Observability Best Practices when running FastAPI in a Lambda

Serverless makes it incredibly easy to get an API up and running in record time. But how do you achieve a high level of observability in a distributed architecture? In this post, we will be looking at instrumenting logging, metrics, and tracing capabil…

Serverless makes it incredibly easy to get an API up and running in record time. But how do you achieve a high level of observability in a distributed architecture? In this post, we will be looking at instrumenting logging, metrics, and tracing capabilities for a FastAPI application using AWS Lambda Powertools for Python.

This guide assumes that you know how to build and deploy a SAM application using the AWS SAM CLI. If you need a refresher on that, please refer to my previous blog post.

0. Table of Contents

1. What is observability?
2. Sample FastAPI Application
3. Checking the baseline
4. AWS Lambda Powertools Python
5. Logging
6. Metrics
7. Tracing
8. Conclusion

1. What is observability?

The rise of cloud and serverless computing has made a big impact on how we build, scale, and manage applications. Developers can now independently build and deploy applications with unmatched agility and speed. With distributed applications becoming more and more common, we must take care to not lose observability in our systems and services. Before long, we will get to a point where a single request traverses our architecture through any number of functions, containers, queues, and other services.

By implementing best practices when it comes to logging and metrics, as well as being able to trace a request through the entire system, we can gain a better understanding of how our applications are performing. The goal of observability is to instrument our applications so that when an issue occurs somewhere in our distributed system, we can quickly locate the root cause of the issue. Good observability also gives great insights into how our applications and services are being used, valuable insights that can guide future business decisions.

Observability can be achieved in several ways, and in this article you will learn how to instrument a FastAPI application, running inside a Lambda function, using the AWS Lambda Powertools for Python library. We will focus on implementing three different capabilities in our application.

  1. Structured Logging where we add contextual information to logs, such as request and correlation IDs, Lambda and FastAPI context, service name, exception information, and more.
  2. Metrics so that we can track how our application is used and how it is performing.
  3. Tracing so that we can trace requests as they travel through our systems.

2. Sample FastAPI Application

The application architecture

In this example we will use a simple FastAPI application, a Pets API that allows you to add, remove, list, fetch and update pets that are stored in DynamoDB. We will use a REST API Gateway to act as a proxy in front of the application.

As shown in the AWS Docs, HTTP APIs do not (at the time of writing) support tracing with AWS X-ray, which is why we have to use a REST API instead.

Example files

The required files to follow along in this example are available on GitHub, or here on my blog.

3. Checking the baseline

Before we add more advanced features to the application, let’s deploy the API in its current state to establish a baseline. While testing the API, I have used this Python script. The script uses boto3 to fetch the API URL from the CloudFormation Stack outputs. It then runs a sequence of requests towards the API, touching all the endpoints. You could also issue requests with cURL, through the Swagger UI available at https://API_GATEWAY_URL/STAGE_NAME/docs, or with something like Postman. The choice is yours.

After making a few requests, navigate to the CloudWatch console and go to Logs Insights. You should see a log group called /aws/lambda/FUNCTION_NAME in the Select log group(s) dropdown.

CloudWatch logs without any bells and whistles

The resulting log records contain almost no useful information, which is because right now we are only using simple print(...) statements in the code. Let’s change that, shall we?

4. AWS Lambda Powertools Python

Lambda Powertools Python is a library that comes fully packed with utilities that make it easy to adopt best practices when it comes to the observability of AWS Lambda functions. We will focus on the three core utilities, which are: Logger, Metrics and Tracer.

First, in requirements.txt, add the library.

mangum
fastapi
boto3
pydantic
+ aws-lambda-powertools

In the application code, you will need to import and initialize the three utilities. In a new file example/src/app/utils.py, add the following:

from aws_lambda_powertools import Logger, Metrics, Tracer
from aws_lambda_powertools.metrics import MetricUnit  # noqa: F401


logger: Logger = Logger()
metrics: Metrics = Metrics()
tracer: Tracer = Tracer()

5. Logging

Let’s start by adding some structured logging to the application. We will use the Logger utility from the Lambda Powertools library for this.

Upgrade from print statements

In the dynamo.py file, add from .utils import logger and change all the print(...) statements to logger.info(...). After a few requests, the experience in CloudWatch should have gotten a tad better.

CloudWatch logs with some bells and whistles

Here we can see the power of structured logging. CloudWatch has identified fields such as level, location, message, service, and so on. We can now use these fields to filter queries. If we wanted to see only error logs, we could add a filter to the query:

fields @timestamp, message
| sort @timestamp desc
| limit 200
| filter level = "ERROR"

What is that service_undefined value in the service field you ask? That’s right, we forgot one thing in the SAM template. We can control what service name Lambda Powertools uses by setting environment variables in a function. If you have a lot of different services, being able to easily filter logs by service name will be critical when you are troubleshooting a problem in production.

Function:
  Type: AWS::Serverless::Function
  Properties:
    ...
    Environment:
      Variables:
+       POWERTOOLS_SERVICE_NAME: FastAPIPowerToolsExample
+       POWERTOOLS_METRICS_NAMESPACE: FastAPINamespace
        TABLE_NAME: !Ref Table
    ...

That wasn’t so hard right? Just a few lines of code, and we now have nicely structured logs in the FastAPI application. Let’s explore some more features in the Lambda Powertools Logger.

Lambda context

Powertools makes it easy to add information from the Lambda context to logs with the inject_lambda_context decorator, like this:

@logger.inject_lambda_context
def handler(event, context):
    ...

But we do not have a handler function, do we? We have a Mangum object wrapping the FastAPI application. Luckily, the Mangum object acts as a handler function, so we can just add the following in example/src/app/__init__.py:

from .utils import logger  # add this import

...

handler = Mangum(app)
# add this line
handler = logger.inject_lambda_context(handler, clear_state=True)

Here, we set the parameter clear_state=True to clear the state on each invocation. This is useful if you want to ensure that the logs are not polluted with the state from previous invocations.

So, one more line of code. What did that line give us in the CloudWatch console?

CloudWatch logs with Lambda context

Some information about the Lambda context, such as function name, memory configuration, as well as whether the invocation required a cold start or not. Nice!

Correlation ID

Correlation IDs can be used to uniquely identify a request as it traverses multiple services. This is a key component to being able to correlate logs from different services in distributed systems.

In our example, the API will accept an optional X-Correlation-Id header, and if it is not present, it will use the request ID from the Lambda context. The correlation ID can then be added to requests towards downstream services (if any), to be able to get a complete view of the request flow. The Logger utility includes two helper functions to handle correlation IDs, logger.set_correlation_id("ID") and logger.get_correlation_id().

Since we want to extract the correlation ID in every request, as well as return it in every response, we will implement this using a FastAPI middleware.

In example/src/app/__init__.py, add the following:

from fastapi import FastAPI, HTTPException, Request
...

@app.middleware("http")
async def add_correlation_id(request: Request, call_next):
    # Get correlation id from X-Correlation-Id header
    corr_id = request.headers.get("x-correlation-id")
    if not corr_id:
        # If empty, use request id from aws context
        corr_id = request.scope["aws.context"].aws_request_id

    # Add correlation id to logs
    logger.set_correlation_id(corr_id)

    response = await call_next(request)

    # Return correlation header in response
    response.headers["X-Correlation-Id"] = corr_id
    return response

Hit the API with a few requests. If you supply a X-Correlation-Id header in the request, you should see the same X-Correlation-Id in the response. If you do not supply one, one will be generated for you. If the application makes further requests to downstream services, the correlation ID could be retrieved with logger.get_correlation_id() and passed on to the downstream service. Navigate to CloudWatch and take a look at the shiny new correlation_id field.

CloudWatch logs with correlation ID

FastAPI context

Let’s add some information about the FastAPI request context to the logs. On the top of my head, it would be nice to be able to filter logs by the request path, request route, and the request method. Since we have the Request object in the middleware created above, couldn’t we add this functionality there? We can’t. This is because middlewares are executed before any routing is made in the FastAPI application. So if we want to add the routes as they are declared in the application, such as /pets/{pet_id}, we need to use a custom APIRoute class.

Add a new file, example/src/app/router.py, and add the following:

from fastapi import Request, Response
from fastapi.routing import APIRoute
from typing import Callable
from .utils import logger


class LoggerRouteHandler(APIRoute):
    def get_route_handler(self) -> Callable:
        original_route_handler = super().get_route_handler()

        async def route_handler(request: Request) -> Response:
            # Add fastapi context to logs
            ctx = {
                "path": request.url.path,
                "route": self.path,
                "method": request.method,
            }
            logger.append_keys(fastapi=ctx)
            logger.info("Received request")

            return await original_route_handler(request)

        return route_handler

Here we add a few fields to the logger so that e.g. every future logger.info(...) call includes those fields. If you do not fully understand how the custom APIRoute works, please read the FastAPI documentation. He explains it a lot better than I possibly could.

Now, to use the custom APIRoute class, we need to add the following in example/src/app/__init__.py:

...
from .router import LoggerRouteHandler  # add this import


app = FastAPI()
# add this line
app.router.route_class = LoggerRouteHandler

...

Let’s see what we have in CloudWatch this time.

CloudWatch logs with FastAPI context

Look at that beauty. We now have filterable fields for the method, route, and actual path. Want to get all logs that occurred when someone fetched a specific pet with ID 123456? Simply add a filter to the query like | filter fastapi.method = "GET" and fastapi.path = "pets/123456". Want to see all logs from all calls to the DELETE /pets/{pet_id} route? You know what to do.

Exceptions

The Logger utility can also help us understand and debug errors easier by logging exceptions with logger.exception(...). To illustrate how Lambda Powertools can help us here, let us first add an endpoint in which we inject an error. Add the following code to example/src/app/__init__.py:

@app.get("/fail")
def fail():
    some_dict = {}
    return some_dict["missing_key"]

Call the endpoint, you should receive a 500 Internal Server Error. Head over to CloudWatch and see how the exception was logged.

CloudWatch logs with exception without any filterable keys

Looks like we are missing a lot of information here, such as correlation ID, route, service, and all the other fields we added previously. Let’s fix that, by adding a custom exception handler.

Creating an exception handler that catches all exceptions does not seem to be fully supported. I’ve tried just creating an exception handler for exceptions of type Exception, and it seems to enter the handler fine and the logging works. However, it does not catch the exception fully, so the exception is still propagated and finally caught in the Mangum event loop. It also does not enter the middleware that sets the correlation ID in the response.

The issue seems to be resolved by adding a starlette.exceptions.ExceptionMiddleware, though I’m unsure if this can have any unforeseen side effects. Use at your own risk! 🙂

Further information on FastAPI GitHub Issue and Starlette GitHub Issue:

In example/src/app/__init__.py, add the following:

from fastapi import FastAPI, HTTPException
# add the following imports
from fastapi.responses import JSONResponse
from starlette.exceptions import ExceptionMiddleware

...

app = FastAPI()
app.router.route_class = LoggerRouteHandler
# add the below code
# +++
app.add_middleware(ExceptionMiddleware, handlers=app.exception_handlers)


@app.exception_handler(Exception)
async def unhandled_exception_handler(request, err):
    logger.exception("Unhandled exception")
    return JSONResponse(status_code=500, content={"detail": "Internal Server Error"})

# +++

Call the /fail endpoint again and head over to CloudWatch.

CloudWatch logs with exception and filterable keys

Now we have a lot of relevant information in logs whenever an unhandled exception occurs, which makes it easy to find what caused a request to fail if you for example have the correlation ID at hand.

The logs are now a lot more useful, and we did not need to add that much code to the application. Most of the code resides in either middleware, exception handlers, or the custom APIRoute class. The changes to the actual business logic have been minimal, which shows that Lambda Powertools can easily be added to existing applications to improve logging practices.

6. Metrics

Let’s explore the next core utility in Lambda Powertools, the Metrics utility. This utility lets you easily push metrics to CloudWatch by taking care of all the necessary boilerplate. It works asynchronously by using Amazon CloudWatch Embedded Metrics Format, by logging the metrics to stdout. It also aggregates all metrics from each invocation to save on the number of calls to CloudWatch.

There is some terminology to be aware of here. CloudWatch metrics are grouped in containers called namespaces. If an application comprises multiple services, you could for example use the same namespace for all of them to group all the metrics for that application. Metrics can also have dimensions, which are key-value pairs added as metadata to metrics which allows you to filter and aggregate metrics depending on the dimension values.

The default configuration for the Metrics utility utilizes the two environment variables POWERTOOLS_METRICS_NAMESPACE and POWERTOOLS_SERVICE_NAME, where the former specifies the metric namespace and the latter adds a dimension service=POWERTOOLS_SERVICE_NAME to all metrics.

Instrumenting

If you followed along with the guide so far you should already have a metrics object in example/src/app/utils.py. There we used the default configuration, which sets the namespace and service dimension from environment variables. You could also specify those explicitly, like this:

from aws_lambda_powertools import Metrics

metrics = Metrics() # Sets metric namespace and service dimension via environment variables
metrics = Metrics(namespace="Namespace", service="ServiceA") # Sets metric namespace to "Namespace" and service dimension to "ServiceA"

If you want to add another dimension to all metrics, such as environment, you can do so with the set_default_dimensions method:

from aws_lambda_powertools import Metrics

metrics = Metrics()
metrics.set_default_dimensions(environment="development")


Now, to enable the functionality in the application, add the following to example/src/app/__init__.py:

...
# Add metrics to the following import
from .utils import logger, metrics

...

handler = Mangum(app)
handler = logger.inject_lambda_context(handler, clear_state=True)
# Add the following (last to properly flush metrics):
handler = metrics.log_metrics(handler, capture_cold_start_metric=True)

Make sure that the log_metrics decorator is added last to the handler function.

Adding metrics

Let’s add the following metrics to the application:

  • Number of times a pet is created
  • Number of times an unhandled exception occurs

Created pets metric

To add a metric that counts how many times a pet is created, we can add the following to the POST /pets route. We add it after the call to DynamoDB, to only record a metric if the pet was created successfully.

@app.post("/pets", status_code=201, response_model=models.PetResponse)
def post_pet(payload: models.CreatePayload):
    res = dynamo.create_pet(kind=payload.kind, name=payload.name)
    # Add the following line
    metrics.add_metric(name="CreatedPets", unit=MetricUnit.Count, value=1)
    return res

Unhandled exception metric

To count the number of unhandled exceptions, we can add the following to the exception handler:

@app.exception_handler(Exception)
async def validation_exception_handler(request, err):
    # add the following line
    metrics.add_metric(name="UnhandledExceptions", unit=MetricUnit.Count, value=1)
    logger.exception("Unhandled exception")
    return JSONResponse(status_code=500, content={"detail": "Internal Server Error"})

Using different dimensions

When using CloudWatch EMF, all metrics in a document must have the same dimensions. So, all calls to add_metric will generate metrics with identical dimensions, service plus any additional default dimensions you’ve added. To add different dimensions for specific metrics, we need to use single_metric. We might for example want to have a metric named RequestCount, that has both service and route dimensions.

First, in example/src/app/utils.py, update the following import to include single_metric:

from aws_lambda_powertools import Logger, Metrics, Tracer
# Add single_metric to the following import
from aws_lambda_powertools.metrics import MetricUnit, single_metric  # noqa: F401

Then, in example/src/app/router.py, add the following code:

from fastapi import Request, Response
from fastapi.routing import APIRoute
from typing import Callable
# Add single_metric to the following import
from .utils import metrics, MetricUnit, single_metric


class LoggerRouteHandler(APIRoute):
    def get_route_handler(self) -> Callable:
        original_route_handler = super().get_route_handler()

        async def route_handler(request: Request) -> Response:
            # Add fastapi context to logs
            ctx = {
                "path": request.url.path,
                "route": self.path,
                "method": request.method,
            }
            logger.append_keys(fastapi=ctx)
            logger.info("Received request")

            # Add the following:
            # +++
            # Add count metric with method + route as dimension
            with single_metric(name="RequestCount", unit=MetricUnit.Count, value=1) as metric:
                metric.add_dimension(name="route", value=f"{request.method} {self.path}")
            # +++

            return await original_route_handler(request)

        return route_handler

CloudWatch showtime

Let’s hit the API a few times on different routes, as well as the /fail endpoint. Then, navigate to CloudWatch and click on All metrics on the left-hand side. Then, click on the Namespace you specified. You should see the following.

CloudWatch metrics by dimensions

First, click on function_name, service. Here you should see a metric that counts the number of times a function has been subject to a cold start which is automatically added by the Lambda Powertools metrics utility when you add the capture_cold_start_metric=True parameter.

CloudWatch metrics with number of cold starts

Next, click on service under the namespace. Here you will be able to see the number of pets that have been created, as well as the number of unhandled exceptions.

CloudWatch metrics with exceptions and created pets

Finally, check out route, service. Here you should have metrics describing the total amount of requests to each route.

CloudWatch metrics with route calls

This showcases how easy it is to add metrics to a FastAPI Lambda application when using Lambda Powertools to take care of the heavy lifting.

7. Tracing

Now, let’s move on to the final core utility, the Tracer. The utility makes it easy to instrument your Lambda function to capture traces and send them to AWS X-Ray. AWS X-Ray is a distributed tracing system that can help you analyze and debug distributed applications, and provides a way to track requests as they travel through your different services.

Powertools also has the convenience of auto-patching modules supported by X-Ray, such as boto3. This will automatically create segments in your traces whenever you use boto3 to call DynamoDB or another AWS service from within your Lambda function, and likewise for other supported modules such as requests.

Instrumenting

First, we need to enable tracing on the Lambda function and API Gateway. In the SAM template, add the following:

Function:
  Type: AWS::Serverless::Function
  Properties:
+   Tracing: Active
    MemorySize: 128
...

RestApi:
  Type: AWS::Serverless::Api
  Properties:
    StageName: Prod
+   TracingEnabled: true

Then, as with the other utilities, we need to decorate our handler. Add the following to example/src/app/__init__.py:

...
# Add tracer to the following import
from .utils import tracer, logger, metrics, MetricUnit

...

handler = Mangum(app)

# Add the following lines
# +++
handler.__name__ = "handler"  # tracer requires __name__ to be set
handler = tracer.capture_lambda_handler(handler)
# +++

handler = logger.inject_lambda_context(handler, clear_state=True)
handler = metrics.log_metrics(handler, capture_cold_start_metric=True)

If you want to measure a specific method call and add it to your traces, you can decorate functions with the capture_method decorator. For illustrative purposes, add the following to example/src/app/dynamo.py:

...
# Add the following import
from .utils import tracer

...

@tracer.capture_method  # add this
def create_pet(kind: str, name: str) -> dict:
...

@tracer.capture_method  # and this
def get_pet(pet_id: str) -> dict:
...

@tracer.capture_method  # this as well
def update_pet(pet_id: str, kind: str = None, name: str = None):
...

@tracer.capture_method  # and this
def list_pets(next_token: str = None) -> dict:
...

@tracer.capture_method # and finally this
def delete_pet(pet_id: str):
...

Deploy the API, and send a couple of request to it. Mix it up a little and hit the /fail endpoint a few times, as well as try to generate some 4xx responses by trying to access non-existent pets, or create pets with invalid payloads. Go to the CloudWatch console and click on Service Map on the left-hand side.

AWS X-Ray service map

Here you can see that AWS X-Ray generated a service map of our entire architecture. We can see that a client has made a request to the API Gateway, which forwarded the request to the Lambda service, which in turn invoked a Lambda function, which then made further requests to the DynamoDB Table. As the system grows, the service map will be an invaluable tool to get a full picture of the entire system. For each node in the map, it is also possible to see the percentage of successful requests, average latency, and other useful metrics.

You can also drill down into individual traces. Below you can see timeline for a request to create a pet.

AWS X-Ray segments timeline

Advanced features

There are a lot of advanced features in the tracer utility, such as adding metadata and annotations, recording responses, and capturing exceptions.

For example, when we raise PetNotFoundError, it will be visible in the relevant segment details.

AWS X-Ray exception details

We could also add the correlation ID as an annotation, to allow us to filter and query traces by correlation ID. In the add_correlation_id middleware, add tracer.put_annotation(key="correlation_id", value=corr_id):

@app.middleware("http")
async def add_correlation_id(request: Request, call_next):
    # Get correlation id from X-Correlation-Id header
    corr_id = request.headers.get("x-correlation-id")
    if not corr_id:
        # If empty, use request id from aws context
        corr_id = request.scope["aws.context"].aws_request_id

    # Add correlation id to logs
    logger.set_correlation_id(corr_id)

    # Add correlation id to traces
    tracer.put_annotation(key="correlation_id", value=corr_id)

    response = await call_next(request)

    # Return correlation header in response
    response.headers["X-Correlation-Id"] = corr_id
    return response

Now, by drilling down into the ##handler segment, you should see the correlation ID in the correlation_id annotation.

AWS X-Ray segment with correlation ID annotation

If you want to find traces for a specific ID, you can use the query functionality in the console.

AWS X-Ray query by correlation ID

8. Conclusion

That’s all for now. Hopefully, you have learned a thing or two about how you can use the Lambda Powertools library to implement best practices when it comes to observability for FastAPI applications (and Lambda functions in general). AWS CloudWatch and AWS X-Ray are wonderful tools to help you analyze and monitor serverless applications. Lambda Powertools makes it incredibly easy to start using those services, allowing you to focus on your application logic while leaving the implementation of metrics, logs, and traces to the library.


Print Share Comment Cite Upload Translate
APA
Elias Brange | Sciencx (2024-03-29T11:39:23+00:00) » Observability Best Practices when running FastAPI in a Lambda. Retrieved from https://www.scien.cx/2022/04/11/observability-best-practices-when-running-fastapi-in-a-lambda/.
MLA
" » Observability Best Practices when running FastAPI in a Lambda." Elias Brange | Sciencx - Monday April 11, 2022, https://www.scien.cx/2022/04/11/observability-best-practices-when-running-fastapi-in-a-lambda/
HARVARD
Elias Brange | Sciencx Monday April 11, 2022 » Observability Best Practices when running FastAPI in a Lambda., viewed 2024-03-29T11:39:23+00:00,<https://www.scien.cx/2022/04/11/observability-best-practices-when-running-fastapi-in-a-lambda/>
VANCOUVER
Elias Brange | Sciencx - » Observability Best Practices when running FastAPI in a Lambda. [Internet]. [Accessed 2024-03-29T11:39:23+00:00]. Available from: https://www.scien.cx/2022/04/11/observability-best-practices-when-running-fastapi-in-a-lambda/
CHICAGO
" » Observability Best Practices when running FastAPI in a Lambda." Elias Brange | Sciencx - Accessed 2024-03-29T11:39:23+00:00. https://www.scien.cx/2022/04/11/observability-best-practices-when-running-fastapi-in-a-lambda/
IEEE
" » Observability Best Practices when running FastAPI in a Lambda." Elias Brange | Sciencx [Online]. Available: https://www.scien.cx/2022/04/11/observability-best-practices-when-running-fastapi-in-a-lambda/. [Accessed: 2024-03-29T11:39:23+00:00]
rf:citation
» Observability Best Practices when running FastAPI in a Lambda | Elias Brange | Sciencx | https://www.scien.cx/2022/04/11/observability-best-practices-when-running-fastapi-in-a-lambda/ | 2024-03-29T11:39:23+00:00
https://github.com/addpipe/simple-recorderjs-demo