Last week in scylladb.git master (issue #287; 2025-08-03)

This short report brings to light some interesting commits to scylladb.git master from the last week. Commits in the a1d7722c6d..1c25aa891b range are covered.

There were 111 non-merge commits from 23 authors in that period. Some notable commits:


The main data structure holding vnode tokens was changed to avoid large allocations, which can produce stalls on very large clusters.

After replacing a node, the load balancer will immediately check available space, to avoid the up to 60 second delay from regularly scheduled checks that can delay load balancing.

Tablets now support LightWeight Transactions. Paxos state is stored in a side table allocated on demand for each table using LWT, rather than system.paxos.

Compaction start/end messages were demoted to DEBUG level, as they were deemed too noisy.

Repair will send smaller messages for partition differences with many small rows (like tobmstone-only rows), reducing stalls.

The CQL port listener will now first half-close the receiving side of the connection, drain all responses, then close the sending side. This prevents queries from being lost during graceful shutdown.

The writing side of the sstable Trie index writer was merged. Trie indexes will still not be written.

When the coordinator for a LWT transaction cannot find a co-located shard (because the coordinator was not also a replica), it will now allocate a random (but consistent) shard for processing, rather than always falling back to shard 0 and overloading it.


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