feat(metrics): add bound instruments behind experimental feature flag by bryantbiggs · Pull Request #3421 · open-telemetry/opentelemetry-rust

bryantbiggs · 2026-03-14T20:36:53Z

Splits from #3392, per maintainer feedback. Built on top of #3420 (in-place delta collection, now merged). Refs #1374.

Summary

Adds BoundCounter and BoundHistogram to the public API behind the experimental_metrics_bound_instruments feature flag. These types cache the resolved aggregator reference for a fixed attribute set, allowing subsequent measurements to bypass per-call sort, dedup, hash, and HashMap lookup entirely.

let counter = meter.u64_counter("requests").build();
let bound = counter.bind(&[KeyValue::new("method", "GET")]);
bound.add(1); // ~1.9ns — no attribute lookup

Architecture

TrackerEntry bound_count

bound_count (introduced in #3420) tracks how many live BoundCounter/BoundHistogram handles reference an entry. Entries with bound_count > 0 are never evicted during delta collection.

Cardinality overflow handling

ValueMap::bind() returns Option> — None when at the cardinality limit. Bound handle types use a Direct/Fallback enum:

Direct: normal case — dedicated tracker, single-digit ns, no map lookup per call
Fallback: at overflow — delegates every call to the unbound Measure::call() path

Feature flag

All public API, traits, noop impls, and internal plumbing are gated behind experimental_metrics_bound_instruments.

Benchmark Results

Apple M4 Max, 16 cores (12 performance + 4 efficiency), macOS 15.4:

Benchmark	Time	vs Unbound
Counter Unbound	53.2 ns	—
Counter Bound	1.87 ns	~28x faster
Histogram Unbound	58.6 ns	—
Histogram Bound	6.57 ns	~8.9x faster

EC2 Counter Results (from #3392)

Instance	Unbound	Bound	Speedup
c7i.large	86.2ns	7.47ns	11.5x
c7i.xlarge	88.9ns	7.62ns	11.7x
c7i.4xlarge	90.1ns	7.65ns	11.8x
c7a.large	77.4ns	2.71ns	28.6x
c7a.xlarge	75.0ns	2.71ns	27.7x
c7a.4xlarge	75.0ns	2.71ns	27.7x
c7g.large	111.8ns	6.87ns	16.3x
c7g.xlarge	111.4ns	6.84ns	16.3x
c7g.4xlarge	112.9ns	6.73ns	16.8x

EC2 Histogram Results (from #3392)

Instance	Unbound	Bound	Speedup
c7i.large	97.1ns	17.1ns	5.7x
c7i.xlarge	99.0ns	17.9ns	5.5x
c7i.4xlarge	98.5ns	18.4ns	5.4x
c7a.large	92.5ns	13.7ns	6.8x
c7a.xlarge	92.4ns	13.7ns	6.7x
c7a.4xlarge	92.5ns	13.7ns	6.8x
c7g.large	133.2ns	27.7ns	4.8x
c7g.xlarge	133.5ns	27.6ns	4.8x
c7g.4xlarge	133.8ns	27.2ns	4.9x

Test Coverage

15 tests covering:

Cumulative and delta temporality for both Counter and Histogram
Bound + unbound sharing the same data point (same attribute set)
Idle delta cycles (no update = no export, but handle persists)
Cardinality overflow fallback to unbound path (Counter and Histogram)
Recovery after delta eviction frees cardinality space
Overflow fallback handles working across multiple delta cycles
Drop enabling eviction of stale entries
Multiple bound handles sharing the same tracker (ref-counting)
Binding with empty attributes

Test plan

All 164 metrics tests pass (149 existing + 15 new)
Compiles with and without the feature flag
Zero regression on unbound measurement path

codecov · 2026-03-14T21:02:23Z

Codecov Report

❌ Patch coverage is 91.98347% with 97 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.1%. Comparing base (c992379) to head (5bfcd99).

Files with missing lines	Patch %	Lines
opentelemetry-sdk/src/metrics/mod.rs	93.7%	57 Missing ⚠️
opentelemetry/src/metrics/noop.rs	0.0%	8 Missing ⚠️
...pentelemetry-sdk/src/metrics/internal/aggregate.rs	0.0%	6 Missing ⚠️
opentelemetry-sdk/src/metrics/noop.rs	0.0%	5 Missing ⚠️
opentelemetry/src/metrics/instruments/histogram.rs	60.0%	4 Missing ⚠️
opentelemetry/src/metrics/instruments/mod.rs	0.0%	4 Missing ⚠️
opentelemetry-sdk/src/metrics/internal/mod.rs	95.5%	3 Missing ⚠️
opentelemetry/src/metrics/instruments/counter.rs	66.6%	3 Missing ⚠️
...entelemetry-sdk/src/metrics/internal/last_value.rs	96.5%	2 Missing ⚠️
...emetry-sdk/src/metrics/internal/precomputed_sum.rs	96.4%	2 Missing ⚠️
... and 3 more

Additional details and impacted files

@@           Coverage Diff           @@
##            main   #3421     +/-   ##
=======================================
+ Coverage   83.7%   84.1%   +0.3%     
=======================================
  Files        126     126             
  Lines      25386   26592   +1206     
=======================================
+ Hits       21255   22364   +1109     
- Misses      4131    4228     +97

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

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

cijothomas · 2026-04-25T01:36:42Z

#3420 is merged. @bryantbiggs can you rebase? The spec is progressing in parallel (open-telemetry/opentelemetry-specification#5050), so this is perfect time to get this merged!

bryantbiggs · 2026-04-29T15:46:44Z

Pushed a round of fixes addressing all five comments; reply inline on the bound_count one with my reasoning for keeping ref-counting.

Two extras worth flagging that aren't tied to a comment: ValueMap::bind() now inserts both unsorted and sorted keys for new entries (mirrors measure()'s pattern so unbound calls hit the fast path regardless of which side created the tracker), and BoundCounter/BoundHistogram are now Arc-backed and Clone so a single bound state can be shared across threads without re-binding.

Add BoundCounter and BoundHistogram types that cache resolved aggregator references for a fixed attribute set. Created via Counter::bind() and Histogram::bind(), bound instruments bypass per-call attribute lookup for significant performance improvements (~28x for counters, ~9x for histograms). Architecture: - TrackerEntry.bound_count tracks live handles, preventing eviction - Direct/Fallback enum handles cardinality overflow gracefully - Unsupported aggregators (ExpoHistogram, LastValue, PrecomputedSum) fall back to unbound path instead of panicking All public API, traits, and internal plumbing are gated behind the experimental_metrics_bound_instruments feature flag. Includes 15 tests covering cumulative/delta temporality, overflow fallback, recovery after eviction, bound+unbound sharing, idle cycles, drop semantics, and empty attributes. Splits from open-telemetry#3392. Refs open-telemetry#1374.

cfg-gated imports must sort before non-gated imports of the same module per rustfmt's stable import ordering rules.

Add bound_count field to TrackerEntry to track live bound instrument handles. Entries with bound_count > 0 are never evicted during delta collection, ensuring bound handles always point to a live tracker. Moved from open-telemetry#3420 per review feedback — bound_count belongs with the bound instruments feature, not the delta collection refactor.

- Make no_attribute_tracker an Arc<TrackerEntry> so bind(&[]) returns a handle to the same tracker that measure(&[]) uses, preventing duplicate data points for empty attribute sets - Re-check bound_count under write lock in collect_and_reset() to prevent a TOCTOU race where bind() increments bound_count after collect() reads it as 0 under a concurrent read lock - Use Ordering::Release for has_been_updated stores in bound handles, matching the unbound measure() path

CI exercises bound instruments via --all-features, so the testing feature does not need to enable experimental_metrics_bound_instruments or pull in the tokio runtime additions.

Foundational OpenTelemetry guarantee is that telemetry must never crash the host application. A custom SyncInstrument impl that does not override bind() now returns a NoopBoundSyncInstrument and emits an otel_debug log, matching the silent-drop behavior of other default impls.

When bind() creates a new tracker, insert it under both the original and sorted attribute orderings, mirroring measure()'s insert pattern. Without this, an unbound measure() call with the same attrs in non-sorted order would miss the fast path and re-sort on every call when the bound side created the tracker first.

Switch from Box<dyn ...> to Arc<dyn ...> and derive Clone so a single bound state can be shared across threads or modules without re-binding. Drop semantics are unchanged: bound_count is decremented when the last clone is dropped, matching Rust's RAII expectations for handle types. Mirrors the existing Clone behavior on Counter and Histogram.

Tests: * bound_histogram_empty_attributes_shares_with_unbound mirrors the existing counter version for histograms. * bound_counter_view_filters_attributes_at_bind_time and a histogram analog confirm a view filters attributes at bind() time, with bound and unbound recordings collapsing onto the same filtered data point. Bench: * Counter_Bound_With_View_Delta confirms view filtering happens at bind time, leaving the hot path at unfiltered-bound speed. Updated bench required-features and added a note that criterion does not enforce regression on its own. CHANGELOG entries for both crates updated to mention Clone, the noop trait default, and the converged bound/unbound semantics.

The previous Fallback path re-routed every bound.add() through the unbound measure() path, creating a ~25x perf cliff at the cardinality limit (~50ns vs ~1.8ns) and letting attribution drift across delta eviction cycles as space freed and refilled. ValueMap::bind() now looks up or lazily creates the overflow TrackerEntry at the limit and returns a direct handle to it. The bound handle writes directly to overflow via a single atomic update for its lifetime — perf parity with a normal bind. To recover after delta eviction frees space, drop the handle and rebind. Spec-aligned with open-telemetry/opentelemetry-specification#5050: overflow data lands in the otel.metric.overflow=true bucket (MUST) and per-call attribute lookup is bypassed (SHOULD). Bench (Apple M4 Max): Counter_Bound_AtOverflow_Delta: 1.82 ns Histogram_Bound_AtOverflow_Delta: 6.58 ns

Removes BoundFallbackHandle and the fallback parameter on Measure::bind, leaving every aggregator on the same direct-bind shape: bind() returns a BoundXxxHandle holding an Arc<TrackerEntry> that writes directly to the aggregator without per-call attribute lookup. LastValue, PrecomputedSum, and ExpoHistogram now match the Sum / Histogram pattern. PrecomputedSum's bind() is unreachable from user code today (async-only) but the impl exists so the trait is uniform and future Gauge / Observable bind() extensions are mechanical. The rare RwLock-poisoned case now yields a NoopBoundMeasure that drops measurements silently — mirroring measure()'s own poison handling. Test coverage added: - Bound ExponentialHistogram via View: delta with NaN/inf filter, bind-at-overflow attribution, persistence across delta eviction, view filter applied at bind time, and cumulative accumulation. - Cumulative bind-at-overflow tests for Counter and Histogram. - LastValue::bind and PrecomputedSum::bind unit tests exercising Measure / BoundMeasure traits and bound_count Drop semantics. - ValueMap::bind under a poisoned trackers RwLock returns None so the caller produces a NoopBoundMeasure rather than panicking.

cijothomas

LGTM!
Thanks for working on this. Spec PR is still pending, but give this is gated under feature-flag, okay to merge. I
ts aligned with current spec PR.

bryantbiggs requested a review from a team as a code owner March 14, 2026 20:36

bryantbiggs mentioned this pull request Mar 14, 2026

feat(metrics): add bound instruments behind experimental feature flag #3392

Closed

bryantbiggs force-pushed the feat/bound-instruments branch from db9517e to fb725ae Compare March 14, 2026 20:52

bryantbiggs force-pushed the feat/bound-instruments branch 3 times, most recently from f8a03dc to 08649e8 Compare March 15, 2026 00:42

cijothomas mentioned this pull request Apr 25, 2026

Bound instruments open-telemetry/opentelemetry-specification#5050

Open

bryantbiggs force-pushed the feat/bound-instruments branch from 264009b to e8cac00 Compare April 25, 2026 01:53