Prometheus exporter puts label values under the wrong label names when two instruments share a name

[jbyers44]

Describe your environment

OS: macOS
Python version: 3.10.18
Package version: opentelemetry-sdk 1.44.0, opentelemetry-exporter-prometheus 0.65b0.

The issue also reproduces on main as of today.

What happened?

The exporter puts label values under the wrong names when two instruments from different meters share a name, unit, and description but not attribute keys. The family id is name|description|unit, so both go into one family. _get_or_create_family returns the existing family, which has the first metric's label names. The second metric's values go into those names by position.

This regression started with #4869, which removed label keys from the family id to fix duplicate HELP/TYPE lines (#4868). 1.39.1 / 0.60b1 exports both series correctly. 1.40.0 / 0.61b0 mislabels the second one.

We hit this bug with opentelemetry-instrumentation-httpx and opentelemetry-instrumentation-requests. Both emit http.client.duration with the same unit and description and different keys. The result is series like http_method="200".

Steps to Reproduce

from opentelemetry.exporter.prometheus import PrometheusMetricReader
from opentelemetry.sdk.metrics import MeterProvider
from prometheus_client import REGISTRY, generate_latest

provider = MeterProvider(metric_readers=[PrometheusMetricReader()])

first = provider.get_meter("scope.first").create_histogram("request.duration", unit="ms", description="Duration.")
second = provider.get_meter("scope.second").create_histogram("request.duration", unit="ms", description="Duration.")

first.record(1, {"method": "GET", "status": "200"})
second.record(1, {"host": "example.com", "method": "POST", "status": "500"})

for line in generate_latest(REGISTRY).decode().splitlines():
    if line.startswith("request_duration_milliseconds_count"):
        print(line)

Expected Result

Each value under its own label, with one HELP and one TYPE line:

request_duration_milliseconds_count{host="",method="GET",otel_scope_name="scope.first",otel_scope_schema_url="",otel_scope_version="",status="200"} 1.0
request_duration_milliseconds_count{host="example.com",method="POST",otel_scope_name="scope.second",otel_scope_schema_url="",otel_scope_version="",status="500"} 1.0

Actual Result

request_duration_milliseconds_count{method="GET",otel_scope_name="scope.first",otel_scope_schema_url="",otel_scope_version="",status="200"} 1.0
request_duration_milliseconds_count{method="example.com",otel_scope_name="POST",otel_scope_schema_url="scope.second",otel_scope_version="",status=""} 1.0

The series has no host label, method gets the host, otel_scope_name gets the method, and status is empty. If second records first, the second series is correct and the exporter mislabels the first series.

Additional context

#4413 fixed the same kind of bug for points with different attribute sets within one instrument. Its test, test_multiple_data_points_with_different_label_sets, still passes, since every point there comes from one metric.

Would you like to implement a fix?

Yes

Tip

_{React with 👍 to help prioritize this issue. Please use comments to provide useful context, avoiding +1 or me too, to help us triage it. Learn more here.}