[alexluong]

A batched retry lookup fails as a whole when one returned row can't be decoded (event metadata that isn't a JSON object of strings, or response data that isn't JSON). Every retry in that query gets the error and is received again in a new batch, so one bad row counts a receive against every retry it shares a batch with and can push them into the retry DLQ. Without batching, only the retry whose own row is bad fails.

Outpost writes these columns itself, so a bad row needs an outside write or a bug. This change is defensive: a bad row now fails only its own retry. Query errors (timeout, store down) still fail the whole query, so they don't turn into extra per-key queries.

1. driver.KeyErrors: ListLatestAttempts returns the other keys' results along with the per-key decode errors.
2. ClickHouse and Postgres scans keep rows they can't decode, with the error, instead of aborting. ListAttempt on both still fails on such a row. Postgres reads event_metadata as raw JSON and decodes it after the scan, so a bad value no longer fails Scan itself.
3. Batcher: a key error fails only that key's waiters.

🤖 Generated with Claude Code