Akışları Sınırlandırma ve Kırpma

Unless told otherwise, a Redis Stream keeps every entry you append. For high-throughput producers this means unbounded memory growth. Capping and trimming let you keep streams to a manageable size.

The cleanest approach is to cap during XADD using MAXLEN. Redis evicts the oldest entries to keep the length at or below the limit.

Exact trimming can be costly. The ~ modifier requests approximate trimming, which is far cheaper because Redis only removes whole macro-nodes.

• MAXLEN 1000 exact
• MAXLEN ~ 1000 approximate, may keep slightly more

You can trim an existing stream with XTRIM, using the same options as the XADD capping clause.

Instead of length, you can trim by entry ID. MINID removes all entries older than a given ID, which maps naturally to time since IDs start with a millisecond timestamp.

Because the first part of a stream ID is a Unix millisecond timestamp, MINID lets you keep only the last hour or day of events. Compute the cutoff timestamp in your app and pass it in.

The LIMIT option (with approximate trimming) caps how many entries a single trim removes, smoothing out latency spikes on very large streams.

Use XLEN to see the current entry count and XINFO STREAM for detailed metadata including first and last IDs.

Trimming removes entries regardless of whether consumer groups have read them. If a slow consumer falls behind the trim point, those messages are lost for it. Size your cap to cover your slowest consumer's lag.

Use MAXLEN when you care about count and bounded memory. Use MINID when you care about a time window. Prefer the ~ approximate form in production for performance unless exact bounds are required.

Capping at add time is usually best: it keeps memory steady without a separate cleanup job. Reserve XTRIM for ad-hoc cleanups or when retention rules change.

You learned to control stream growth with MAXLEN capping, MINID time-window trimming, the cheaper ~ approximate mode, the LIMIT option to bound eviction work, and the caveat that trimming can drop messages slow consumers have not yet read.