Last 2 weeks in scylla-cluster-tests.git master (issue #135; 2026-05-08)

This short report brings to light some interesting commits to scylla-cluster-tests.git master from the last 2 weeks. Commits in the 4d0c7f8b…9e778747 range are covered.

There were 74 non-merge commits from 13 authors in that period. Some notable commits:

SCT gained first-class Apache Cassandra cluster support (docker based).

Building on the new Cassandra plumbing, Scylla-migrator CS-to-Scylla migration tests landed — a SparkMigratorTest with small (~30 min, 100K rows) and scale (hours, 500GB / 4 × 125M rows) variants that drives an EMR job from a Cassandra source to a Scylla target.

OCI gained multi-DC support through a new OciVcnPeering class that wires up Dynamic Routing Gateways, DRG-VCN attachments, Remote Peering Connections and route/security rules between every region pair, exposed via a configure-oci-peering Hydra command.

A callback-based node configuration mechanism was added: an after_config hook stored on each node and run after config_setup() but before node_startup() , eliminating the ~30–60s restart overhead in nemesis operations that previously had to mutate config and bounce the node.

A vector-search end-to-end test was added for XCloud that drives disk utilization to the auto-scale threshold with cassandra-stress, then continuously inserts vectors and validates ANN queries throughout the resize, failing if any index lookup returns incorrect results.

Backtrace decoding was moved to an external servicebacktraces.scylladb.com is now tried first, with automatic fallback to local addr2line . The previous flow loaded 1+ GB of DWARF debug info on the monitor node and OOM-killed the 8 GB VM in longevity tests.

K8s multitenancy support was removed. The feature is preserved at docs/shelved/MULTITENANTK8s.md for future re-implementation.

Latte was bumped to 0.46.0 (via 0.45.0 the same week), bringing a tokio -alt-timer perf upgrade with sinusoidal-stream optimization, removal of the vulnerable rustls-webpki 0.101.7 dependency, and serialization/deserialization perf improvements that eliminate intermediate representations. On top of the bumps, latte enterprise perf test configs for tablets and vnodes were added (Jenkins pipelines, test-case YAMLs, latency thresholds, gradual load steps), and perf_gradual_throttle_steps now accepts a dict format with threads/concurrency/rate per step to drive latte’s multi-parameter config.

Gemini was bumped to v2.3.8 via v2.3.7, fixing a stmtlogger-induced hang and a partitions UUID leak, and earlier resolving false row-difference reports caused by pointer-vs-value comparison on the oracle side.

A native backup test for Google Cloud was added, extending the existing native-backup infrastructure with GCS object storage configuration. Manager sanity gained a multi-keyspace data validation workload driven by latte that covers every CQL data type across 2 keyspaces × 2 tables (with SI and MV), enabling deterministic column-by-column validation after backup → truncate → restore.

A set of 2025 codebase health audit reports was checked in, covering 6 SCT subsystems (remote execution, provisioning, config/REST/tester, cluster core, nemesis, skills), surfacing 7 critical, 9 high, 22 medium and 7 low severity issues for follow-up.

See you in the next issue of last week in scylla-cluster-tests.git master!