see the SQL & PRQL reference): no JOIN across collections yet (planned), no subqueries, CTE (WITH), window functions or set operations; PRQL s-strings are rejected outright and pipelines that would lower into a CTE fail closed with a positioned error.
Scripting is aggregation- and search-scoped — scripts run in pipeline aggregations, scripted_metric, as the value source of the single-value metric aggs, in a search request’s script_fields, and in the scripted queries (script_score, script filters, _script sort; Rhai, sandboxed, see the scripts reference). A script filter applies to the collected window (aggregations are not narrowed); per-segment script filtering is not implemented.
CDC is poll-based — the change feed is a pull API bounded by WAL retention; a consumer that falls behind purged_below must run a full resync.

Tenancy is namespace-scoped 1:1 — a key’s tenant maps to a single namespace; there is no multi-namespace tenant grouping yet.