What yugabyte/yugabyte-db shipped

This week: xCluster no longer silently diverges on partial/expression indexes after ALTER COLUMN TYPE, object lock releases wait when catalog versions cannot be read, and the upsert guard now checks partitions. Plus Connection Manager and CI fixes.

Changes this week in yugabyte-db: ALTER COLUMN TYPE with expression and partial index reuse no longer causes silent xCluster divergence when the reuse setting differs between sides. Exclusive object lock releases are deferred rather than sent without catalog versions, legacy-mode transactions take local object locks during promotion, and the upsert guard checks the target partition for each row. Also: heartbeat handling keeps transactions alive until sub-transaction rollback completes, the pooled sequence currval/lastval fallback is removed, Connection Manager debug builds compile again, and pr-lint now fails on unformatted managed Java.

This week: restores no longer lose values from columns brought back or leave orphaned vector index storage, vector index flush failures now surface instead of hanging. Merge scan costing fixed, xCluster schema check tolerates renamed indexed columns, and Go and Java dependency CV

This week in yugabyte-db: point-in-time restores now forget columns they bring back, so compaction stops discarding their values, and tservers clean up vector index storage that restores and database clones leave behind. Vector index flush failures now reach the callers that flush before checkpoints or intent deletes instead of blocking tablet shutdown. Also in: merge scan costing charges per row for the k-way merge, xCluster schema validation tolerates renamed indexed columns, and dependency upgrades clear critical CVEs in yugabyted-ui and YBA.

This week: DISTINCT no longer returns duplicates under batched nested loop joins, FK VALIDATE stops waiting on concurrent deletes, YSQL migrations retry on serialization failure, and new universes get their first master leader faster. YBA adds a v2 login API.

This week in the repository: DISTINCT no longer returns duplicates under batched nested loop joins, VALIDATE CONSTRAINT on a foreign key no longer waits on concurrent deletes, and YSQL migrations retry on serialization failure instead of failing the upgrade. Cluster creation is faster, with quick master leader election now on by default and the first YSQL lease requested immediately. Replica placement candidates are built once per table, yugabyted sets glibc.pthread.rseq for the servers it starts, and YBA adds a v2 login API plus fixes for rollback of failed K8s edits.

Weekly update: pg_cron leader failover fixes, in-flight cron jobs allowed to finish on graceful stepdown, master accepts restarted full tablet reports, yb-admin help lists operations, and yb_tablegroup_size() reports colocated tablegroup size.

This week in yugabyte-db: pg_cron leader failover is fixed so a new leader waits out the old lease, and in-flight jobs can finish when the leader steps down gracefully. The master now accepts a restarted full tablet report on leader change. yb-admin help lists operations instead of dumping gflags. A new yb_tablegroup_size() function reports the size of a tablegroup's colocation parent tablet. Vector index restore and hidden-index removal fixes cover xCluster DR and PITR cases, and the ASH buffer is no longer zeroed in fresh shared memory.

Cluster startup gets faster this week: quick master leader election, tserver initdb overlapping the wait for master, an immediate first YSQL lease, and fewer postgres starts per node. Plus fewer postgres starts, heartbeats deadlock fix, and last_autoanalyze in pg_stat_user_tables

This week's changes focus on how fast a cluster and its connections come up. yb-master now defaults quick_leader_election on, so an RF3 universe does not wait a full election timeout for its first master leader on create or restart. The tserver runs its local initdb while waiting for the master rather than after TabletServer::Init, skips config probes and the postgres version check during local initdb, and requests its first YSQL lease immediately instead of waiting a full refresh interval. Tests shut down ExternalMiniCluster tablet servers in parallel. Elsewhere: retryable request ids come from sharded lock free allocators, strings::Substitute and StringPrintf are replaced by Format() in src/yb, free memory is returned to the OS after connection startup, a heartbeat deadlock with colocate

Catalog preload now frees memory per row instead of batching, and cache metrics update outside locks. SST statistics export to Prometheus for monitoring.

This week we reduced catalog preload memory overhead by freeing scanned rows immediately rather than batching deletions, lowering peak memory during first connections after DDL. Cache statistics now update outside mutex locks to reduce contention. We also added Prometheus exports for SST garbage statistics, enabling operators to build dashboards and study compaction thresholds.

New in yugabyte-db: faster DDL on large catalogs with adaptive response budgets, placement block replica limits, thin client UINT32 support, reduced lock contention in operation tracking, and per-drive fsync metrics.

This week's yugabyte-db updates focus on performance and reliability improvements. Catalog preloading now adapts to response size budgets instead of fixed row limits, reducing latency after DDL on large clusters. New placement block constraints give operators fine-grained control over replica distribution. The thin client gains support for UINT32 columns to open tables keyed by oids. Lock contention improvements reduce blocking in operation tracking. Per-drive write throughput metrics alongside fsync latency provide clearer device performance visibility.

Distributed tracing now flows across RPC boundaries, thread hops, and shared memory exchanges. Better error codes and safety checks improve debugging and prevent silent failures.

This week's work strengthens distributed tracing throughout YugabyteDB. Trace context now propagates across RPC boundaries, thread boundaries, and the PG-tserver shared memory path—enabling operators to follow requests through the entire system. We also improved error reporting with proper PostgreSQL error codes, added resource limits for xCluster operations, and added safety checks to fail fast when critical components are missing.