[RELEASE] ScyllaDB 2026.2.3

Release Notes
, ,
1

The ScyllaDB team is pleased to announce the release of ScyllaDB 2026.2.3, a production-ready patch release for ScyllaDB 2026.2 Short Term Support (STS) Minor Feature Release.

More information on ScyllaDB Version Support Policy is available here.

Related Links

Bug Fixes

The following issues are fixed in this release:

Alternator

  • Failed Alternator requests that return an api_error result are now correctly marked as errors in the audit log. Previously, such requests were audited as successful, meaning audit records did not accurately reflect the outcome of failed operations.
    scylladb#30764

  • Alternator write requests could intermittently fail with internal server errors after tablet migration. This was caused by missing column mapping metadata on the destination node. The issue is now resolved.
    scylladb#30748

Compaction

  • A stale-snapshot race between the split replacer and the regular compaction picker is now fixed. Under certain conditions, a split compaction and a concurrent regular compaction could observe inconsistent snapshots of the SSTable set, leading to undefined behavior. Two related race conditions are addressed in this fix.
    scylladb#30788

  • A race between compaction split bypass and regular compaction selection is now resolved. A split bypass compaction racing with a concurrently selected regular compaction could result in incorrect compaction behavior. The fix ensures proper serialization between these two compaction paths.
    scylladb#30829

Materialized Views (MV)

  • A node abort could occur during shutdown due to an incorrect teardown sequence in the materialized view subsystem, where internal resources could be accessed after being destroyed. The fix ensures proper sequencing of shutdown operations.
    scylladb#30923

Service Levels

  • The service level self-healing mechanism now correctly waits for replicas before applying self-healing updates. Previously, the coordinator could proceed without confirming that all replicas had received the updated service level configuration, leading to inconsistent enforcement of service levels across the cluster.
    scylladb#30855

SSTables

  • The trie writer now handles failed memory allocations gracefully. When a large allocation in the trie writer failed, the failure was not properly caught, causing an unhandled exception and a node abort. Allocation failures are now detected and handled correctly, improving node stability under memory pressure.
    scylladb#30867

  • SSTables could have two internal copies of their feature set go out of sync in certain code paths, potentially leading to unexpected behavior, such as incorrect results in materialized view reads. The issue is now resolved.
    scylladb#30899

Storage

  • S3 error classification is now hardened for non-AWS-compatible responses. Previously, the S3 client used a fragile error classification strategy that could misidentify errors returned by S3-compatible storage services that do not fully adhere to AWS error response formats. Unknown errors are now handled gracefully, and the build no longer fails due to the missing generated header utils/s3/aws_error_definitions_generated.hh.
    scylladb#30877

Topology

  • Requests could intermittently fail with stale topology errors when a tablet change occurred during request processing. The storage proxy now maintains a stable view of the replication topology for the full duration of a request, preventing mid-request topology refreshes from rejecting in-flight operations.
    scylladb#30849

  • The topology coordinator now skips the RF (Replication Factor) change rebuild retry when tablet balancing is disabled. Previously, when tablet balancing was disabled, the coordinator would unnecessarily retry the rebuild step after an RF change, causing redundant work and potentially delaying the completion of topology operations.
    scylladb#30791

  • Internal cluster coordination history could grow unbounded in clusters with frequent topology operations, eventually causing internal query failures. The retention period for this history is now reduced from one week to one hour, preventing memory pressure and query limit violations.
    scylladb#30790

Vector Search

  • HTTP 503 (Service Unavailable) responses from the ANN (Approximate Nearest Neighbor) service are now handled correctly. Previously, receiving a 503 during an ANN request could result in an unhandled error path, causing degraded behavior or a node abort. The fix ensures 503 responses are treated as retryable service unavailability rather than fatal errors.
    scylladb#30859