How do I measure flake rate reliably?

Tag test runs with commit SHA and runner ID; a failure followed by a pass with no code change counts as a flake. Aggregate from JUnit across 7 days and compute flake_count / total_runs per test. Anything >2% goes to quarantine immediately.

Should I add retries to tests?

Prefer zero retries. If you must, cap at 1 and only around known-transient boundaries (browser startup, network handshake). Never use retries to mask data races, timeouts, or order-dependence.

Do I need Bazel to get fast pipelines?

No. Bazel helps at scale, but you can get 80% with path filters, sharding, aggressive caches (pnpm/pip/Gradle), BuildKit cache-to/from, and moving heavy suites post-merge behind a merge queue.

Release-engineering · Nov 8, 2025 · 10 minute read

The CI Flake Diet: 10‑Minute Pipelines, Lower CFR, Faster Recovery

Your CI isn’t slow—it’s noisy. Kill flake, shrink pipeline time, and move your DORA metrics in the right direction without hiring an army.

Back to all posts

The CI Flake Diet: 10‑Minute Pipelines, Lower CFR, Faster Recovery

Key takeaways

Implementation checklist