mirror of
https://github.com/windmill-labs/windmill.git
synced 2026-09-07 00:01:49 +00:00
Ruben FiszelandClaude Opus 4.8
2318481f4f
* fix(embeddings): retry on failed init instead of disabling for a day The embeddings init task treated failure and success identically, so a transient error at startup left the embeddings DB uninitialized and every /query_hub_scripts and /query_resource_types call returned "Embeddings db not initialized": - If ModelInstance::new() failed (a transient HuggingFace/network error surviving its own 5 download retries), the spawned task exited for good and embeddings never came up until the next process restart. - If update_embeddings_db (hub fetch + fill_db) failed, the refresh loop still slept the full HUB_EMBEDDINGS_PULLING_INTERVAL_SECS (default 24h) before retrying. Wrap model init in a retry loop, and make update_embeddings_db report success so the refresh loop backs off by HUB_EMBEDDINGS_RETRY_INTERVAL_SECS (default 60s, env-configurable) on failure instead of the full pulling interval. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(embeddings): bound retry backoff to pulling interval A fixed 60s retry loop was fine for the transient case but spun forever in an environment where init can never succeed (air-gapped instance with embeddings left enabled): an error log + HuggingFace fetch storm every ~60s indefinitely, a behavior regression versus failing once and going quiet. Decay the retry with exponential backoff capped at HUB_EMBEDDINGS_PULLING_INTERVAL_SECS, resetting on success. A transient blip still recovers within ~60s; a permanently-broken env settles into ~1 attempt per pulling interval (default: 1/day). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>