Bug: a NaN or infinite metric value makes the whole EMF line invalid JSON #8796

Expected Behaviour

add_metric should not accept a value CloudWatch can't store (nan, inf, -inf), the same way it rejects a value that isn't a number, so the other metrics in the same flush still get published. AWS's own aws-embedded-metrics-python raises Metric value must be finite for these.

Current Behaviour

The value check is isinstance(value, numbers.Number), which lets nan and ±inf through. At flush json.dumps writes them as bare NaN / Infinity, which isn't JSON, so the EMF line for the whole flush (every metric, the dimensions and the metadata) is no longer valid JSON. A nan from a division like total / count is enough to hit it.

I haven't sent one to CloudWatch; locally a strict JSON parser rejects the line. The same check is in metrics/base.py, used by single_metric.

Code snippet

import io, json, contextlib
from aws_lambda_powertools import Metrics
from aws_lambda_powertools.metrics import MetricUnit

metrics = Metrics(namespace="Shop", service="checkout")
metrics.add_metric(name="Orders", unit=MetricUnit.Count, value=1)
metrics.add_metric(name="AvgBasket", unit=MetricUnit.Count, value=float("nan"))

out = io.StringIO()
with contextlib.redirect_stdout(out):
    metrics.flush_metrics()

print(out.getvalue())
# {..."service":"checkout","Orders":[1.0],"AvgBasket":[NaN]}

def reject(constant):
    raise ValueError(constant)

json.loads(out.getvalue(), parse_constant=reject)  # ValueError: NaN

Steps to Reproduce

1. Add a metric with value float("nan") or float("inf") next to a normal metric.
2. Flush.
3. The printed EMF line contains NaN / Infinity and is not valid JSON.

Possible Solution

Check math.isfinite(value) next to the existing type check in both add_metric methods. Since ADR 0006 asks for a warning before raising on input that is accepted today, it could warn with PowertoolsUserWarning and skip the value for now, then raise MetricValueError in the next major. I can open a PR once you've had a look.

Did you run the reproduction yourself?

Yes, I ran it and saw the failure

Powertools for AWS Lambda (Python) version

latest (develop, 3.40.0)

AWS Lambda function runtime

3.12

Packaging format used

PyPi

Debugging logs

No response