Flask instrumentation: http.server.active_requests gauge leaks on exception (missing try/finally)

[liyaka]

Description

The Flask instrumentation's _rewrapped_app function increments active_requests_counter before calling wsgi_app() but the decrement is not wrapped in try/finally. If an exception propagates out of wsgi_app(), the counter is never decremented, causing a permanent gauge leak.

This is particularly impactful when using this metric for Kubernetes HPA (Horizontal Pod Autoscaler) — leaked counters cause HPA to see phantom load and refuse to scale down.

Location

instrumentation/opentelemetry-instrumentation-flask/src/opentelemetry/instrumentation/flask/__init__.py

Lines 351–425 in _rewrapped_app:

active_requests_counter.add(1, active_requests_count_attrs)    # L351 - increment
# ...
result = wsgi_app(wrapped_app_environ, _start_response)        # L398 - if this raises...
# ...duration recording (L400-424)...
active_requests_counter.add(-1, active_requests_count_attrs)   # L425 - NEVER REACHED
return result

Expected behavior

The decrement should always execute, regardless of whether wsgi_app() raises an exception.

The WSGI instrumentation already has the correct pattern

instrumentation/opentelemetry-instrumentation-wsgi/src/opentelemetry/instrumentation/wsgi/__init__.py:

try:
    with trace.use_span(span):
        iterable = self.wsgi(environ, start_response)
        return _end_span_after_iterating(iterable, span, token)
except Exception as ex:
    raise
finally:
    self.active_requests_counter.add(-1, active_requests_count_attrs)  # always runs

Suggested fix

Wrap lines 398–425 in try/finally:

active_requests_counter.add(1, active_requests_count_attrs)
request_route = None
# ...
try:
    result = wsgi_app(wrapped_app_environ, _start_response)
    # ...duration histogram recording...
    return result
finally:
    active_requests_counter.add(-1, active_requests_count_attrs)

How to reproduce

1. Run a Flask app with opentelemetry-instrument and gunicorn
2. Send requests that trigger exceptions propagating past Flask's error handlers (e.g., OOM-killed gunicorn workers, or middleware errors)
3. Observe http.server.active_requests gauge — it increments but never decrements for crashed requests
4. The gauge stays permanently elevated until the process is restarted

Impact

• HPA scaling: Kubernetes HPA using this metric sees phantom active requests → refuses to scale down → wasted resources
• Monitoring: Dashboards show incorrect active request counts
• Alerting: False positives on active request alerts

Environment

• opentelemetry-instrumentation-flask==0.57b0 (also confirmed present on main / 0.62b0)
• gunicorn with --workers 1 --threads N --worker-class gthread
• OTEL SDK 1.36.0, OTLP HTTP exporter, delta temporality

[alliasgher]

Opened #4432 for this.