Describe your environment

OS: Windows 11
Python version: 3.12.7
SDK version: 1.42.1
API version: 1.42.1

What happened?

When a NaN value is recorded on a metric instrument (e.g. counter.add(float("nan"))), it is silently folded into the running sum and permanently poisons that aggregation. Every subsequent export emits nan for that instrument, even if all future measurements are valid.

The root cause: SumAggregation guards against negative values with if amount < 0, but float("nan") < 0 evaluates to False in Python (all IEEE 754 comparisons against NaN return False), so NaN bypasses the guard and corrupts the running total.

Steps to Reproduce

import math
from opentelemetry import metrics
from opentelemetry.sdk.metrics import MeterProvider
from opentelemetry.sdk.metrics.export import InMemoryMetricReader

reader = InMemoryMetricReader()
provider = MeterProvider(metric_readers=[reader])
meter = provider.get_meter("test")

counter = meter.create_counter("test_counter")
counter.add(1)
counter.add(float("nan"))  # poisons the aggregation
counter.add(5)             # valid but sum is now nan

metrics_data = reader.get_metrics_data()
point = (
    metrics_data.resource_metrics[0]
    .scope_metrics[0]
    .metrics[0]
    .data.data_points[0]
)
print(f"sum = {point.value}")           # sum = nan
print(f"is nan = {math.isnan(point.value)}")  # is nan = True

Expected Result

The NaN measurement is dropped before aggregation. The counter continues accumulating valid measurements only. Final sum = 6.0.

Actual Result

sum = nan
is nan = True

The aggregation is permanently corrupted , all subsequent exports for this instrument emit nan with no recovery path.

Additional context

The Java SDK hit the exact same issue in open-telemetry/opentelemetry-java#5846 and resolved it in PR #5859 (merged Sep 2023) by explicitly dropping NaN measurements at the SDK level before any aggregation occurs.

The proposed fix for Python would be a math.isnan / not math.isfinite guard early in the aggregate() methods. one check per aggregation type, prevents permanent state corruption. Happy to implement this if the direction looks good.

Would you like to implement a fix?

Yes