/uvicorn-gunicorn-fastapi-docker

Docker image with Uvicorn managed by Gunicorn for high-performance FastAPI web applications in Python with performance auto-tuning. Optionally with Alpine Linux.

Primary LanguagePythonMIT LicenseMIT

Test Deploy

Supported tags and respective Dockerfile links

Discouraged tags

To learn more about why Alpine images are discouraged for Python read the note at the end: 🚨 Alpine Python Warning.

Deprecated tags

🚨 These tags are no longer supported or maintained, they are removed from the GitHub repository, but the last version pushed is still available in Docker Hub.

  • python3.6
  • python3.6-alpine3.8

The last versions with date tags for Python 3.6 are:

  • python3.6-2022-11-25
  • python3.6-alpine3.8-2022-11-25

Note: There are tags for each build date. If you need to "pin" the Docker image version you use, you can select one of those tags. E.g. tiangolo/uvicorn-gunicorn-fastapi:python3.7-2019-10-15.

uvicorn-gunicorn-fastapi

Docker image with Uvicorn managed by Gunicorn for high-performance FastAPI web applications in Python with performance auto-tuning. Optionally in a slim version or based on Alpine Linux.

GitHub repo: https://github.com/tiangolo/uvicorn-gunicorn-fastapi-docker

Docker Hub image: https://hub.docker.com/r/tiangolo/uvicorn-gunicorn-fastapi/

Description

FastAPI has shown to be a Python web framework with one of the best performances, as measured by third-party benchmarks, thanks to being based on and powered by Starlette.

The achievable performance is on par with (and in many cases superior to) Go and Node.js frameworks.

This image has an auto-tuning mechanism included to start a number of worker processes based on the available CPU cores. That way you can just add your code and get high performance automatically, which is useful in simple deployments.

🚨 WARNING: You Probably Don't Need this Docker Image

You are probably using Kubernetes or similar tools. In that case, you probably don't need this image (or any other similar base image). You are probably better off building a Docker image from scratch as explained in the docs for FastAPI in Containers - Docker: Build a Docker Image for FastAPI.


If you have a cluster of machines with Kubernetes, Docker Swarm Mode, Nomad, or other similar complex system to manage distributed containers on multiple machines, then you will probably want to handle replication at the cluster level instead of using a process manager (like Gunicorn with Uvicorn workers) in each container, which is what this Docker image does.

In those cases (e.g. using Kubernetes) you would probably want to build a Docker image from scratch, installing your dependencies, and running a single Uvicorn process instead of this image.

For example, your Dockerfile could look like:

FROM python:3.9

WORKDIR /code

COPY ./requirements.txt /code/requirements.txt

RUN pip install --no-cache-dir --upgrade -r /code/requirements.txt

COPY ./app /code/app

CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "80"]

You can read more about this in the FastAPI documentation about: FastAPI in Containers - Docker.

When to Use this Docker Image

A Simple App

You could want a process manager like Gunicorn running Uvicorn workers in the container if your application is simple enough that you don't need (at least not yet) to fine-tune the number of processes too much, and you can just use an automated default, and you are running it on a single server, not a cluster.

Docker Compose

You could be deploying to a single server (not a cluster) with Docker Compose, so you wouldn't have an easy way to manage replication of containers (with Docker Compose) while preserving the shared network and load balancing.

Then you could want to have a single container with a Gunicorn process manager starting several Uvicorn worker processes inside, as this Docker image does.

Prometheus and Other Reasons

You could also have other reasons that would make it easier to have a single container with multiple processes instead of having multiple containers with a single process in each of them.

For example (depending on your setup) you could have some tool like a Prometheus exporter in the same container that should have access to each of the requests that come.

In this case, if you had multiple containers, by default, when Prometheus came to read the metrics, it would get the ones for a single container each time (for the container that handled that particular request), instead of getting the accumulated metrics for all the replicated containers.

Then, in that case, it could be simpler to have one container with multiple processes, and a local tool (e.g. a Prometheus exporter) on the same container collecting Prometheus metrics for all the internal processes and exposing those metrics on that single container.


Read more about it all in the FastAPI documentation about: FastAPI in Containers - Docker.

Technical Details

Uvicorn

Uvicorn is a lightning-fast "ASGI" server.

It runs asynchronous Python web code in a single process.

Gunicorn

You can use Gunicorn to start and manage multiple Uvicorn worker processes.

That way, you get the best of concurrency and parallelism in simple deployments.

FastAPI

FastAPI is a modern, fast (high-performance), web framework for building APIs with Python.

The key features are:

  • Fast: Very high performance, on par with NodeJS and Go (thanks to Starlette and Pydantic).
  • Fast to code: Increase the speed to develop features by about 200% to 300% *.
  • Less bugs: Reduce about 40% of human (developer) induced errors. *
  • Intuitive: Great editor support. Completion everywhere. Less time debugging.
  • Easy: Designed to be easy to use and learn. Less time reading docs.
  • Short: Minimize code duplication. Multiple features from each parameter declaration. Less bugs.
  • Robust: Get production-ready code. With automatic interactive documentation.
  • Standards-based: Based on (and fully compatible with) the open standards for APIs: OpenAPI (previously known as Swagger) and JSON Schema.

* estimation based on tests on an internal development team, building production applications.

tiangolo/uvicorn-gunicorn-fastapi

This image will set a sensible configuration based on the server it is running on (the amount of CPU cores available) without making sacrifices.

It has sensible defaults, but you can configure it with environment variables or override the configuration files.

There is also a slim version and another one based on Alpine Linux. If you want one of those, use one of the tags from above.

tiangolo/uvicorn-gunicorn

This image (tiangolo/uvicorn-gunicorn-fastapi) is based on tiangolo/uvicorn-gunicorn.

That image is what actually does all the work.

This image just installs FastAPI and has the documentation specifically targeted at FastAPI.

If you feel confident about your knowledge of Uvicorn, Gunicorn and ASGI, you can use that image directly.

tiangolo/uvicorn-gunicorn-starlette

There is a sibling Docker image: tiangolo/uvicorn-gunicorn-starlette

If you are creating a new Starlette web application and you want to discard all the additional features from FastAPI you should use tiangolo/uvicorn-gunicorn-starlette instead.

Note: FastAPI is based on Starlette and adds several features on top of it. Useful for APIs and other cases: data validation, data conversion, documentation with OpenAPI, dependency injection, security/authentication and others.

How to use

You don't need to clone the GitHub repo.

You can use this image as a base image for other images.

Assuming you have a file requirements.txt, you could have a Dockerfile like this:

FROM tiangolo/uvicorn-gunicorn-fastapi:python3.11

COPY ./requirements.txt /app/requirements.txt

RUN pip install --no-cache-dir --upgrade -r /app/requirements.txt

COPY ./app /app

It will expect a file at /app/app/main.py.

Or otherwise a file at /app/main.py.

And will expect it to contain a variable app with your FastAPI application.

Then you can build your image from the directory that has your Dockerfile, e.g:

docker build -t myimage ./

Quick Start

Build your Image

  • Go to your project directory.
  • Create a Dockerfile with:
FROM tiangolo/uvicorn-gunicorn-fastapi:python3.11

COPY ./requirements.txt /app/requirements.txt

RUN pip install --no-cache-dir --upgrade -r /app/requirements.txt

COPY ./app /app
  • Create an app directory and enter in it.
  • Create a main.py file with:
from fastapi import FastAPI

app = FastAPI()


@app.get("/")
def read_root():
    return {"Hello": "World"}


@app.get("/items/{item_id}")
def read_item(item_id: int, q: str = None):
    return {"item_id": item_id, "q": q}
  • You should now have a directory structure like:
.
├── app
│   └── main.py
└── Dockerfile
  • Go to the project directory (in where your Dockerfile is, containing your app directory).
  • Build your FastAPI image:
docker build -t myimage .
  • Run a container based on your image:
docker run -d --name mycontainer -p 80:80 myimage

Now you have an optimized FastAPI server in a Docker container. Auto-tuned for your current server (and number of CPU cores).

Check it

You should be able to check it in your Docker container's URL, for example: http://192.168.99.100/items/5?q=somequery or http://127.0.0.1/items/5?q=somequery (or equivalent, using your Docker host).

You will see something like:

{"item_id": 5, "q": "somequery"}

Interactive API docs

Now you can go to http://192.168.99.100/docs or http://127.0.0.1/docs (or equivalent, using your Docker host).

You will see the automatic interactive API documentation (provided by Swagger UI):

Swagger UI

Alternative API docs

And you can also go to http://192.168.99.100/redoc or http://127.0.0.1/redoc(or equivalent, using your Docker host).

You will see the alternative automatic documentation (provided by ReDoc):

ReDoc

Dependencies and packages

You will probably also want to add any dependencies for your app and pin them to a specific version, probably including Uvicorn, Gunicorn, and FastAPI.

This way you can make sure your app always works as expected.

You could install packages with pip commands in your Dockerfile, using a requirements.txt, or even using Poetry.

And then you can upgrade those dependencies in a controlled way, running your tests, making sure that everything works, but without breaking your production application if some new version is not compatible.

Using Poetry

Here's a small example of one of the ways you could install your dependencies making sure you have a pinned version for each package.

Let's say you have a project managed with Poetry, so, you have your package dependencies in a file pyproject.toml. And possibly a file poetry.lock.

Then you could have a Dockerfile using Docker multi-stage building with:

FROM python:3.9 as requirements-stage

WORKDIR /tmp

RUN pip install poetry

COPY ./pyproject.toml ./poetry.lock* /tmp/

RUN poetry export -f requirements.txt --output requirements.txt --without-hashes

FROM tiangolo/uvicorn-gunicorn-fastapi:python3.11

COPY --from=requirements-stage /tmp/requirements.txt /app/requirements.txt

RUN pip install --no-cache-dir --upgrade -r /app/requirements.txt

COPY ./app /app

That will:

  • Install poetry and configure it for running inside of the Docker container.
  • Copy your application requirements.
    • Because it uses ./poetry.lock* (ending with a *), it won't crash if that file is not available yet.
  • Install the dependencies.
  • Then copy your app code.

It's important to copy the app code after installing the dependencies, that way you can take advantage of Docker's cache. That way it won't have to install everything from scratch every time you update your application files, only when you add new dependencies.

This also applies for any other way you use to install your dependencies. If you use a requirements.txt, copy it alone and install all the dependencies on the top of the Dockerfile, and add your app code after it.

Advanced usage

Environment variables

These are the environment variables that you can set in the container to configure it and their default values:

MODULE_NAME

The Python "module" (file) to be imported by Gunicorn, this module would contain the actual application in a variable.

By default:

  • app.main if there's a file /app/app/main.py or
  • main if there's a file /app/main.py

For example, if your main file was at /app/custom_app/custom_main.py, you could set it like:

docker run -d -p 80:80 -e MODULE_NAME="custom_app.custom_main" myimage

VARIABLE_NAME

The variable inside of the Python module that contains the FastAPI application.

By default:

  • app

For example, if your main Python file has something like:

from fastapi import FastAPI

api = FastAPI()


@api.get("/")
def read_root():
    return {"Hello": "World"}

In this case api would be the variable with the FastAPI application. You could set it like:

docker run -d -p 80:80 -e VARIABLE_NAME="api" myimage

APP_MODULE

The string with the Python module and the variable name passed to Gunicorn.

By default, set based on the variables MODULE_NAME and VARIABLE_NAME:

  • app.main:app or
  • main:app

You can set it like:

docker run -d -p 80:80 -e APP_MODULE="custom_app.custom_main:api" myimage

GUNICORN_CONF

The path to a Gunicorn Python configuration file.

By default:

  • /app/gunicorn_conf.py if it exists
  • /app/app/gunicorn_conf.py if it exists
  • /gunicorn_conf.py (the included default)

You can set it like:

docker run -d -p 80:80 -e GUNICORN_CONF="/app/custom_gunicorn_conf.py" myimage

You can use the config file from the base image as a starting point for yours.

WORKERS_PER_CORE

This image will check how many CPU cores are available in the current server running your container.

It will set the number of workers to the number of CPU cores multiplied by this value.

By default:

  • 1

You can set it like:

docker run -d -p 80:80 -e WORKERS_PER_CORE="3" myimage

If you used the value 3 in a server with 2 CPU cores, it would run 6 worker processes.

You can use floating point values too.

So, for example, if you have a big server (let's say, with 8 CPU cores) running several applications, and you have a FastAPI application that you know won't need high performance. And you don't want to waste server resources. You could make it use 0.5 workers per CPU core. For example:

docker run -d -p 80:80 -e WORKERS_PER_CORE="0.5" myimage

In a server with 8 CPU cores, this would make it start only 4 worker processes.

Note: By default, if WORKERS_PER_CORE is 1 and the server has only 1 CPU core, instead of starting 1 single worker, it will start 2. This is to avoid bad performance and blocking applications (server application) on small machines (server machine/cloud/etc). This can be overridden using WEB_CONCURRENCY.

MAX_WORKERS

Set the maximum number of workers to use.

You can use it to let the image compute the number of workers automatically but making sure it's limited to a maximum.

This can be useful, for example, if each worker uses a database connection and your database has a maximum limit of open connections.

By default it's not set, meaning that it's unlimited.

You can set it like:

docker run -d -p 80:80 -e MAX_WORKERS="24" myimage

This would make the image start at most 24 workers, independent of how many CPU cores are available in the server.

WEB_CONCURRENCY

Override the automatic definition of number of workers.

By default:

  • Set to the number of CPU cores in the current server multiplied by the environment variable WORKERS_PER_CORE. So, in a server with 2 cores, by default it will be set to 2.

You can set it like:

docker run -d -p 80:80 -e WEB_CONCURRENCY="2" myimage

This would make the image start 2 worker processes, independent of how many CPU cores are available in the server.

HOST

The "host" used by Gunicorn, the IP where Gunicorn will listen for requests.

It is the host inside of the container.

So, for example, if you set this variable to 127.0.0.1, it will only be available inside the container, not in the host running it.

It's is provided for completeness, but you probably shouldn't change it.

By default:

  • 0.0.0.0

PORT

The port the container should listen on.

If you are running your container in a restrictive environment that forces you to use some specific port (like 8080) you can set it with this variable.

By default:

  • 80

You can set it like:

docker run -d -p 80:8080 -e PORT="8080" myimage

BIND

The actual host and port passed to Gunicorn.

By default, set based on the variables HOST and PORT.

So, if you didn't change anything, it will be set by default to:

  • 0.0.0.0:80

You can set it like:

docker run -d -p 80:8080 -e BIND="0.0.0.0:8080" myimage

LOG_LEVEL

The log level for Gunicorn.

One of:

  • debug
  • info
  • warning
  • error
  • critical

By default, set to info.

If you need to squeeze more performance sacrificing logging, set it to warning, for example:

You can set it like:

docker run -d -p 80:8080 -e LOG_LEVEL="warning" myimage

WORKER_CLASS

The class to be used by Gunicorn for the workers.

By default, set to uvicorn.workers.UvicornWorker.

The fact that it uses Uvicorn is what allows using ASGI frameworks like FastAPI, and that is also what provides the maximum performance.

You probably shouldn't change it.

But if for some reason you need to use the alternative Uvicorn worker: uvicorn.workers.UvicornH11Worker you can set it with this environment variable.

You can set it like:

docker run -d -p 80:8080 -e WORKER_CLASS="uvicorn.workers.UvicornH11Worker" myimage

TIMEOUT

Workers silent for more than this many seconds are killed and restarted.

Read more about it in the Gunicorn docs: timeout.

By default, set to 120.

Notice that Uvicorn and ASGI frameworks like FastAPI are async, not sync. So it's probably safe to have higher timeouts than for sync workers.

You can set it like:

docker run -d -p 80:8080 -e TIMEOUT="20" myimage

KEEP_ALIVE

The number of seconds to wait for requests on a Keep-Alive connection.

Read more about it in the Gunicorn docs: keepalive.

By default, set to 2.

You can set it like:

docker run -d -p 80:8080 -e KEEP_ALIVE="20" myimage

GRACEFUL_TIMEOUT

Timeout for graceful workers restart.

Read more about it in the Gunicorn docs: graceful-timeout.

By default, set to 120.

You can set it like:

docker run -d -p 80:8080 -e GRACEFUL_TIMEOUT="20" myimage

ACCESS_LOG

The access log file to write to.

By default "-", which means stdout (print in the Docker logs).

If you want to disable ACCESS_LOG, set it to an empty value.

For example, you could disable it with:

docker run -d -p 80:8080 -e ACCESS_LOG= myimage

ERROR_LOG

The error log file to write to.

By default "-", which means stderr (print in the Docker logs).

If you want to disable ERROR_LOG, set it to an empty value.

For example, you could disable it with:

docker run -d -p 80:8080 -e ERROR_LOG= myimage

GUNICORN_CMD_ARGS

Any additional command line settings for Gunicorn can be passed in the GUNICORN_CMD_ARGS environment variable.

Read more about it in the Gunicorn docs: Settings.

These settings will have precedence over the other environment variables and any Gunicorn config file.

For example, if you have a custom TLS/SSL certificate that you want to use, you could copy them to the Docker image or mount them in the container, and set --keyfile and --certfile to the location of the files, for example:

docker run -d -p 80:8080 -e GUNICORN_CMD_ARGS="--keyfile=/secrets/key.pem --certfile=/secrets/cert.pem" -e PORT=443 myimage

Note: instead of handling TLS/SSL yourself and configuring it in the container, it's recommended to use a "TLS Termination Proxy" like Traefik. You can read more about it in the FastAPI documentation about HTTPS.

PRE_START_PATH

The path where to find the pre-start script.

By default, set to /app/prestart.sh.

You can set it like:

docker run -d -p 80:8080 -e PRE_START_PATH="/custom/script.sh" myimage

Custom Gunicorn configuration file

The image includes a default Gunicorn Python config file at /gunicorn_conf.py.

It uses the environment variables declared above to set all the configurations.

You can override it by including a file in:

  • /app/gunicorn_conf.py
  • /app/app/gunicorn_conf.py
  • /gunicorn_conf.py

Custom /app/prestart.sh

If you need to run anything before starting the app, you can add a file prestart.sh to the directory /app. The image will automatically detect and run it before starting everything.

For example, if you want to add Alembic SQL migrations (with SQLALchemy), you could create a ./app/prestart.sh file in your code directory (that will be copied by your Dockerfile) with:

#! /usr/bin/env bash

# Let the DB start
sleep 10;
# Run migrations
alembic upgrade head

and it would wait 10 seconds to give the database some time to start and then run that alembic command.

If you need to run a Python script before starting the app, you could make the /app/prestart.sh file run your Python script, with something like:

#! /usr/bin/env bash

# Run custom Python script before starting
python /app/my_custom_prestart_script.py

You can customize the location of the prestart script with the environment variable PRE_START_PATH described above.

Development live reload

The default program that is run is at /start.sh. It does everything described above.

There's also a version for development with live auto-reload at:

/start-reload.sh

Details

For development, it's useful to be able to mount the contents of the application code inside of the container as a Docker "host volume", to be able to change the code and test it live, without having to build the image every time.

In that case, it's also useful to run the server with live auto-reload, so that it re-starts automatically at every code change.

The additional script /start-reload.sh runs Uvicorn alone (without Gunicorn) and in a single process.

It is ideal for development.

Usage

For example, instead of running:

docker run -d -p 80:80 myimage

You could run:

docker run -d -p 80:80 -v $(pwd):/app myimage /start-reload.sh
  • -v $(pwd):/app: means that the directory $(pwd) should be mounted as a volume inside of the container at /app.
    • $(pwd): runs pwd ("print working directory") and puts it as part of the string.
  • /start-reload.sh: adding something (like /start-reload.sh) at the end of the command, replaces the default "command" with this one. In this case, it replaces the default (/start.sh) with the development alternative /start-reload.sh.

Development live reload - Technical Details

As /start-reload.sh doesn't run with Gunicorn, any of the configurations you put in a gunicorn_conf.py file won't apply.

But these environment variables will work the same as described above:

  • MODULE_NAME
  • VARIABLE_NAME
  • APP_MODULE
  • HOST
  • PORT
  • LOG_LEVEL

🚨 Alpine Python Warning

In short: You probably shouldn't use Alpine for Python projects, instead use the slim Docker image versions.


Do you want more details? Continue reading 👇

Alpine is more useful for other languages where you build a static binary in one Docker image stage (using multi-stage Docker building) and then copy it to a simple Alpine image, and then just execute that binary. For example, using Go.

But for Python, as Alpine doesn't use the standard tooling used for building Python extensions, when installing packages, in many cases Python (pip) won't find a precompiled installable package (a "wheel") for Alpine. And after debugging lots of strange errors you will realize that you have to install a lot of extra tooling and build a lot of dependencies just to use some of these common Python packages. 😩

This means that, although the original Alpine image might have been small, you end up with a an image with a size comparable to the size you would have gotten if you had just used a standard Python image (based on Debian), or in some cases even larger. 🤯

And in all those cases, it will take much longer to build, consuming much more resources, building dependencies for longer, and also increasing its carbon footprint, as you are using more CPU time and energy for each build. 🌳

If you want slim Python images, you should instead try and use the slim versions that are still based on Debian, but are smaller. 🤓

Tests

All the image tags, configurations, environment variables and application options are tested.

Release Notes

Latest Changes

0.7.0

Highlights of this release:

  • Support for Python 3.10 and 3.11.
  • Deprecation of Python 3.6.
    • The last Python 3.6 image tag was pushed and is available in Docker Hub, but it won't be updated or maintained anymore.
    • The last image with a date tag is python3.6-2022-11-25.
  • Upgraded versions of all the dependencies.

Features

  • ✨ Add support for Python 3.10 and 3.11. PR #220 by @tiangolo.
  • ✨ Add Python 3.9 and Python 3.9 Alpine. PR #67 by @graue70.

Breaking Changes

Upgrades

  • ⬆️ Upgrade FastAPI and Uvicorn versions. PR #212 by @tiangolo.
  • ⬆️ Upgrade packages to the last version that supports Python 3.6. PR #207 by @tiangolo.

Docs

  • 📝 Add note to discourage Alpine with Python. PR #122 by @tiangolo.
  • 📝 Add warning for Kubernetes, when to use this image. PR #121 by @tiangolo.
  • ✏ Fix typo, repeated word on README. PR #96 by @shelbylsmith.

Internal

0.6.0

  • Add docs about installing and pinning dependencies. PR #41.
  • Add slim version. PR #40.
  • Update and refactor bringing all the new features from the base image. Includes:
    • Centralize, simplify, and deduplicate code and setup
    • Move CI to GitHub actions
    • Add Python 3.8 (and Alpine)
    • Add new configs and docs:
      • WORKER_CLASS
      • TIMEOUT
      • KEEP_ALIVE
      • GRACEFUL_TIMEOUT
      • ACCESS_LOG
      • ERROR_LOG
      • GUNICORN_CMD_ARGS
      • MAX_WORKERS
    • PR #39.
  • Disable pip cache during installation. PR #38.
  • Migrate local development from Pipenv to Poetry. PR #34.
  • Add docs for custom PRE_START_PATH env var. PR #33.

0.5.0

  • Refactor tests to use env vars and add image tags for each build date, like tiangolo/uvicorn-gunicorn-fastapi:python3.7-2019-10-15. PR #17.
  • Upgrade Travis. PR #9.

0.4.0

  • Add support for live auto-reload with an additional custom script /start-reload.sh, check the updated documentation. PR #6 in parent image.

0.3.0

  • Set WORKERS_PER_CORE by default to 1, as it shows to have the best performance on benchmarks.
  • Make the default web concurrency, when WEB_CONCURRENCY is not set, to a minimum of 2 workers. This is to avoid bad performance and blocking applications (server application) on small machines (server machine/cloud/etc). This can be overridden using WEB_CONCURRENCY. This applies for example in the case where WORKERS_PER_CORE is set to 1 (the default) and the server has only 1 CPU core. PR #6 and PR #5 in parent image.

0.2.0

  • Make /start.sh run independently, reading and generating used default environment variables. And remove /entrypoint.sh as it doesn't modify anything in the system, only reads environment variables. PR #4 in parent image.

0.1.0

  • Add support for /app/prestart.sh.

License

This project is licensed under the terms of the MIT license.