fix(flask): wrap wsgi_app call in try/except to prevent active_requests gauge leak

[alliasgher]

Description

Wrap the wsgi_app call in try/finally so that http.server.active_requests is decremented when the wrapped WSGI app raises. Without this, every exception leaves the gauge permanently above the real number of in-flight requests, causing steady drift upward over the process lifetime.

Fixes #4403

Type of change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

Added test_active_requests_decremented_on_error in test_programmatic.py which drives a handler that raises and asserts the gauge ends at 0.

Checklist

• Changelog entry added under Unreleased / Fixed

[liyaka]

Thanks for the quick fix! We hit this exact bug in production — the leaked gauge was preventing our Kubernetes HPA from scaling down.

One suggestion: consider using try/finally instead of try/except to keep the decrement in a single code path. With the current approach, active_requests_counter.add(-1, ...) exists in two places (the except block and line 458), which is easy to get out of sync during future refactors.

The WSGI instrumentation already uses this pattern:

# wsgi/__init__.py
try:
    ...
    return _end_span_after_iterating(iterable, span, token)
except Exception as ex:
    raise
finally:
    self.active_requests_counter.add(-1, active_requests_count_attrs)

For Flask it would look like:

try:
    result = wsgi_app(wrapped_app_environ, _start_response)
    # ...duration recording unchanged...
    return result
finally:
    active_requests_counter.add(-1, active_requests_count_attrs)

This also removes the original decrement at line 458, so there's exactly one decrement path.

Also worth adding a test to prevent regression — I have one in #4437 (test_active_requests_counter_decremented_on_error) that could be adapted here.

[alliasgher]

Good catch — updated to use try/finally so the decrement is in a single code path regardless of whether wsgi_app succeeds or raises @liyaka

[alliasgher]

Added the regression test test_active_requests_counter_decremented_on_error. It hits /hello/500 (which raises ValueError internally), collects http.server.active_requests, and asserts the value is back to 0 after both requests complete.

[MikeGoldsmith]

Thanks for this fix @alliasgher - looks good to me.

[liyaka]

Any estimation for this fix to be released?
thank you!

[alliasgher]

Pushed bb17ad0 to fix broken links in CHANGELOG.md that were failing check-links CI.