The ScyllaDB team is pleased to announce the release of ScyllaDB 2026.3.6, a production-ready patch release for ScyllaDB 2026.3 Short Term Support (STS) Minor Feature Release.
More information on ScyllaDB Version Support Policy is available here.
Related Links
Bug Fixes
The following issues are fixed in this release.
Guardrails
-
The
write_consistency_levels_disallowedguardrail also applied to internal writes, such as those that persist group 0 (schema and topology) state, compaction history, and tracing data. As a result, disallowing consistency levelONEcould block schema changes and node startup, and disallowingANYstopped tracing from recording data. The guardrail, and the related per-consistency-level write metric, now apply only to client writes, so the setting can safely be used to restrict client consistency levels.
scylladb#32485 -
On tables with
large_data_guardrails_enabled = true, a node could restart during startup, after anALTER TABLE, or after an off-strategy compaction or tablet split. Tables that use the large data guardrail now remain stable across restarts, schema changes, and compaction.
scylladb#32825
Materialized Views (MV)
-
On tables with
large_data_guardrails_enabled = true, a write rejected by the large data guardrail could still update the table’s materialized views and secondary indexes, leaving them out of sync with the base table (for example, a view row without a base row, or a missing index entry). The guardrail now checks the write before the views are updated, so a rejected write leaves the base table, its views, and its indexes unchanged.
scylladb#32822 -
While a materialized view was being built, the node could stop applying schema and topology changes while it waited for streaming into the base table (such as repair, tablet migration, or rebuild) to finish, which in some cases could lead to a deadlock. Schema and topology changes are now applied normally while such streaming is in progress.
scylladb#32508
Monitoring
- Tablet repair logged several per-shard and per-request messages at the info level, which could flood the system journal and cause other log messages to be suppressed. These messages are now logged at the debug level, while messages that report an actual hints or batchlog flush stay at the info level, reducing log volume during tablet repair without losing relevant information.
scylladb#32924
Performance
- On every topology change, including each tablet migration stage, a sanity check read the entire current CDC generation, which caused reactor stalls on large clusters. The check now reads a single row, removing these stalls during tablet migrations and other topology changes.
scylladb#32886
Repair
- With vnodes, or with tablets that were not migrated, the records that repair and tablet cleanup save to
system.commitlog_cleanupswere not removed until the node restarted, so the table could keep growing, and removing the records at startup could stall the node. Obsolete records are now removed as new ones are saved, which keeps the table small and improves node startup time after repairs.
scylladb#32823