ory / lumen / 24402038199

74%

Build Type

push

github

Committed by web-flow

Commit Message

fix(embedder): re-probe unhealthy servers and resolve dimension mismatch on failover (#132)

* fix(embedder): re-probe unhealthy servers and resolve dimension mismatch on failover

Two related bugs in FailoverEmbedder when the primary embedding server is down:

1. Once initServers() marked all servers unhealthy, the checked flag prevented
   re-probing on subsequent Embed() calls — even after the server came back up.
   Fix: re-probe when active < 0 and a 30-second cooldown has elapsed. Also
   reset active = -1 on failover exhaustion so the re-probe path triggers.

2. ModelName() and Dimensions() returned stale server-0 values before the first
   Embed() call. When failover later selected a different server with different
   dimensions, the Indexer/Store was already created with wrong dimensions,
   causing sqlite-vec errors. Fix: trigger eager initServers() in ModelName()
   and Dimensions() so they reflect the server that will actually handle
   embeddings.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor(embedder): extract maybeReprobe helper to deduplicate init/reprobe logic

Dimensions(), ModelName(), and Embed() all had identical init/reprobe
blocks. Extract into maybeReprobe(log bool) called under the existing
mutex, with the log parameter controlling reprobe diagnostics in Embed().

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Coverage Stats

15 of 15 new or added lines in 1 file covered. (100.0%)

6 existing lines in 2 files now uncovered.

3274 of 4528 relevant lines covered (72.31%)

706.32 hits per line