After the ingest-bug streak, the obvious move: make those bugs stay dead. The full format matrix is wired into CI now -- every accepted input format, CSV through parquet through the numpy and *vecs families, uploaded to a live cluster, ingested, and queried for exact expected results. Core subset on every push, whole matrix nightly. Ingest regressions get a red light with their name on it.
Making that sustainable was most of the work. Per-run collection namespacing so runs can't collide, create-confirm and upload retries so infrastructure blips don't read as format bugs, hard timeouts so a hang fails fast instead of wedging the runner for an hour. Right-sized fixtures made the suite appx 8x faster with the same exact-match property. A slow flaky gate is a gate people learn to ignore.
Two side quests from the same week. Our docs-example tests were hitting real third-party APIs, which means every CI run was slow-motion load-testing Wikipedia and a couple of museum archives from one shared egress IP. Sorry, Wikipedia. Mocked now, with recorded fixtures -- the docs tests test our code. And release builds went from fat to thin LTO, which parallelizes the link-time tail that had been serializing every deploy. I also actually bumped the deployed version, so the win shipped to real machines instead of living in CI as a benchmark.
June ended with the gates green. Least glamorous sentence I'll write all year.