Redis 8.4 extends XREADGROUP with optional CLAIM min-idle-time: one command reclaims idle pending stream entries, then spends the remaining COUNT budget on new messages (>), collapsing the old XPENDING → XCLAIM/XAUTOCLAIM → XREADGROUP recovery loop into a single round trip (blog, Sergey Georgiev, 2026-05-26; docs).
This post covers Streams consumer-group CLAIM only, not other Redis surface area.
What shipped
When CLAIM is set, the command does two things in order, sharing one COUNT budget (blog, docs):
- Claim first — reclaim PEL entries idle ≥
min-idle-time(orphans prioritized). If claims fillCOUNT, return immediately. - Then read — remaining budget goes to new entries (
>), same as classicXREADGROUP.
With CLAIM, responses include idle time (ms) and delivery count so clients can tell reclaimed vs fresh and build retry / dead-letter caps without a separate XPENDING. Lock wording to the blog: idle >0 = reclaimed / 0 = fresh; delivery count 0 = new / ≥1 = claimed.
XACK is still required after successful processing—CLAIM simplifies recovery; it does not replace acknowledgements (blog).
Docs synopsis order for optional tokens: [CLAIM min-idle-time] [NOACK] before STREAMS (docs).
Behavior notes
BLOCK: can wake on new entries or when the next pending entry ages pastmin-idle-time(reactive reclaim wakeup).- Ignored when the stream ID is not
>(e.g. replaying own pending)—standard response shape, no CLAIM extras (docs). - Compatibility: fully optional; mix CLAIM and non-CLAIM consumers in one group. An internal
streamNACKlinked-list opt (replacing a time-ordered rax index) does not change protocol/RDB/AOF formats (blog).
Vendor claims
All figures below are Redis-reported—not independently verified (blog):
- Vs
XAUTOCLAIMon their stress setup (20k PEL / 1k idle / COUNT=1000): avg claim latency 54.671 ms → 2.426 ms — “up to 22.5× faster on average.” Frame as Redis-reported / up to / that workload; blog caveats that speedup scales with PEL size ÷ idle fraction—small or mostly-idle PELs see much smaller wins. - Linked-list vs rax (memtier, 2M msgs): 4,935 → 6,321 ops/sec (+28% throughput; blog also cites −22% avg / −21% P99 latency).
- Earlier rax index overhead
18.6 B/entry (8.7% on their 200k-PEL memory test)—vendor measurement; later removed by the linked-list opt.
These are not universal speedups, and the post makes no comparison with Kafka or NATS.
Who should care
Teams running agent job queues / event pipelines on Redis Streams who still hand-roll reclaim loops should start at the Redis blog and XREADGROUP docs.

The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment