mirror of
https://github.com/openobserve/openobserve.git
synced 2026-11-06 00:09:22 +00:00
Sub-cluster internal node failures (h2 errors) were silently dropped — partial_err was never transmitted to the super cluster leader, causing degraded results to be cached as if they were complete. Adds PreCustomMessage::PartialErrRefEarly + PartialErrRef to carry partial_err over the flight wire independently of ScanStats (which can be disabled at deployment level). Sent twice for reliability: once before first batch (catches fast-failing nodes), once after all batches (final accumulated value). Decoder dedupes on receive. Also elevates node gRPC error log from info → error.