merlimat opened a new pull request, #25938: URL: https://github.com/apache/pulsar/pull/25938
This PIP proposes automatic segment split/merge for scalable topics, a sub-PIP of [PIP-468: Scalable Topic Controller](https://github.com/apache/pulsar/blob/master/pip/pip-468.md). ## Motivation PIP-468 gives the controller the ability to split and merge segments, but only on explicit operator request. This PIP makes the controller scale a topic automatically, eliminating the partition-count-style operational toil scalable topics were meant to remove. ## Key design points - **Splits are fast; merges are lazy.** Splits fire quickly with only a short `splitCooldown` (1m) to coalesce bursts; merges are gated by a longer `mergeCooldown` (5m) + cold `mergeWindow`. - **The controller reacts, it does not poll.** New stream/checkpoint consumers register directly with the controller, so consumer-count scale-up is event-driven (within seconds). Traffic load is *pushed* into the metadata store by each segment's owning broker (only on a significant change, default ±25%) and read by the controller on a 60s tick — no per-tick RPC fan-out. - **Pure decision function.** `AutoScalePolicyEvaluator.decide(layout, load, consumerCounts, policy, now)` → `Split | Merge | None`, unit-testable in isolation. - **Bounded churn.** `maxSegments` (64) caps growth; `maxDagDepth` (10) caps merges-in-lineage to stop split↔merge flip-flopping (splits always remain available); threshold hysteresis between split and merge. - **Default-on cluster-wide** with per-namespace (`Policies`) and per-topic (`ScalableTopicMetadata`) overrides, following the existing `autoTopicCreationOverride` pattern. The PIP also specifies the new `SegmentLoadStats` metadata record, the `broker.conf` knobs, the admin client API, and the observability surface. ### Links - Parent PIP: [PIP-468](https://github.com/apache/pulsar/blob/master/pip/pip-468.md) - Grand-parent PIP: [PIP-460](https://github.com/apache/pulsar/blob/master/pip/pip-460.md) -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
