Skip redundant Vitess rebuilds and record per-task timings

[arthurschreiber]

Context

Rebuilding Vitess from source is the single most expensive step of a benchmark run (several minutes). When one commit is benchmarked across the ~7 workloads, the cluster is torn down and recreated between each workload even though the exact same binaries are needed, so make install runs ~7 times for identical output.

Production timing over a 4.3-day sample shows the cost: first-of-config runs took 430–566s vs 110–254s for repeats of the same config — roughly ~25–30 wasted minutes per benchmarked commit.

Changes

Skip the rebuild when nothing changed

vitess_build now records the installed ref+Go version in /usr/local/vitess-build.ref and skips the git fetch and make install when it already matches and the binaries are still present. The marker is removed before a build and written only after a fully successful install, so an interrupted build can't cause the next run to wrongly skip.

To let the binaries survive between workloads of the same commit, the vtctld/vtgate/vttablet teardown no longer removes them from /usr/local/bin (a rebuild overwrites them, so nothing accumulates).

The when: guard uses | bool to avoid the ansible string-truthiness trap (a set_fact result of "False" is otherwise truthy).

Per-task timing (observability)

Adds ansible/ansible.cfg enabling the profile_tasks callback, wired via ANSIBLE_CONFIG (ansible-playbook does not otherwise discover a config from the per-execution playbook directory). Every execution's exec-stdout.log now ends with a per-task timing breakdown — the data needed to evaluate further optimizations (e.g. snapshotting the TPCC dataset).

Testing

• ansible-lint (v25.12.2, matching CI, production profile) — 0 failures ✅
• go build ./go/... + golangci-lint run ./go/infra/... (v2.12.2) — 0 issues ✅

Expected effect

Config switches within one commit drop from ~430–566s to roughly the repeat cost plus cluster start (~2–3 min).

🤖 Generated with Claude Code