Hey everyone, we shipped several fixes for the piece version/loading issues some of you hit recently. These were platform infrastructure issues, not problems with the pieces themselves.
What changed:
• Broken piece versions now migrate automatically. Flows pinned to known bad versions are moved to the fixed release when loaded, without manual upgrades or republishing.
• Worker piece installs now self-heal. We fixed install layouts that could cause PieceNotFoundError, and workers now verify cached pieces before use and reinstall them automatically if needed.
• Hung steps now report proper timeout errors instead of cryptic internal RPC failures.
• Worker warm-up is faster. Pieces are now pre-installed based on distinct piece versions rather than flow count, significantly improving startup on larger instances.
For automatic piece upgrades, we run compatibility checks before upgrading, record exact before/after versions in the audit log, and can revert affected flows if needed.
Everything is already live on Cloud. For self-hosted, most of these fixes are in 0.90.4 (out now), and the remaining two (proper timeout errors on hung steps, and the automatic HTTP piece 0.11.11 to 0.11.16 migration) ship in the next release.
Sorry for the disruption, and thanks to everyone who reported these issues.