Last week in scylladb.git master (issue #326; 2026-07-19)

This short report brings to light some interesting commits to scylladb.git master from the last week. Commits in the 9cda315bbab..14339e70165 range are covered.

There were 109 non-merge commits from 27 authors in that period. Some notable commits:


The version number was updated to 2026.4.0-dev, marking the start of the 2026.3 stabilization cycle and begining of 2026.4 development.

The sstable writer learned to deduplicate compression dictionaries on sstable write, instead of just on sstable opens. This reduces memory consumption.

Fast bootstrap for tablet raft groups (for strongly consistent tables) will no longer bias leadership to the same nodes.

A materialized view race between view creation and an update was fixed.

The bundled cqlsh was updated to version v6.0.37-scylla.

Tablet migration now also migrates the schema history, so that paxos mutations stored in the paxos table can refer to column mappings from older schema versions.

The minimum hardware requirements on x86 were upgraded from Westmere-class CPUs to x86-64-v3. This lets the compiler use the AVX instruction set.

Nodes can now join the cluster while unrelated tablet migrations are in progress.

The nodetool command to stop sstable reshape now stops reshape on all tablets, instead of working on a single tablet.

Alternator, ScyllaDB’s implementation of the DynamoDB API, now runs connection setup in a separate service level. This reduces interference from newly connecting clients.

The perf-simple-query benchmarking tool is now shard-aware to reflect how ScyllaDB is typically used.

Locking of the topology for Paxos requests was simplified and improved. This prevents spurious stale topology errors.

The layout of sstables on object storage buckets was restructured.

The compaction manager fixed a race condition between tablet split and picking candidates for compaction.

The topology history kept by the cluster was reduced from a week to an hour to reduce problems transferring long histories.

When cancelling many pending writes (typically because a node went down), we are now more careful to avoid stalls.


See you in the next issue of last week in scylladb.git master!