opentelemetry_wrapper

a wrapper around opentelemetry and opentelemetry-instrumentation-* to make life a bit easier

design principles

safe

• never crash the application
  • fallback to non-opentelemetry if necessary
  • ignore input over raising an exception
  • possible exception: it may be better to fail at startup than to run with known bad config
    • https://opentelemetry.io/docs/specs/otel/error-handling/#basic-error-handling-principles
  • make dangerous things impossible (by design) > difficult (guardrails) > unusual (defaults) > understood (docs)
    • there should be a "best practice" somewhere in that statement
• hard to get wrong
  • idempotent, even when you instrument the same thing in different ways from different places
  • "be conservative in what you send, be liberal in what you accept"
  • note: easy and simple mean different things

unsurprising

• simple, idiomatic, succinct, and pretty
  • decorators over wrappers
  • wrappers over context managers
  • context mangers over manually managing spans anything else (note: this is still a todo)
• reasonable documented defaults
  • magic may be hard to understand, but it's better than being irritating
• emits zero logs/spans (of its own) at runtime
  • fail silently over flailing noisily
  • drowning out real logs can be worse than being useless (e.g., you could crash fluentd - ask me how I know)

helpful

• machine-readable first, human-readable a close second
  • no newlines in logs or spans
  • no whitespace-delimited ambiguity
  • json all the things (within reason)
• provide all available application context
  • we want to know all we can about what's going on and where
  • code introspection and runtime analysis if we can make it fast enough

usage

TL;DR:

1. call instrument_all() to instrument logging and requests
2. instrument your FastAPI app using instrument_fastapi_app(FastAPI(...))
3. use @instrument_decorate on any function or class you want to monitor
4. set OTEL_WRAPPER_DISABLED=true in your CICD tests, especially if you're using pytest

env vars

Note:

The service.name and service.namespace can also be set via OTEL_RESOURCE_ATTRIBUTES, but any settings there will be overwritten by OTEL_SERVICE_NAME and OTEL_SERVICE_NAMESPACE. For more details, read the original OpenTelemetry docs

If you don't have a serviceAccount mounted in your pod (i.e. automountServiceAccountToken: false), then you should manually set OTEL_SERVICE_NAMESPACE using the following yaml:

 env:
   - name: OTEL_SERVICE_NAMESPACE
     valueFrom:
       fieldRef:
         fieldPath: metadata.namespace

K8s docs

@instrument_decorate decorator for functions and classes

• decorating a function creates a span whenever the function is called
  • the span name is set to the function or class name, and attributes are added for the filename, line number, etc
• also works for async functions
• also works for classes and dataclasses
  • creates a span for every method call, including init (or post_init), new, and call
  • creates a span for every property get, set, or delete
  • if this is too verbose, feel free to decorate specific methods in the class instead
• it is not recommended to use this on functions that are called thousands of times or classes you create thousands of instances for (e.g. pydantic classes, usually), since it can be excessively noisy

from opentelemetry_wrapper import instrument_decorate


@instrument_decorate
def square(x):
    return x * x


@instrument_decorate
async def cube(x):
    return x * x * x


@instrument_decorate
class Thing:
    def __init__(self):
        self.x = 1

instrumenting the builtin logging module

• sets a root logger handler (or more than one) that can output logs to the console or to a file path
• logs are output as json by default to make them easier to work with downstream, set print_json=False to disable
  • the verbose parameter has no effect on json output, which is maximally verbose
  • otherwise, it defaults to True - set to False to print less text to the console
• the log level can be specified via the level arg, but defaults to whatever was set via the OTEL_LOG_LEVEL env var

import logging
from opentelemetry_wrapper import instrument_logging

instrument_logging()

logging.warning('...')

instrumenting fastapi

• Warning: using opentelemetry.instrumentation.fastapi.FastAPIInstrumentor does not work as intended
• use instrument_fastapi_app, for example:

from fastapi import FastAPI
from opentelemetry_wrapper import instrument_fastapi_app

app = instrument_fastapi_app(FastAPI(...))

Capturing headers:

• to avoid capturing any headers as spans, set a blank header, like OTEL_HEADER_ATTRIBUTES=,
• OTEL_HEADER_ATTRIBUTES will attempt to parse base64 and json, and will flatten the data structure
• OTEL_INSTRUMENTATION_HTTP_CAPTURE_HEADERS_SERVER_REQUEST behaves similarly, but does not flatten the structure

prometheus

• set (for example) OTEL_EXPORTER_PROMETHEUS_ENDPOINT=/metrics
• remember to whitelist /metrics in OPA
• repeated and trailing slashes are normalized (e.g. ///metrics/ -> /metrics)
• if you set OTEL_EXPORTER_PROMETHEUS_ENDPOINT=/metrics/, a redirect will be set up from /metrics to /metrics/
• if you set the port instead, the created wsgi app accepts any endpoint
• if you set up both, the endpoint will be created in fastapi and a separate wsgi app will also be created
• the /graph endpoint is not available, you'll need to actually run prometheus to get that

features

• Make instrumentation (more) idempotent:
  • you can call the instrument functions unlimited times from multiple places in your codebase, and it'll work the same
  • e.g., a class definition, a defined method of the class, a class instance, and a method from the instance
• Make re-instrumentation of logging actually work when passing in a new format string
• Make logging print as a one-line JSON dict by default, with a lot of magic to convert stuff to valid json
• logs and spans contain info about which thread / process and which file / function / line of code it came from
  • and the k8s namespace and pod, if applicable, otherwise the local pc name
• Provide support for decorating functions and classes
• Provide support for instrumentation of dataclasses
  • NOTE: Global instrumentation needs to be run before any dataclasses are initialized
  • Otherwise, use the decorator on each class as usual (since it is idempotent anyway)
• Add global instrumentation of FastAPI
  • sometimes works even after apps are created for some reason, likely due to how Uvicorn runs in a new process
  • but somehow sometimes doesn't work in prod, for equally unknown reasons
  • probably best to instrument each app instance
• Logs OIDC http headers as span attributes for FastAPI
• Creates OTLP exporters if specific env vars (below) are set
  • Pushes logs, metrics, and traces to the OTEL endpoint, if configured
  • Note: only logs and traces are printed to console, metrics are too noisy