Repository navigation
Separate batch acceptance from Redis transport health #105
Description
PR #91 is conflicting and combines several independent memory, threading, API, and reconnect changes. The remaining batch-path problem should be implemented on current main after #104 rather than by reviving that branch.
Problem
The batch APIs can infer connection health from the number of accepted stream writes. An all-server-rejected batch may spuriously enter reconnect even though Redis is healthy, while a mixed accepted/rejected batch can hide the transport state.
Required change
- Represent transport/connection failure separately from per-item Redis acceptance.
- Do not reconnect for server-side
ReplyErroror for an all-rejected explicit-ID batch. - Preserve and document the partial-result contract for mixed accepted/rejected batches.
- Follow the agreed redis-plus-plus reconnect policy for genuine connection loss.
Acceptance
- One rejected add leaves the existing connection usable.
- An all-rejected batch does not reconnect.
- A mixed batch returns the documented partial result without claiming transport failure.
- A genuine disconnect follows the intended reconnect/retry policy.
- Implement as a focused PR on current
main; do not merge Fix thread safety, memory bugs, and error handling #91 wholesale.
Activity
derekste commentedon Sep 29, 2026
Reproduced at
0bb9c35e7480e5d01afd58a5ed5ce27f3191f226on private Redis: a single duplicate ID returnedRA_NOT_CONNECTEDand caused two new connections; an all-rejected duplicate/backward batch also caused two connections. The caller's connection was healthy in both cases.Fix PR: #109, stacked after #108. Native macOS lifecycle/write/fault tests pass (3/3). They cover accepted/rejected mixed results across all batch APIs, genuine socket loss after successful Redis acceptance, recovery even when another batch item succeeded, and recorded-command evidence that nothing is replayed. Full Linux validation is running.
Keep this issue open until the fix merges and passes validation. Related IOC investigation: fermi-ad/redis-pvxs-ioc#24; neither this reproduction nor the cursor fixes establish the cause of the historical #73 allocator crash.
bigsamich commentedon Sep 30, 2026
Verified against #109 (
904e0ef) and the stack head. All four behavioural criteria hold:RA_REJECTEDwith 0 new connections, and the connection stays usable. Onmainit returnsRA_NOT_CONNECTEDand opens 2 new connections.CLIENT KILLreturnsRA_NOT_CONNECTED, reconnects, and replays nothing.Two notes:
Fixes #105in Separate rejected stream writes from transport failures #109 only takes effect if Separate rejected stream writes from transport failures #109 targetsmainwhen it merges. It currently targetsdev/subscription-lifecycle.getStreamSnapshot, still reconnect on a rejection. Filed as Read, admin and health paths treat server error replies as connection loss (read-side counterpart of #105) #123.