[Cosmos] Replace per-client schedulers with shared CosmosSchedulers to fix thread scaling

[xinlian12]

Problem

PR testing revealed that global-ep-mgr and partition-availability-staleness-check thread counts increase linearly with tenant/client count because both GlobalEndpointManager and GlobalPartitionEndpointManagerForPerPartitionCircuitBreaker create per-instance Schedulers.newSingle() schedulers.

With N clients -> N dedicated threads for each component -> 2N extra threads just for background location refresh and circuit breaker staleness checks.

Solution

Replace per-instance Schedulers.newSingle() with shared static BoundedElastic schedulers in CosmosSchedulers, following the existing pattern used for COSMOS_PARALLEL, TRANSPORT_RESPONSE_BOUNDED_ELASTIC, etc.

Changes

CosmosSchedulers.java

• Added GLOBAL_ENDPOINT_MANAGER_BOUNDED_ELASTIC shared scheduler
• Added PARTITION_AVAILABILITY_CHECK_BOUNDED_ELASTIC shared scheduler

GlobalEndpointManager.java

• Replaced per-instance Schedulers.newSingle(CosmosDaemonThreadFactory) with CosmosSchedulers.GLOBAL_ENDPOINT_MANAGER_BOUNDED_ELASTIC
• Track background refresh Disposable via AtomicReference with getAndSet() to atomically clean up old subscriptions on concurrent calls
• close() cancels the tracked subscription instead of disposing the scheduler

GlobalPartitionEndpointManagerForPerPartitionCircuitBreaker.java

• Replaced per-instance Schedulers.newSingle("partition-availability-staleness-check") with CosmosSchedulers.PARTITION_AVAILABILITY_CHECK_BOUNDED_ELASTIC
• Track recovery Disposable via AtomicReference for consistent cleanup on close()

Design Decisions

• BoundedElastic over Single -- supports concurrent background tasks from multiple clients; threads auto-reclaim with 60s TTL
• Disposable tracking -- shared schedulers cannot be disposed per-client, so background subscriptions are tracked and cancelled individually in close()
• AtomicReference.getAndSet() -- prevents Disposable leaks when startRefreshLocationTimerAsync() is called concurrently
• Existing isClosed guards in both classes provide additional protection against post-close work

---

Benchmark Results Thread Scaling Fix Validation

Config: Gateway mode, ReadThroughput, concurrency=64, 10min per run, accounts lx1-lx28 (cycling modulo 28), host-pinned
Branches: upstream/main vs xinlian12/fix/shared-schedulers-thread-scaling

1. Throughput: main vs fix (H1 ReadThroughput, steady-state)

Tenants	main (ops/s)	fix (ops/s)	Delta
4	33,895	33,762	-0%
16	30,779	30,398	-1%
64	29,212	29,279	+0%
256	28,427	27,798	-2%
896	FAIL	27,009	fix succeeds

H2:

Tenants	main (ops/s)	fix (ops/s)	Delta
4	31,199	30,959	-1%
16	28,323	28,305	-0%
64	27,175	26,757	-2%
256	25,261	25,791	+2%
896	18,855	20,766	+10%

2. Thread Scaling (PEAK, H1 ReadThroughput)

Tenants	main	fix	Reduction
4	218	223	-2%
16	255	267	-5%
64	354	364	-3%
256	748	552	26%
896	1,748	544	69%

3. Thread Pool Breakdown (PEAK, H1 ReadThroughput)

Branch	T	part-avail	global-ep	glob-ep-bounded	bench-disp	transport	bulk-exec	reactor-ep	TOTAL
main	4	4	4	0	68	48	38	16	232
fix	4	4	0	5	67	47	41	16	236
main	16	16	16	0	67	43	39	16	255
fix	16	16	0	30	67	52	39	16	267
main	64	64	64	0	67	45	40	16	354
fix	64	64	0	65	67	49	43	16	364
main	256	256	256	0	67	54	42	16	749
fix	256	160	0	160	67	48	42	16	553
main	896	814	815	0	0	15	42	16	1,760
fix	896	160	0	160	67	43	42	16	548

4. Key Findings

• global-ep-mgr eliminated: 0 per-client threads across all tenant counts (was 1:1 on main)
• partition-avail capped at ~160: shared BoundedElastic pool reuses threads with 60s TTL (was 1:1 on main)
• global-endpoint-manager-bounded-elastic capped at ~160: replacement shared pool
• Thread count flat at 256-896t: fix holds at ~548 threads regardless (main grows 748-1,760)
• No throughput regression at 4-256t: -2% to +2% (within noise)
• 896t-H1 now works: main failed with client creation timeout, fix succeeds with 27K ops/s
• 896t-H2: +10% throughput improvement: 18,855 to 20,766 ops/s from reduced thread overhead

[xinlian12]

@sdkReviewAgent

[Copilot]

Pull request overview

This PR addresses a thread-scaling issue in the Cosmos Java SDK where per-client Schedulers.newSingle() usage caused background thread counts to grow linearly with the number of client instances. It introduces shared schedulers in CosmosSchedulers and updates background work to run on those shared schedulers instead of allocating dedicated per-client threads.

Changes:

• Added shared bounded-elastic schedulers to CosmosSchedulers for Global Endpoint Manager refresh and per-partition availability checks.
• Updated GlobalEndpointManager background refresh to use the shared scheduler and removed per-instance scheduler disposal.
• Updated GlobalPartitionEndpointManagerForPerPartitionCircuitBreaker to use the shared scheduler and removed per-instance scheduler disposal.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
sdk/cosmos/azure-cosmos/src/main/java/com/azure/cosmos/implementation/CosmosSchedulers.java	Adds shared bounded-elastic schedulers for endpoint refresh and partition availability checks.
sdk/cosmos/azure-cosmos/src/main/java/com/azure/cosmos/implementation/GlobalEndpointManager.java	Switches background location refresh work from per-instance single scheduler to shared bounded-elastic scheduler.
sdk/cosmos/azure-cosmos/src/main/java/com/azure/cosmos/implementation/perPartitionCircuitBreaker/GlobalPartitionEndpointManagerForPerPartitionCircuitBreaker.java	Switches staleness check work from per-instance single scheduler to shared bounded-elastic scheduler.

[xinlian12]

✅ Review complete (47:35)

No new comments — existing review coverage is sufficient.

_{Steps: ✓ context, correctness, cross-sdk, design, history, past-prs, synthesis, test-coverage}

[xinlian12]

/azp run java - cosmos - tests

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).