proxpi

PyPI caching mirror

Host a proxy PyPI mirror server with caching
- Cache the index (project list and projects' file list)
- Cache the project files
Support multiple indices
Set index cache times-to-live (individually for each index)
Set files cache max-size on disk
Manually invalidate index cache

Usage

Start server

Choose between running inside Docker container if you want to run in a known-working environment, or outside via a Python app (instructions here are for the Flask development server) if you want more control over the environment.

Docker

Uses a Gunicorn WSGI server

docker run -p 5000:5000 epicwink/proxpi

Without arguments, runs with 2 threads. If passing arguments, make sure to bind to an exported address (or all with 0.0.0.0) on port 5000 (ie --bind 0.0.0.0:5000).

Compose

Alternatively, use Docker Compose

docker compose up

Local

Install

pip install proxpi

Install coloredlogs as well to get coloured logging

Run server

FLASK_APP=proxpi.server flask run

See flask run --help for more information on address and port binding, and certificate specification to use HTTPS. Alternatively, bring your own WSGI server.

Use proxy

Use PIP's index-URL flag to install packages via the proxy

pip install --index-url http://127.0.0.1:5000/index/ simplejson

Cache invalidation

Either head to http://127.0.0.1:5000/ in the browser, or run:

curl -X DELETE http://127.0.0.1:5000/cache/simplejson
curl -X DELETE http://127.0.0.1:5000/cache/list

If you need to invalidate a locally cached file, restart the server: files should never change in a package index.

Environment variables

PROXPI_INDEX_URL: index URL, default: https://pypi.org/simple/
PROXPI_INDEX_TTL: index cache time-to-live in seconds, default: 30 minutes. Disable index-cache by setting this to 0
PROXPI_EXTRA_INDEX_URLS: extra index URLs (comma-separated)
PROXPI_EXTRA_INDEX_TTLS: corresponding extra index cache times-to-live in seconds (comma-separated), default: 3 minutes, cache disabled when 0
PROXPI_CACHE_SIZE: size of downloaded project files cache (bytes), default 5GB. Disable files-cache by setting this to 0
PROXPI_CACHE_DIR: downloaded project files cache directory path, default: a new temporary directory
PROXPI_BINARY_FILE_MIME_TYPE=1: force file-response content-type to "application/octet-stream" instead of letting Flask guess it. This may be needed if your package installer (eg Poetry) mishandles responses with declared encoding.
PROXPI_DISABLE_INDEX_SSL_VERIFICATION=1: don't verify any index SSL certificates
PROXPI_DOWNLOAD_TIMEOUT: time (in seconds) before proxpi will redirect to the proxied index server for file downloads instead of waiting for the download, default: 0.9
PROXPI_CONNECT_TIMEOUT: time (in seconds) proxpi will wait for a socket to connect to the index server before requests raises a ConnectTimeout error to prevent indefinite blocking, default: none, or 3.1 if read-timeout provided
PROXPI_READ_TIMEOUT: time (in seconds) proxpi will wait for chunks of data from the index server before requests raises a ReadTimeout error to prevent indefinite blocking, default: none, or 20 if connect-timeout provided

Considerations with CI

proxpi was designed with three goals (particularly for continuous integration (CI)):

to reduce load on PyPI package serving
to reduce pip install times
not require modification to the current workflow

Specifically, proxpi was designed to run for CI services such as Travis, Jenkins, GitLab CI, Azure Pipelines and GitHub Actions.

proxpi works by caching index requests (ie which versions, wheel-types, etc are available for a given project, the index cache) and the project files themselves (to a local directory, the package cache). This means they will cache identical requests after the first request, and will be useless for just one pip install.

Cache persistence

As a basic end-user of these services, for at least most of these services you won't be able to keep a proxpi server running between multiple invocations of your project(s) CI pipeline: CI invocations are designed to be independent. This means the best that you can do is start the cache for just the current job.

A more advanced user of these CI services can bring their own runner (personally, my needs are for running GitLab CI). This means you can run proxpi on a fully-controlled server (eg EC2 instance), and proxy PyPI requests (during a pip command) through the local cache. See the instructions below.

Hopefully, in the future these CI services will all implement their own transparent caching for PyPI. For example, Azure already has Azure Artifacts which provides much more functionality than proxpi, but won't reduce pip install times for CI services not using Azure.

GitLab CI instructions

This implementation leverages the index URL configurable of pip and Docker networks. This is to be run on a server you have console access to.

Create a Docker bridge network

docker network create gitlab-runner-network

Start a GitLab CI Docker runner using their documentation

Run the proxpi Docker container

docker run \
  --detach \
  --network gitlab-runner-network \
  --volume proxpi-cache:/var/cache/proxpi \
  --env PROXPI_CACHE_DIR=/var/cache/proxpi \
  --name proxpi epicwink/proxpi:latest

You don't need to expose a port (the -p flag) as we'll be using an internal Docker network.

Set pip's index URL to the proxpi server by setting it in the runner environment. Set runners[0].docker.network_mode to gitlab-runner-network. Add PIP_INDEX_URL=http://proxpi:5000/index/ and PIP_TRUSTED_HOST=proxpi to runners.environment in the GitLab CI runner configuration TOML. For example, you may end up with the following configuration:

[[runners]]
  name = "awesome-ci-01"
  url = "https://gitlab.com/"
  token = "SECRET"
  executor = "docker"
  environment = [
    "DOCKER_TLS_CERTDIR=/certs",
    "PIP_INDEX_URL=http://proxpi:5000/index/",
    "PIP_TRUSTED_HOST=proxpi",
  ]

[[runners.docker]]
  network_mode = "gitlab-runner-network"
  ...

This is designed to not require any changes to the GitLab CI project configuration (ie gitlab-ci.yml), unless it already sets the index URL for some reason (if that's the case, you're probably already using a cache).

Another option is to set up a proxy, but that's more effort than the above method.

Alternatives

simpleindex: routes URLs to multiple indices (including PyPI), supports local (or S3 with a plygin) directory of packages, no caching without custom plugins
bandersnatch: mirrors one index (eg PyPI), storing packages locally, or on S3 with a plugin. Manual update, no proxy
devpi: heavyweight, runs a full index (or multiple) in addition to mirroring (in place of proxying), supports proxying (with inheritance), supports package upload, server replication and fail-over
pypiserver: serves local directory of packages, proxy to PyPI when not-found, supports package upload, no caching
PyPI Cloud: serves local or cloud-storage directory of packages, with redirecting/cached proxying to indexes, authentication and authorisation.
pypiprivate: serves local (or S3-hosted) directory of packages, no proxy to package indices (including PyPI)
Pulp: generic content repository, can host multiple ecosystems' packages. Python package index plugin supports local/S3 mirrors, package upload, proxying to multiple indices, no caching
pip2pi: manual syncing of specific packages, no proxy
nginx_pypi_cache: caching proxy using nginx, single index
Flask-Pypi-Proxy: unmaintained, no cache size limit, no caching index pages
http.server: standard-library, hosts directory exactly as laid out, no proxy to package indices (eg PyPI)
Apache with mod_rewrite: I'm not familiar with Apache, but it likely has the capability to proxy and cache (with eg mod_cache_disk)
Gemfury: hosted, managed. Private index is not free, documentation doesn't say anything about proxying

ToddG/proxpi