Gunicorn with multiple workers breaks metrics

[abstractOwl]

Describe your environment Python 3.8.5 opentelemetry-distro[otlp]==0.44b.0 Gunicorn 22.0.0

Steps to reproduce

1. Run a Gunicorn Flask app as described in the documentation, specifying multiple workers.
2. Visit a route that emits metrics several times
3. View metrics in metrics portal

What is the expected behavior? Counter will increase monotonically for each visit to the route

What is the actual behavior? Counter metric jumps between values from each worker because meter state is not synced between workers.

Example from console exporter (the worker.id resource attribute is set differently for each worker):

{
    "resource_metrics": [
        {
            "resource": {
                "attributes": {
                    "telemetry.sdk.language": "python",
                    "telemetry.sdk.name": "opentelemetry",
                    "telemetry.sdk.version": "1.23.0",
                    "service.instance.id": "localhost",
                    "worker.id": "885db",
                    "service.name": "otel-demo",
                    "telemetry.auto.version": "0.44b0"
                },
[…]
                    "metrics": [
                        {
                            "name": "test.counter",
                            "description": "",
                            "unit": "",
                            "data": {
                                "data_points": [
                                    {
                                        "attributes": {},
                                        "start_time_unix_nano": 1714162262042565771,
                                        "time_unix_nano": 1714162322001856183,
                                        "value": 11
                                    }
                                ],
                                "aggregation_temporality": 2,
                                "is_monotonic": true
                            }
                        }
                    ],

{
    "resource_metrics": [
        {
            "resource": {
                "attributes": {
                    "telemetry.sdk.language": "python",
                    "telemetry.sdk.name": "opentelemetry",
                    "telemetry.sdk.version": "1.23.0",
                    "service.instance.id": "localhost",
                    "worker.id": "818d7",
                    "service.name": "otel-demo",
                    "telemetry.auto.version": "0.44b0"
                },
                "schema_url": ""
            },

[…]

                    "metrics": [
                        {
                            "name": "test.counter",
                            "description": "",
                            "unit": "",
                            "data": {
                                "data_points": [
                                    {
                                        "attributes": {},
                                        "start_time_unix_nano": 1714162262869065166,
                                        "time_unix_nano": 1714162322111501132,
                                        "value": 12
                                    }
                                ],
                                "aggregation_temporality": 2,
                                "is_monotonic": true
                            }
                        }
                    ],

Additional context This issue is caused by the way each Gunicorn worker needs to be instrumented separately due to the fork process model. I think most developers would expect that multiple workers on the same host be treated as the same instance when it comes to metrics.

A hacky workaround for this would be to use [hostname]:[pid] as service.instance.id and then aggregate on a separate service hostname resource attribute in your metrics portal, but that doesn't seem like a great solution.

[abstractOwl]

I see that this issue was referenced previously in this closed issue: https://github.com/open-telemetry/opentelemetry-python/issues/93

[ocelotl]

@abstractOwl appending the pid could be a valid solution, curious why you think it is a hacky approach to this issue?

[abstractOwl]

@ocelotl I'd prefer a solution where we didn't have to add an extra filter to all of our queries - it breaks many of the default metrics views/generated dashboards that are offered by observability vendors and is overall an inconvenience.

[bowenng]

Any thoughts on adding aggregation capabilities in the Prometheus exportor in the OTEL Collector?

I observed that as long as each forked process has its own resource specified, the "debug" exporter actually prints out the correct metrics generated per resource (I.e. per process). It seems to be theoretically possible to aggregate among all the resources in the Prometheus exporter.

I think the issue is that the Prometheus exporter is ignoring the resource and is not doing any aggregation.

[xrmx]

Had a discussion internally about this and given this excerpt from note [1] from https://opentelemetry.io/docs/specs/semconv/attributes-registry/service/:

For applications running behind an application server (like unicorn), we do not recommend using one identifier for all processes participating in the application.
Instead, it’s recommended each division (e.g. a worker thread in unicorn) to have its own instance.id.

Shouldn't we add a way to set a sensible service.instance.id by default?

Other sdks:

• java, set it by default: https://github.com/open-telemetry/opentelemetry-java/issues/5103
• js, set it by default: https://github.com/open-telemetry/opentelemetry-js/pull/4626