This short report brings to light some interesting commits to scylla-cluster-tests.git master from the 2 last weeks. Commits in the 769e0dda…34b7873e range are covered.
There were 128 non-merge commits from 20 authors in that period. Some notable commits:
Minicloud: SCT now has core support for a local QEMU/KVM cloud emulator that serves the EC2 and GCE compute APIs. Every knob is a documented minicloud_* config option, configurations/minicloud.yaml delivers the mandatory params, and an overview doc explains how to run against it. Jenkins support followed with artifact, longevity and rolling-upgrade jobs on a nested-virtualization sct-runner, where minicloud: true in the jenkinsfile is the whole opt-in.
GCE tests now survive capacity stockouts. Provisioning relocates instead of failing when a zone or region is exhausted. Turned on by default in defaults/gce_config.yaml.
Nemesis classes gained a precheck() feasibility hook that returns a human-readable reason to exclude a disruption permanently. The abort-decommission-tablets nemesis is the first user of the hook.
MV disruptions now collect diagnostics.
The monkey module split continued, with SLA nemeses and encryption nemeses each moved into their own module.
Nested config keys are exportable from bash now that SCT accepts SCT_<FIELD>__sub_key as an equivalent of the dot form. The dot notation stays supported, and remains the only way to express sub-keys that contain a hyphen.
Test-case labeling gained a validated team_ownership taxonomy field on the TestMetadata model, with core-test-infra, core-qa, storage-qa, cloud-qa and qa-tools as the accepted values.
The docker backend can simulate racks through the image entrypoint --rack argument.
Manager backup testing covers a second WORM mode. GCE bucket creation was split into object-lock and event-based-hold variants, which made room for a new event-based-hold backup test.
On the performance side, four GCE tests on z4d-highmem-16 instances were added — predefined throughput steps and 2.5TB elasticity latency with nemesis, in standardlssd and highlssd variants. Cost reduction also continued here, with perf tests moved from a weekly to a bi-weekly trigger now that microbenchmarks run after PGO builds.
Stress tool images moved forward. scylla-bench went to v1.3.2, which writes per-row value data by default and adds -no-random-data to opt out of it. gemini v2.6.1 with bugfixes, while latte 0.50.0 replaces the DynamoDB driver with the Alternator driver and can read binary data files in workload scripts.
A contribution handbook was written over the period, covering governance and roles, branches, CI/CD and security, and reshaped after review.
Vnode-to-tablet migration got its own coverage through a 4h utilization longevity config with nemesis exclusions.
See you in the next issue of last week in scylla-cluster-tests.git master!