This short report brings to light some interesting commits to scylla-cluster-tests.git master from the last week. Commits in the 1e7cf313…f65f8d14 range are covered.
There were 42 non-merge commits from 14 authors in that period. Some notable commits:
A new longevity test reproduces paxos table growth by writing a large space of unique primary keys with LWT and then removing the base rows with plain DELETE statements. Each successful CAS leaves a $paxos entry behind while the base table empties on disk.
The asymmetric cluster longevity test moved to GCE, loading 300M rows from a single mixed stress command at 45k ops/s on six 8-vCPU database nodes.
An integration test checks that the default monitor images exist in AWS and GCE. It resolves ami_id_monitor in all eight supported AWS regions and looks up gce_image_monitor in the project parsed out of the image URL.
SCT runs scylla-doctor 1.14, which handles LimitNOFILE=infinity from systemd, connects to IPv6 seed hosts, and writes journalctl dumps to the system temp directory. The release also adds a disk-performance-exceeded collector and analyzer, an authentication analyzer, and a service levels collector.
The minicloud backend gained minicloud_scylla_reserve_memory, which passes --reserve-memory to scylla-server so the guest OS keeps enough memory for sshd and the tools SCT runs on the node. It is empty by default and set to 3G for the rolling upgrade scenario.
A mini-plan for GCE network config alignment scopes the three gaps that remain, and the first of them landed: ip_type: ipv6 is rejected at config time on gce, with the offending address named and ipv4, aws or oci offered instead.
The health check optimization plan was rebuilt around retry delays as the dominant cost, with backoff and a cluster-level short-circuit ahead of cheaper check modes. The first instrumentation run covered 43 gates and 260 per-node samples across two 6-node clusters.
docs/performance/ holds a reference set for the performance suite: the test catalog, the phase skeleton every perf test shares, the results and regression path through HDR histograms and Argus, and the trigger YAML semantics. Three runbooks cover adding an instance type, a stress tool or a whole new perf test.
The perf weekly status report skill tracks latency-2.5tb-i8g-elasticity under the i8g Tablets category and drops the seven i4i jobs.
See you in the next issue of last week in scylla-cluster-tests.git master!