What scylladb/scylladb shipped

This week: logstor reports live record bytes per tablet, LWT no longer times out when a coordinator's clock lags, group 0 recovery with strongly consistent tables is fixed, and the large data guardrail no longer crashes on restart or ALTER TABLE.

Work merged this week in ScyllaDB: logstor tablets now report their live record bytes as size. A coordinator whose clock is behind another's can raise its next ballot above a rejecting one, so LWT on a touched key no longer times out until clocks catch up. Recovery from group 0 majority loss with strongly consistent tables is fixed and tested. system.commitlog_cleanups no longer stalls startup or grows without bound. The large data guardrail no longer crashes on restart, ALTER TABLE or after off-strategy compaction, and it now checks a write before pushing view updates. Repaired-set compactions consult a snapshot of the tombstone gc state, and hashing bytes and keys uses one-shot XXH3.

This week: periodic sstable scrub to catch silent corruption, hints no longer lost on decommission or when encryption sidecars are left behind, nodetool -H fixed, release builds define NDEBUG, and more strongly consistent write path work.

This week in ScyllaDB: compaction now periodically validates all sstables, so corruption hiding in idle data can be found. Hints are no longer lost when a decommissioning node is banned, and hint segments move together with their encryption sidecars. On the operational side, system.commitlog_cleanups no longer stalls startup or grows without bound, object-storage keyspaces no longer abort on shutdown, nodetool tablestats and compactionstats accept -H as a plain flag, and release builds now define NDEBUG. Strongly consistent tables saw further work on routing, error propagation and commitlog-backed raft persistence writes.

2026.4.0-rc0 is out. This week: vnodes-to-tablets migration covers materialized views and secondary indexes; DROP of an indexed column is rejected; MUTATION_FRAGMENTS() accepts token(); malformed sstables are quarantined instead of retried forever.

The first 2026.4.0 release candidate is tagged. Vnodes-to-tablets migration now handles keyspaces with materialized views and secondary indexes, giving them their own table maps and tablet counts. Dropping a column used by a vector, fulltext or pattern index is now rejected rather than leaving an index on a missing column. MUTATION_FRAGMENTS() accepts token() in selectors and restrictions. Bad INSERT JSON arguments return InvalidRequest instead of a server error, grouped queries stop at their LIMIT, and regular compactions quarantine malformed sstables instead of looping.

This week: Alternator DescribeTable reports WarmThroughput and GSI ProvisionedThroughput, STCS becomes an alias of ICS, LIKE queries served by a pattern index, Raft log bounded on followers, and node replacement during vnodes-to-tablets migration.

This week in ScyllaDB: Alternator's DescribeTable now reports WarmThroughput and GSI ProvisionedThroughput, addressing two gaps that blocked DynamoDB tooling. Size-Tiered Compaction Strategy becomes an alias of Incremental Compaction Strategy, which bounds temporary space per shard. A new pattern index answers LIKE queries without ALLOW FILTERING. The Raft log is now bounded on followers, and a node that dies during vnodes-to-tablets migration can be replaced. Alternator also implements DescribeExport, and raft metrics for strongly consistent tablets are exported per table.

Tablet routing now waits for migration to finish, view GC waits for the view to be built everywhere, and nodetool stop COMPACTION works again. Object-storage keyspaces are documented as Preview, and logstor gains frozen-mutation writes plus fixed disk space stats.

This week's changes focus on making background work behave predictably. Tablet routing information is withheld until a vnodes-to-tablets migration is finalized, so clients are not routed to tablets while a keyspace still replicates by vnodes. Tombstone GC on a view is suspended until the view has been built on every node. Gossip digest sends abort on shutdown instead of hanging, chunked S3 downloads are paced and bounded, and disabling compaction no longer fails on an already stopped manager. nodetool stop COMPACTION covers major compaction again. Object-storage keyspaces are now documented as a Preview feature, logstor can write a frozen mutation without unfreezing it, and disk space statistics for logstor tables are fixed.

ScyllaDB 2026.2.9 and 2026.3.3 candidates released. Fixed nodetool compaction control, topology requests, commitlog replay, and added BM25_HIGHLIGHT() for search excerpt generation.

ScyllaDB released 2026.2.9 and 2026.3.3 candidates this week. Key improvements include fixing nodetool stop COMPACTION for major compaction, adding BM25_HIGHLIGHT() for full-text search excerpts with matched terms marked, system.custom_indexes virtual table for external search node monitoring, and prevention of mixed-storage clusters. Topology coordination and tablet restore reliability also improved.

This week: raft applier now uses mailboxes, secondary index validation fixed, tablet truncate support added, and strong consistency improvements shipped.

Latest Scylla updates include improved raft applier performance with mailbox-based communication, critical secondary index correctness fixes preventing incorrect query results, enhanced strong consistency for tablet migrations, and new truncate support for strongly consistent tables. Additionally, Alternator now supports the new DynamoDB Vector Search API, and multiple optimizations improve test execution speed.

Week 40: tablet-aware restore improvements, strongly consistent table migration, extended storage testing for tablets, secondary index predicate validation, and generic task conversion. Multiple stability and correctness fixes.

Week 40 shipped significant tablet improvements: inter-node migration for strongly consistent tables, tablet test coverage for S3/GCS storage, and tablet-aware restore fixes. Also improved secondary index correctness, extended view building stability, and continued generic task migration. Multiple backports address crashes in backup operations, streaming concurrency, and collection element handling.