This short report brings to light some interesting commits to scylladb.git master from the last three weeks. Commits in the d5efd1f6768..61e5ec98882 range are covered.
There were 317 non-merge commits from 41 authors in that period. Some notable commits:
The master branch version was set to 2026.3, indicating the beginning of the 2026.3 development cycle and the 2026.2 stabilization cycle.
Restore in tablets mode will now stream sstables directly to the nodes and shards that own the tablets. This greatly improves restore performance.
ScyllaDB now tolerates rows in the system role_permissions table containing all NULLs (apart from the primary key). Such rows could have been created by automatic migration from pre-Raft versions.
The raft implementation now does a better job at disambiguating entries’ commit status.
The Trie index writer is now more careful about avoiding reactor stalls.
Deserialization of large collections was made more efficient.
Alternator is ScyllaDB’s implementation of the DynamoDB API. Alternator Streams are now readable after they have been disabled (data will expire automatically).
Alternator is ScyllaDB’s implementation of the DynamoDB API. Alternator Streams are no longer experimental.
The build system can now collect packaging artifacts into a single directory rather than leaving them scattered around the filesystem.
Repair requests for empty token ranges are now rejected.
User-defined aggregates can define an optional REDUCEFUNC in support of autoparallelization. ScyllaDB now requires EXECUTE permissions on REDUCEFUNC. User-defined aggregates are experimental.
Validation of tuple lengths when constraining a tuple of columns using an IN relation was improved.
A new abort_on_malformed_sstable_error configuration item allows early exit when detecting bad sstables.
The tablet load balancer now applies hysteresis to avoid continuous intra-node rebalancing.
Alternator, ScyllaDB’s implementation of the DynamoDB API, gained more vector search features such as control of the similarity function, an optimized data type, and returning the similarity scores.
The system now recovers from a partial upgrade of the service levels table that could have been left by previous versions.
See you in the next issue of last week in scylladb.git master!