[adityabavadekar]

Adds Enabled() to synchronous instruments, allowing callers to skip expensive work, such as building attributes, when nothing will be recorded. Follow-up to #4313, similar to Tracer::Enabled() (#3357). Spec

Fixes #4719

Changes

• Adds a non-virtual Enabled() to SynchronousInstrument (ABI v2 only).
• SDK instruments read their Meter's enabled flag, so a configurator update reaches all of them with one write. Instruments dropped by all matching views are always disabled.
• No-op instruments return false.
• Adds OPENTELEMETRY_ATOMIC_READ_RELAXED_8, since OPENTELEMETRY_ATOMIC_READ_8 locks the cache line on MSVC.
• Adds tests covering all 8 SDK instruments, runtime configuration updates, Drop views, and no-op instruments.

Benchmarks

Prototypes of each design with meter disabled, all threads sharing one counter. Each call checks Enabled() before building attributes and calling Add(), while 8 background threads export spans over OTLP. GCC 16.2 -O3, i7-1360P, median of 3 runs.

Reading the meter's enabled flag through a pointer is around ~3× faster than a virtual Enabled() call.

A per-instrument flag is similarly fast to read, but configuration updates cost O(n) instead of O(1). With 65,536 instruments, updates took 665 µs versus 11.5 ns with the pointer.

Design	1 thread (ns/call)	16 threads (ns/call)
No Enabled() (build attributes, then Add)	258	499
Virtual Enabled()	2.62	5.42
Per-instrument flag	0.95	1.50
Pointer to Meter flag (this PR)	0.76	1.65

For significant contributions please make sure you have completed the following items:

• CHANGELOG.md updated for non-trivial changes
• Unit tests have been added
• Changes in public API reviewed

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.73%. Comparing base (d4b49e3) to head (404fe1b).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4721      +/-   ##
==========================================
+ Coverage   86.63%   86.73%   +0.10%     
==========================================
  Files         527      527              
  Lines       20830    20858      +28     
==========================================
+ Hits        18045    18089      +44     
+ Misses       2785     2769      -16     

Files with missing lines	Coverage Δ
api/include/opentelemetry/metrics/noop.h	44.16% <100.00%> (+2.60%)	⬆️
...i/include/opentelemetry/metrics/sync_instruments.h	100.00% <100.00%> (ø)
sdk/include/opentelemetry/sdk/metrics/meter.h	62.50% <ø> (ø)
...de/opentelemetry/sdk/metrics/meter_enabled_state.h	100.00% <100.00%> (ø)
...telemetry/sdk/metrics/state/multi_metric_storage.h	92.31% <100.00%> (+0.16%)	⬆️
sdk/src/metrics/meter.cc	89.07% <100.00%> (+5.07%)	⬆️
sdk/src/metrics/sync_instruments.cc	66.12% <100.00%> (+0.66%)	⬆️

🚀 New features to boost your workflow:

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

[mateenali66]

checked the lifetime first. each sdk instrument holds the shared_ptr<MeterEnabledState> that owns its flag, so Enabled() stays valid after the provider is destroyed or the global provider is swapped. asan is clean on both, and on a dropped instrument. tsan is clean with 6 threads calling Enabled() and Add() while the configurator flips 4000 times. abi v1 layouts and vtables are unchanged, v2 only grows SynchronousInstrument by the pointer.

one optional thing. an instrument created from a meter whose provider is already gone has no storage and logs a warning on every Add(), but Enabled() returns true. storage == nullptr || storage->Empty() in GetInstrumentEnabledState would match what the Meter constructor already does without a context.

needs a rebase, only CHANGELOG conflicts. 1.30.0 shipped, so the entry moves under the new Unreleased.

[marcalff]

Proposal:

bool enabled_in_api_ = false;
const bool *enabled_ = &enabled_in_api_;

This:

• disable the flag by default, when no SDK is present
• avoids to call this->SetEnabledFlag(this->DisabledFlag()); from noop.h, since the flag is already disabled.
• avoid header only singletons like the static EnabledFlag() and DisabledFlag(), which are a problem in general.

For instruments implemented in an SDK, SetEnabledFlag() will work as before, pointing to the shared meter flag.

[lalitb]

@adityabavadekar - Could you also include the enabled case in the benchmark, comparing Add() directly with Enabled() followed by Add(), using the same attributes? The disabled results show the benefit clearly; this would show the overhead when metrics are being recorded normally.

[lalitb]

non-blocking - Could we mention that concurrent updates to this flag must use atomic writes? This would help other SDK implementations use it safely.

[lalitb]

LGTM overall. Could you add the enabled-case benchmark before merging, so we can see the overhead when metrics are being recorded normally?