This is a PyPI mirror client according to PEP 381
http://www.python.org/dev/peps/pep-0381/.
The following instructions will place the bandersnatch executable in a
virtualenv under bandersnatch/bin/bandersnatch
.
- bandersnatch requires
>= Python 3.6
This installs the latest stable, released version.
$ python3.6 -m venv bandersnatch
$ bandersnatch/bin/pip install -r https://bitbucket.org/pypa/bandersnatch/raw/stable/requirements.txt
This installs the current development version. Use 'hg up ' and run buildout again to choose a specific release.
$ git clone https://github.com/pypa/bandersnatch
$ cd bandersnatch
$ ./bootstrap.sh
- Run
bandersnatch mirror
- it will create an empty configuration file for you in/etc/bandersnatch.conf
. - Review
/etc/bandersnatch.conf
and adapt to your needs. - Run
bandersnatch mirror
again. It will populate your mirror with the current status of all PyPI packages - roughly 500GiB (2017-02-12). Expect this to grow substantially over time. - Run
bandersnatch mirror
regularly to update your mirror with any intermediate changes.
Configure your webserver to serve the web/
sub-directory of the mirror.
For nginx it should look something like this::
server {
listen 127.0.0.1:80;
server_name <mymirrorname>;
root <path-to-mirror>/web;
autoindex on;
charset utf-8;
}
-
Note that it is a good idea to have your webserver publish the HTML index files correctly with UTF-8 as the charset. The index pages will work without it but if humans look at the pages the characters will end up looking funny.
-
Make sure that the webserver uses UTF-8 to look up unicode path names. nginx gets this right by default - not sure about others.
You need to set up one cron job to run the mirror itself.
Here's a sample that you could place in /etc/cron.d/bandersnatch
:
LC_ALL=en_US.utf8
*/2 * * * * root bandersnatch mirror |& logger -t bandersnatch[mirror]
This assumes that you have a logger
utility installed that will convert the
output of the commands to syslog entries.
bandersnatch does not keep much local state in addition to the mirrored data.
In general you can just keep rerunning bandersnatch mirror
to make it fix
errors.
If you delete the state files then the next run will force it to check everything against the master PyPI::
- delete
./state
file and./todo
if they exist in your mirror directory - run
bandersnatch
mirror to get a full sync
Be aware, that full syncs likely take hours depending on PyPIs performance and your network latency and bandwidth.
You need to run bandersnatch on a case-sensitive filesystem.
OS X natively does this OK even though the filesystem is not strictly case-sensitive and bandersnatch will work fine when running on OS X. However, tarring a bandersnatch data directory and moving it to, e.g. Linux with a case-sensitive filesystem will lead to inconsistencies. You can fix those by deleting the status files and have bandersnatch run a full check on your data.
The PyPI has a quite extensive list of packages that we need to maintain in a flat directory. Filesystems with small limits on the number of sub-directories per directory can run into a problem like this::
2013-07-09 16:11:33,331 ERROR: Error syncing package: zweb@802449 OSError: [Errno 31] Too many links: '../pypi/web/simple/zweb'
Specifically we recommend to avoid using ext3. Ext4 and newer does not have the limitation of 32k sub-directories.
A bandersnatch static mirror is compatible only to the "static", cacheable parts of PyPI that are needed to support package installation. It does not support more dynamic APIs of PyPI that maybe be used by various clients for other purposes.
An example of an unsupported API is PyPI's XML-RPC interface, which is used
when running pip search
.
If you have questions or comments, please submit a bug report to https://github.com/pypa/bandersnatch/issues/new
Everyone interacting in the bandersnatch project's codebases, issue trackers, chat rooms, and mailing lists is expected to follow the PyPA Code of Conduct.
This client is based on the original pep381client by Martin v. Loewis.
Richard Jones was very patient answering questions at PyCon 2013 and made the protocol more reliable by implementing some PyPI enhancements.