[veeral-patel]

What

An autoscaling poller pool now reports its current target in two places:

1. On each poll — PollerScalingInfo.poller_target on the workflow and activity poll requests.
2. On the worker heartbeat — WorkerPollerInfo.poller_target, for all four poller types.

Why

Scaling suggestions ride on dispatched tasks and the matcher is FIFO, so a worker's share of suggestions tracks its share of pollers. Growth is proportional to current size, which preserves imbalance between identical workers instead of correcting it — measured at 74/12/14 percent across three identical workers, with a different winner each run.

Reporting the target gives the server the one number it needs to compare a worker against the fleet mean and withhold suggestions that would push it further from its share.

The poll-request copy is consumed inside the matching decision and discarded. Reporting it on the heartbeat too lets a follow-up read it per worker and export it as a metric, making fleet-wide poller distribution visible without instrumenting every SDK.

How

Poll path — AsyncPoller installs a target supplier on each poll task when it creates that task's PollScaleReportHandle. The poll request is rebuilt per poll, because the target moves. A default no-op on PollTaskAsync keeps pollers with nothing to report from having to implement it.

Heartbeat — PollScaleReportHandle is a local inside AsyncPoller.start() and is not reachable from Worker.buildHeartbeatCallback. PollerTracker already exists to bridge exactly that gap — its javadoc says it "bridges the gap between poll tasks (created inside start()) and heartbeat building" — so the supplier goes there, and buildPollerInfo reads it alongside the in-flight count it already reads.

Each tracker is already the right one per type: WorkflowWorker hands stickyPollerTracker to the sticky poll task and pollerTracker to the normal one, so the normal/sticky split works even though both share one PollerOptions.

Design notes

• AsyncNexusPollTask now overrides setPollerTargetSupplier purely to feed its tracker. It had taken the interface's no-op default because the nexus poll request has no field for a target — but it does have an autoscaler and a heartbeat slot, so the heartbeat reports it.
• Non-autoscaling workers use MultiThreadedPoller, which has no scale report handle, so no supplier is installed and the tracker reports 0 — which the proto defines as unknown.

Dependency, and the submodule bump

Needs temporalio/api#887.

The temporal-serviceclient/src/main/proto submodule moves to vpatel/poller-load-balancing-java-base on the api repo — the same two proto commits cherry-picked onto the api commit this repo already recorded, so the bump is minimal and WorkerPollerInfo is byte-identical to the PR branch.

It is pinned there rather than to the api PR branch deliberately: that branch is based on current api main, and taking it would also pull in Callback.NexusHandler.source_context, which generatePayloadLimitValidator rejects as an unclassified payload-bearing field — a failure unrelated to this change. The submodule should move to a released api version before merge.

Testing

./gradlew :temporal-sdk:compileJava and :temporal-sdk:compileTestJava pass; spotless applied. Not yet exercised end to end against a server — that has been done via the Go SDK (temporalio/sdk-go#2799), which shares the wire contract.