Record error.type on failed collections in PeriodicMetricReader

[TimurRakhmatullin86]

Problem

In PeriodicMetricReader.Scheduled.doRun() the self-observability sample for
otel.sdk.metric_reader.collection.duration is always recorded with a null error:

long startNanoTime = CLOCK.nanoTime();
String error = null;
Collection<MetricData> metricData;
try {
  metricData = collectionRegistration.collectAllMetrics();
} finally {
  long durationNanos = CLOCK.nanoTime() - startNanoTime;
  instrumentation.recordCollection(durationNanos / 1_000_000_000.0, error);
}

error is declared, initialized to null, and never assigned anywhere in the method, so
MetricReaderInstrumentation.recordCollection's if (error != null) branch that attaches the
error.type attribute is unreachable from this caller. When collectAllMetrics() throws, the
finally still records the sample — as a successful collection with no error.type — and
the exception propagates to the outer catch (Throwable). The self-observability histogram
therefore reports error-free collections while the reader is exporting nothing, which is
exactly the signal an operator would use to detect the outage.

The sibling BatchSpanProcessor.exportCurrentBatch() uses the same String error = null; … finally { instrumentation…(…, error); } shape but assigns error on each failure path.

Fix

Set error to the throwable's class name before it propagates:

try {
  metricData = collectionRegistration.collectAllMetrics();
} catch (Throwable t) {
  error = t.getClass().getName();
  throw t;
} finally {
  ...
}

The rethrow keeps the existing control flow — the exception still reaches the same outer
catch (Throwable) — so the only change is that the recorded sample now carries error.type.
(ThrowableUtil.propagateIfFatal is intentionally not added: the outer catch already handles
all throwables, so it would be a no-op here.)

Tests

SdkMeterProviderMetricsTest.collectionFailureIsRecordedWithErrorType: registers a
MetricProducer that throws on the first produce() then succeeds, flushes twice, and asserts
the exported otel.sdk.metric_reader.collection.duration point carries
error.type = java.lang.IllegalStateException. It fails against the current code (the point has
no error.type) and passes with the fix. Full :sdk:metrics:test suite and
:sdk:metrics:spotlessCheck pass; errorprone/NullAway clean.

[opentelemetry-pr-dashboard]

Pull request dashboard status

Status last refreshed: 2026-07-21 21:49:28 UTC.

• Waiting on: Reviewers
• Next step: Review the latest changes.

This automated status or its linked feedback items may be incorrect. If something looks wrong, report it with the result you expected.