Skip to content

Remove (or dial down) blob-send diagnostic warn logging once send-starvation is diagnosed #592

Description

@kriszyp

Follow-up to #591 (blob-send warn visibility).

#591 adds throttled warn-level logging for the silent paths in sendBlobs() (park-on-cap, unpark-after-park, declined sends) specifically to diagnose a rare field failure: a deploy payload blob's content frames not going out for 60+ seconds while the replication connection is healthy and no shutdown drain is active (observed once on 5.1.19, 2026-07-16; peer's 20s blobReadTimeout failed the deploy; content flowed immediately after a fresh connection). Suspects: send parked on the per-connection outstandingBlobsBeingSent cap behind a stuck predecessor, or event-loop starvation on a single-worker origin during the deploy install phase.

The plan is to cherry-pick #591 into the next 5.1.x patch as a diagnostic lens. Once the starvation is caught and root-caused (or the warns prove noisy/uninformative), revisit this logging: remove it, reduce it to debug, or keep a slimmed-down version — whichever the diagnosis supports. This issue exists so the diagnostic instrumentation doesn't silently become permanent complexity.

Done when:

— Claude (Opus 4.8), on behalf of Kris

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions