Complete, reproducible comparisons of Tako, nginx, HAProxy, Envoy and Caddy.
The current 27 October full-suite report contains all 135 cases, throughput, latency, errors, CPU, memory, channels and workflows. RESULTS.md summarizes the findings and measurement limits. Every chart uses this dataset; older VM results remain in the archive.
Use the free standard Ubuntu runner and the pinned full-suite workflow. Each repetition compares all five proxies in a two-CPU, 8 GiB, no-swap container. Proxy, app and load generator share that budget. Read the harness guide for source pins, matched backend limits, checks and artifact contents.
gh workflow run full-suite.yml -R tako-sh/performance --ref '<performance-branch>'
gh run list -R tako-sh/performance --branch '<performance-branch>'
gh run watch '<run-id>' -R tako-sh/performance --exit-status
gh run download '<run-id>' -R tako-sh/performance --dir '<artifact-directory>'Manual dispatch requires the workflow to exist on the default branch. Before that, use its PR-triggered run. Verify the run's exact source and harness revisions. Do not compare a narrow experiment with historical charts as a full refresh.
To regenerate the current report after collecting its artifacts:
python3 results/20261027-full-suite/analyze.py results/20261027-full-suite
python3 results/20261027-full-suite/render.py results/20261027-full-suiteThe dated analyzer checks all identities and raw metrics, preserves the three recorded sampling caveats, and marks the invalid CPU aggregate unavailable. It does not edit the raw data. Rendering requires Matplotlib.
- Merged buffer-lifetime validation: 188 complete correctness cases, held/streaming memory and paired performance.
- Initial buffer experiments.
- TLS offload investigation.
- Memory harness, buffer harness, and production validation.
- Historical VM report; remote scripts and proxy configs remain available for dedicated test machines.
Raw failures and superseded runs are retained. Public reports exclude credentials, private hosts and runtime logs. Load-balanced mode requires a larger or multi-node testbed and is outside this single-host suite.