[Fusion] Batch path segment pool access and keep the local pool across requests

[tobias-tengler]

Every result path borrows an int[] from the process-wide PathSegmentPool, and every rent and return takes its SpinLock. A request with nested lookups needs hundreds to thousands of paths, so with a few busy cores the threads spin on that lock, and OperationPlanContext.Clean returns every array through it once more.

The shared pool now rents and returns arrays in batches of 32, and the request-local pool stays alive in the pooled FetchResultStore: after a request it marks its arrays free again and returns only what exceeds PathSegmentLocalPoolCapacity (64 by default, 1 the minimum). A request touches the shared lock a few times instead of twice per path. The local pool shrinks its bookkeeping after a burst and drops its arrays when the shared pool has been replaced; the shared pool rejects arrays of another length without unbalancing its in-use count. A result document is detached from the local pool when the store hands its memory to the operation result, so a path read after execution is allocated and never touches another request's arrays. The local pool drops returns once every owned array is free, so a straggling return after a reset cannot hand the same array to two paths, and the shared pool holds its lock for at most 32 arrays per rent or return.

Benchmarks

Each thread is one pooled request context with its own local pool, running requests back to back. A request rents the given number of paths, returns half and ends. Mean is the time per request on one thread.

Path segments, 1 thread, 32 paths per request

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Rent, return, reset	main	588.2 ns	1.00	51 B	1.00
Rent, return, reset	PR	145.6 ns	0.25	3 B	0.06

Path segments, 1 thread, 1024 paths per request

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Rent, return, reset	main	17.034 µs	1.00	51 B	1.00
Rent, return, reset	PR	11.093 µs	0.65	3 B	0.06

Path segments, 4 threads, 32 paths per request

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Rent, return, reset	main	11.678 µs	1.00	197 B	1.00
Rent, return, reset	PR	0.191 µs	0.02	5 B	0.03

Path segments, 4 threads, 1024 paths per request

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Rent, return, reset	main	253.519 µs	1.00	200 B	1.00
Rent, return, reset	PR	47.240 µs	0.19	5 B	0.03

Path segments, 16 threads, 32 paths per request

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Rent, return, reset	main	29.349 µs	1.00	778 B	1.00
Rent, return, reset	PR	1.181 µs	0.04	6 B	0.01

Path segments, 16 threads, 1024 paths per request

Method	Version	Mean	Ratio	Allocated	Alloc Ratio
Rent, return, reset	main	308.177 µs	1.00	778 B	1.00
Rent, return, reset	PR	188.465 µs	0.61	12 B	0.02

[github-actions]

Patch coverage

98.3% of changed lines covered (174/177)

File	Covered	Changed	Patch %
…/Fusion/src/Fusion.Execution/Text/Json/PathSegmentPool.cs	76	79	96.2% 🟡
…/src/Fusion.Execution/Execution/OperationPlanContext.cs	2	2	100.0% 🟢
…/Execution/Results/FetchResultStore.Pooling.cs	13	13	100.0% 🟢
…/src/Fusion.Execution/Execution/Results/FetchResultStore.cs	1	1	100.0% 🟢
…/src/Fusion.Execution/Text/Json/CompositeResultDocument.cs	2	2	100.0% 🟢
…/src/Fusion.Execution/Text/Json/PathSegmentLocalPool.cs	79	79	100.0% 🟢
…/Fusion/src/Fusion.Execution/Text/Json/PathSegmentMemory.cs	1	1	100.0% 🟢

Uncovered changed lines (JSON)

{
  "sha": "f63946282819db97eee5a86f68c594e7f818a02f",
  "files": [
    { "path": "src/HotChocolate/Fusion/src/Fusion.Execution/Text/Json/PathSegmentPool.cs", "ranges": [[124, 126]] }
  ]
}

---

Project coverage: 59.0% (310997/527501 lines)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.