openai: investigate metrics code relying on span recording

[xrmx]

It looks like our metrics code is relying on tracing being enabled, this is not the case if we run with OTEL_SDK_DISABLED=true.

Here's some tracebacks from running chatbot-rag-app with the sdk disabled:

api-frontend  | Traceback (most recent call last):
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/langchain_openai/chat_models/base.py", line 737, in _stream
api-frontend  |     for chunk in response:
api-frontend  |                  ^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/wrappers.py", line 122, in __next__
api-frontend  |     self.end(exc)
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/wrappers.py", line 78, in end
api-frontend  |     _record_operation_duration_metric(self.operation_duration_metric, self.span, self.start_time)
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/helpers.py", line 212, in _record_operation_duration_metric
api-frontend  |     operation_duration_metric_attrs = _get_attributes_if_set(
api-frontend  |                                       ^^^^^^^^^^^^^^^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/helpers.py", line 189, in _get_attributes_if_set
api-frontend  |     attributes = span.attributes
api-frontend  |                  ^^^^^^^^^^^^^^^
api-frontend  | AttributeError: 'NonRecordingSpan' object has no attribute 'attributes'. Did you mean: 'set_attributes'?
api-frontend  | 
api-frontend  | During handling of the above exception, another exception occurred:
api-frontend  | 
api-frontend  | Traceback (most recent call last):
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/werkzeug/serving.py", line 370, in run_wsgi
api-frontend  |     execute(self.server.app)
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/werkzeug/serving.py", line 333, in execute
api-frontend  |     for data in application_iter:
api-frontend  |                 ^^^^^^^^^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/werkzeug/wsgi.py", line 256, in __next__
api-frontend  |     return self._next()
api-frontend  |            ^^^^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/werkzeug/wrappers/response.py", line 32, in _iter_encoded
api-frontend  |     for item in iterable:
api-frontend  |                 ^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/flask/helpers.py", line 125, in generator
api-frontend  |     yield from gen
api-frontend  |   File "/app/api/chat.py", line 78, in ask_question
api-frontend  |     for chunk in llm.stream(qa_prompt):
api-frontend  |                  ^^^^^^^^^^^^^^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/langchain_core/language_models/chat_models.py", line 415, in stream
api-frontend  |     for chunk in self._stream(messages, stop=stop, **kwargs):
api-frontend  |                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/langchain_openai/chat_models/base.py", line 735, in _stream
api-frontend  |     with context_manager as response:
api-frontend  |          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/wrappers.py", line 108, in __exit__
api-frontend  |     self.end(exc_value)
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/wrappers.py", line 70, in end
api-frontend  |     _record_operation_duration_metric(self.operation_duration_metric, self.span, self.start_time)
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/helpers.py", line 212, in _record_operation_duration_metric
api-frontend  |     operation_duration_metric_attrs = _get_attributes_if_set(
api-frontend  |                                       ^^^^^^^^^^^^^^^^^^^^^^^
api-frontend  |   File "/usr/local/lib/python3.12/site-packages/opentelemetry/instrumentation/openai/helpers.py", line 189, in _get_attributes_if_set
api-frontend  |     attributes = span.attributes
api-frontend  |                  ^^^^^^^^^^^^^^^
api-frontend  | AttributeError: 'NonRecordingSpan' object has no attribute 'attributes'. Did you mean: 'set_attributes'?

[codefromthecrypt]

Raised #54, which works around this.

It is still curious (to me) why instrumentation is installed when OTEL_SDK_DISABLED=true. As a lay person, I would expect disabled to mean also don't load instrumentation, as instrumentation even when no-op can cause latency or errors like this.

[codefromthecrypt]

chatted with @anuraaga offline and it seems not unusual in otel for disabled to mean effectively unsampled. Basically, disabled still allows for propagation.

In that case #54 is all we need I think, that and tests in each new instrumentation to verify they work when "disabled"

[anuraaga]

A bit unrelated to what exactly does disabled mean, actually Span in the API (i.e. instrumentation) doesn't expose attributes, Span is mostly write-only until it's ended. Python's duck typing is allowing the .attributes to resolve when tracing is enabled (I don't know if introducing type checking with pylancing or something could catch it) but I guess it's just being lucky. Instead of or in addition to setting .span, I guess it should be setting the attributes themselves that were computed.

[xrmx]

My alternative version of the fix #56 where I remove the dependency from a recording span in the metrics code.