pypi i dumb-pypi



Build Status PyPI version

dumb-pypi is a simple read-only PyPI index server generator, backed entirely by static files. It is ideal for internal use by organizations that have a bunch of their own packages which they'd like to make available.

You can view an example generated repo.

A rant about static files (and why you should use dumb-pypi)

The main difference between dumb-pypi and other PyPI implementations is that dumb-pypi has no server component. It's just a script that, given a list of Python package names, generates a bunch of static files which you can serve from any webserver, or even directly from S3.

There's something magical about being able to serve a package repository entirely from a tree of static files. It's incredibly easy to make it fast and highly-available when you don't need to worry about running a bunch of application servers (which are serving a bunch of read-only queries that could have just been pre-generated).

Linux distributions have been doing this right for decades. Debian has a system of hundreds of mirrors, and the entire thing is powered entirely by some fancy rsync commands.

For the maintainer of a PyPI repositry, dumb-pypi has some nice properties:

  • File serving is extremely fast. nginx can serve your static files faster than you'd ever need. In practice, there are almost no limits on the number of packages or number of versions per package.

  • It's very simple. There's no complicated WSGI app to deploy, no databases, and no caches. You just need to run the script whenever you have new packages, and your index server is ready in seconds.

For more about why this design was chosen, see the detailed RATIONALE.md in this repo.


To use dumb-pypi, you need two things:

  • A script which generates the index. (That's this project!)

  • A generic webserver to serve the generated index.

    This part is up to you. For example, you might sync the built index into an S3 bucket, and serve it directly from S3. You might run nginx from the built index locally.

My recommended high-availability (but still quite simple) deployment is:

  • Store all of the packages in S3.

  • Have a cronjob (or equivalent) which rebuilds the index based on the packages in S3. This is incredibly fast—it would not be unreasonable to do it every sixty seconds. After building the index, sync it into a separate S3 bucket.

  • Have a webserver (or set of webservers behind a load balancer) running nginx (with the config provided below), with the source being that second S3 bucket.

Generating static files

First, install dumb-pypi somewhere (e.g. into a virtualenv).

By design, dumb-pypi does not require you to have the packages available when building the index. You only need a list of filenames, one per line. For example:


You should also know a URL to access these packages (if you serve them from the same host as the index, it can be a relative URL). For example, it might be https://my-pypi-packages.s3.amazonaws.com/ or ../../pool/.

You can then invoke the script:

$ dumb-pypi \
    --package-list my-packages \
    --packages-url https://my-pypi-packages.s3.amazonaws.com/ \
    --output-dir my-built-index

The built index will be in my-built-index. It's now up to you to figure out how to serve that with a webserver (nginx is a good option — details below!).

Additional options for packages

You can extend the capabilities of your registry using the extended JSON input syntax when providing your package list to dumb-pypi. Instead of using the format listed above of one filename per line, format your file with one JSON object per line, like this:

{"filename": "dumb-init-1.1.2.tar.gz", "hash": "md5=<hash>", "requires_dist": ["cfgv"], "requires_python": ">=3.6", "uploaded_by": "ckuehl", "upload_timestamp": 1512539924}

The filename key is required. All other keys are optional and will be used to provide additional information in your generated repository. This extended information can be useful to determine, for example, who uploaded a package. (Most of this information is useful in the web UI by humans, not by pip.)

Where should you get information about the hash, uploader, etc? That's up to you—dumb-pypi isn't in the business of storing or calculating this data. If you're using S3, one easy option is to store it at upload time as S3 metadata.

Partial rebuild support

If you want to avoid rebuilding your entire registry constantly, you can pass the --previous-package-list (or --previous-package-list-json) argument to dumb-pypi, pointing to the list you used the last time you called dumb-pypi. Only the files relating to changed packages will be rebuilt, saving you time and unnecessary I/O.

The previous package list json is available in the output as packages.json.

You can serve the packages from any static webserver (including directly from S3), but for compatibility with old versions of pip, it's necessary to do a tiny bit of URL rewriting (see RATIONALE.md for full details about the behavior of various pip versions).

In particular, if you want to support old pip versions, you need to apply this logic to package names (taken from PEP 503):

def normalize(name):
    return re.sub(r'[-_.]+', '-', name).lower()

Here is an example nginx config which supports all versions of pip and easy_install:

server {
    location / {
        root /path/to/index;
        set_by_lua $canonical_uri "return string.gsub(string.lower(ngx.var.uri), '[-_.]+', '-')";
        try_files $uri $uri/index.html $canonical_uri $canonical_uri/index.html =404;

If you don't care about easy_install or versions of pip prior to 8.1.2, you can omit the canonical_uri hack.

Using your deployed index server with pip

When running pip, pass -i https://my-pypi-server/simple or set the environment variable PIP_INDEX_URL=https://my-pypi-server/simple.

Known incompatibilities with public PyPI

We try to maintain compatibility with the standard PyPI interface, but there are some incompatibilities currently which are hard to fix due to dumb-pypi's design:

  • While both JSON API endpoints are supported, many keys in the JSON API are not present since they require inspecting packages which dumb-pypi can't do. Some of these, like requires_python and requires_dist, can be passed in as JSON.

  • The per-version JSON API endpoint only includes data about the current requested version and not all versions, unlike public PyPI. In other words, if you access /pypi/<package>/1.0.0/json, you will only see the 1.0.0 release under the releases key and not every release ever made. The regular non-versioned API route (/pypi/<package>/json) will have all releases.


Thanks for contributing! To get started, run make venv and then . venv/bin/activate to source the virtualenv. You should now have a dumb-pypi command on your path using your checked-out version of the code.

To run the tests, call make test. To run an individual test, you can do pytest -k name_of_test tests (with the virtualenv activated).

Jump To