This release improves CI test sharding. Shard timing estimates are now based on the 90th percentile of all CI runs in the last 30 days, rather than the average of runs on the default branch. This means modules that are only exercised on feature branches now get real per-unit estimates instead of falling back to a global median, so shards are balanced more accurately. The p90 estimator is also less sensitive to a single slow or flaky run, so timing estimates stay conservative and stable.
- Test shard timing now uses the p90 of all CI runs over the last 30 days, giving better balanced shards for new or selectively-run modules and reducing the impact of outliers.