Loading
test(perf): add k6 OCI pull metadata test with thresholds
What
perf/k6/seed.jspushes an OCI image with 1, 10, and 100 MiB layers.perf/k6/pull_metadata.jsreads it back, with thresholds that fail a slow run.- This MR just adds the k6 tests which are not run anywhere yet.
Governing ADRs
ADR 003 supplies the traffic shapes behind this use case. It projects throughput rather than latency, so these bars come from measurement.
Testing
Measured on Artifact Registry v2.18.0, one pod at 4000m CPU, 5 users, 20 seconds.
| Threshold | Bar | Measured |
|---|---|---|
manifest GET p95 |
25 ms | 9.49 ms |
blob HEAD p95, each layer size |
15 ms | 1.91 ms to 2.02 ms |
http_reqs count |
> 500 | 27,973 |
http_req_failed rate |
0 | 0 |
Each threshold fails when forced with a tight bar or a short run.
The same run at the chart default, limits.cpu: 200m and requests.cpu: 100m,
two pods with the autoscaler enabled, fails every latency threshold.
| Threshold | Bar | Measured | Result |
|---|---|---|---|
manifest GET p95 |
25 ms | 78.81 ms | fails |
blob HEAD p95, each layer size |
15 ms | 74.01 ms to 75.80 ms | fails |
http_reqs count |
> 500 | 7,821 | passes |
http_req_failed rate |
0 | 0 | passes |
The default serves 3.6 times fewer requests in the same 20 seconds
References
gitlab-org/quality/quality-engineering/team-tasks#4613 (closed) gitlab-org/quality/quality-engineering/team-tasks#4614 (closed)
Edited by Harsha Muralidhar