[novatechflow]

Summary

Produce and Fetch now return a retriable KAFKA_STORAGE_ERROR (56) instead of UNKNOWN_SERVER_ERROR (-1) when S3 is unavailable or an S3 request fails. librdkafka treats -1 on Produce as permanent and drops the message.

Changes

• backpressureErrorCode returned UNKNOWN_SERVER_ERROR for S3StateUnavailable and for the default case, which covers a failed append or read while S3 still looks healthy. Both now return KAFKA_STORAGE_ERROR. Degraded S3 still returns REQUEST_TIMED_OUT.
• Produce sets the partition error message, which clients see on v8 and later: "object storage degraded, retry later", "object storage unavailable, retry later" or "object storage request failed, retry later". It does not include bucket names or the underlying error.
• A partition log that cannot be restored from S3 now also returns the storage error on Produce and Fetch, instead of -1.
• Adds KAFKA_STORAGE_ERROR to pkg/protocol/errors.go, and updates docs/architecture.md and the release notes.

Tests

• TestProduceBackpressureUnavailable and TestFetchBackpressureUnavailable now expect 56.
• TestProduceBackpressureErrorMessages checks the code and message on Produce v9 for degraded, unavailable and append failure.
• TestFetchRestoreFailureReturnsStorageError covers the restore failure path.

All of them fail against main and pass with the fix.

Not covered here: the e2e or live check with a librdkafka producer through a short MinIO outage.

Fixes #255

[PaxMachinaOne]

🤖 PaxMachina automated review (worker-code)

The change correctly maps S3 failures to retriable Kafka errors without exposing sensitive storage details.

No blocking findings.