Disable exemplar reservoir for asynchronous instruments by default

[dashpole]

Fixes #8285

Asynchronous instruments do not accept context, so the default TraceBased exemplar filter can never record an exemplar. Use the DropReservoir when the TraceBased exemplar filter is used, similar to #8211.

Benchmarks

Since callbacks are made as part of collect, there is a significant baseline overhead from collection included in the benchmark.

goos: linux
goarch: amd64
pkg: go.opentelemetry.io/otel/sdk/metric
cpu: AMD EPYC 7B12
                                          │   main.txt   │              new.txt               │
                                          │    sec/op    │   sec/op     vs base               │
AsyncMeasureNewAttributeSet/AlwaysOn-24     3.662µ ± 13%   3.735µ ± 8%        ~ (p=0.699 n=6)
AsyncMeasureNewAttributeSet/TraceBased-24   3.586µ ± 10%   1.195µ ± 5%  -66.69% (p=0.002 n=6)
geomean                                     3.623µ         2.112µ       -41.70%

                                          │   main.txt   │               new.txt               │
                                          │     B/op     │     B/op      vs base               │
AsyncMeasureNewAttributeSet/AlwaysOn-24     3.344Ki ± 0%   3.343Ki ± 0%        ~ (p=0.571 n=6)
AsyncMeasureNewAttributeSet/TraceBased-24    3424.5 ± 0%     592.0 ± 0%  -82.71% (p=0.002 n=6)
geomean                                     3.344Ki        1.390Ki       -58.43%

                                          │  main.txt  │               new.txt               │
                                          │ allocs/op  │ allocs/op   vs base                 │
AsyncMeasureNewAttributeSet/AlwaysOn-24     16.00 ± 0%   16.00 ± 0%        ~ (p=1.000 n=6) ¹
AsyncMeasureNewAttributeSet/TraceBased-24   16.00 ± 0%   12.00 ± 0%  -25.00% (p=0.002 n=6)
geomean                                     16.00        13.86       -13.40%
¹ all samples are equal

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.9%. Comparing base (4086701) to head (d68998d).

Additional details and impacted files

@@          Coverage Diff          @@
##            main   #8286   +/-   ##
=====================================
  Coverage   82.9%   82.9%           
=====================================
  Files        314     314           
  Lines      24985   24989    +4     
=====================================
+ Hits       20730   20736    +6     
+ Misses      3882    3880    -2     
  Partials     373     373           

Files with missing lines	Coverage Δ
sdk/metric/exemplar.go	100.0% <100.0%> (ø)
sdk/metric/pipeline.go	90.3% <100.0%> (+<0.1%)	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

Disables exemplar reservoir allocation for asynchronous instruments when the default exemplar.TraceBasedFilter is configured (since async observations are recorded without a measurement context, trace-based filtering cannot ever sample), reducing per-attribute-set memory/alloc overhead in sdk/metric.

Changes:

• Pass instrument kind into reservoirFunc and return aggregate.DropReservoir for async instruments when exemplar.TraceBasedFilter is used.
• Expand unit tests to cover AlwaysOn/AlwaysOff/TraceBased across sync vs async instrument kinds.
• Add a benchmark covering async instruments with AlwaysOn vs TraceBased exemplar filters, and document the optimization in the changelog.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
sdk/metric/pipeline.go	Passes InstrumentKind into exemplar reservoir factory selection.
sdk/metric/exemplar.go	Drops exemplar reservoirs for async instruments under TraceBasedFilter.
sdk/metric/exemplar_test.go	Adds table-driven coverage for the updated reservoir selection logic.
sdk/metric/benchmark_test.go	Adds benchmark to quantify async attribute-set overhead with different filters.
CHANGELOG.md	Notes the user-visible optimization.

[MrAlias]

🚀

[Dipanshusinghh]

@dashpole I went through the changes across reservoirFunc, the updated tests, and the new benchmark.

The additional check in reservoirFunc to return DropReservoir for async instrument kinds (ObservableCounter, ObservableUpDownCounter, ObservableGauge) when using the TraceBased filter makes sense, since async observations don’t carry context and therefore cannot produce exemplars.

I also verified from the test updates that:

• AlwaysOff correctly skips reservoir creation
• AlwaysOn continues to invoke the provider
• TraceBased behaves as expected for both sync and async instrument kinds

The benchmark results for the async TraceBased case clearly show reduced allocations and memory usage, which aligns well with avoiding unnecessary reservoir creation.

Overall, the change looks consistent with the intended behavior and provides a meaningful optimization without affecting existing functionality.