PEP: 381 Title: Mirroring infrastructure for PyPI Version: $Revision$ Last-Modified: $Date$ Author: Tarek Ziadé , Martin v. Löwis Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 21-March-2009 Python-Version: N.A. Post-History: Abstract ======== This PEP describes a mirroring infrastructure for PyPI. Rationale ========= PyPI is hosting over 6000 projects and is used on a daily basis by people to build applications. Especially systems like `easy_install` and `zc.buildout` make intensive usage of PyPI. For people making intensive use of PyPI, it can act as a single point of failure. People have started to set up some mirrors, both private and public. Those mirrors are active mirrors, which means that they are browsing PyPI to get synced. In order to make the system more reliable, this PEP describes: - the mirror listing and registering at PyPI - the pages a public mirror should maintain. These pages will be used by PyPI, in order to get hit counts and the last modified date. - how a mirror should synchronize with PyPI - how a client can implement a fail-over mechanism Mirror listing and registering ============================== People that wants to mirror PyPI make a proposal on catalog-SIG. When a mirror is proposed on the mailing list, it is manually added in a mirror list in the PyPI application after it has been checked to be compliant with the mirroring rules. The mirror list is provided as a list of host names of the form X.pypi.python.org The values of X are the sequence a,b,c,...,aa,ab,... a.pypi.python.org is the master server; the mirrors start with b. A CNAME record last.pypi.python.org points to the last host name. Mirror operators should use a static address, and report planned changes to that address in advance to distutils-sig. The new mirror also appears at `http://pypi.python.org/mirrors` which is a human-readable page that gives the list of mirrors. This page also explains how to register a new mirror. Statistics page ::::::::::::::: PyPI provides statistics on downloads at `/stats`. This page is calculated daily by PyPI, by reading all mirrors' local stats and summing them. The stats are presented in daily or montly files, under `/stats/days` and `/stats/months`. Each file is a `bzip2` file with these formats: - YYYY-MM-DD.bz2 for daily files - YYYY-MM.bz2 for monthly files Examples: - /stats/days/2008-11-06.bz2 - /stats/days/2008-11-07.bz2 - /stats/days/2008-11-08.bz2 - /stats/months/2008-11.bz2 - /stats/months/2008-10.bz2 Mirror Authenticity =================== With a distributed mirroring system, clients may want to verify that the mirrored copies are authentic. There are multiple threats to consider: 1. the central index may get compromised 2. the central index is assumed to be trusted, but the mirrors might be tampered. 3. a man in the middle between the central index and the end user, or between a mirror and the end user might tamper with datagrams. This specification only deals with the second threat. Some provisions are made to detect man-in-the-middle attacks. To detect the first attack, package authors need to sign their packages using PGP keys, so that users verify that the package comes from the author they trust. The central index provides a DSA key at the URL /serverkey, in the PEM format as generated by "openssl dsa -pubout" (i.e. RFC 3280 SubjectPublicKeyInfo, with the algorithm 1.3.14.3.2.12). This URL must *not* be mirrored, and clients must fetch the official serverkey from PyPI directly, or use the copy that came with the PyPI client software. Mirrors should still download the key, to detect a key rollover. For each package, a mirrored signature is provided at /serversig/. This is the DSA signature of the parallel URL /simple/, in DER form, using SHA-1 with DSA (i.e. as a RFC 3279 Dsa-Sig-Value, created by algorithm 1.2.840.10040.4.3) Clients using a mirror need to perform the following steps to verify a package: 1. download the /simple page, and compute its SHA-1 hash 2. compute the DSA signature of that hash 3. download the corresponding /serversig, and compare it (byte-for-byte) with the value computed in step 2. 4. compute and verify (against the /simple page) the MD-5 hashes of all files they download from the mirror. An implementation of the verification algorithm is available from https://svn.python.org/packages/trunk/pypi/tools/verify.py Verification is not needed when downloading from central index, and should be avoided to reduce the computation overhead. About once a year, the key will be replaced with a new one. Mirrors will have to re-fetch all /serversig pages. Clients using mirrors need to find a trusted copy of the new server key. One way to obtain one is to download it from https://pypi.python.org/serverkey. To detect man-in-the-middle attacks, clients need to verify the SSL server certificate, which will be signed by the CACert authority. Special pages a mirror needs to provide ======================================= A mirror is a subset copy of PyPI, so it provides the same structure by copying it. - simple: rest version of the package index - packages: packages, stored by Python version, and letters - serversig: signatures for the simple pages It also needs to provide two specific elements: - last-modified - local-stats Last modified date :::::::::::::::::: CPAN uses a freshness date system where the mirror's last synchronisation date is made available. For PyPI, each mirror needs to maintain a URL with simple text content that represents the last synchronisation date the mirror maintains. The date is provided in GMT time, using the ISO 8601 format [#iso8601]_. Each mirror will be responsible to maintain its last modified date. This page must be located at : `/last-modified` and must be a text/plain page. Local statistics :::::::::::::::: Each mirror is responsible to count all the downloads that where done via it. This is used by PyPI to sum up all downloads, to be able to display the grand total. These statistics are in CSV-like form, with a header in the first line. It needs to obey PEP 305 [#pep305]_. Basically, it should be readable by Python's `csv` module. The fields in this file are: - package: the distutils id of the package. - filename: the filename that has been downloaded. - useragent: the User-Agent of the client that has downloaded the package. - count: the number of downloads. The content will look like this:: # package,filename,useragent,count zc.buildout,zc.buildout-1.6.0.tgz,MyAgent,142 ... The counting starts the day the mirror is launched, and there is one file per day, compressed using the `bzip2` format. Each file is named like the day. For example, `2008-11-06.bz2` is the file for the 6th of November 2008. They are then provided in a folder called `days`. For example: - /local-stats/days/2008-11-06.bz2 - /local-stats/days/2008-11-07.bz2 - /local-stats/days/2008-11-08.bz2 This page must be located at `/local-stats`. How a mirror should synchronize with PyPI ========================================= A mirroring protocol called `Simple Index` was described and implemented by Martin v. Loewis and Jim Fulton, based on how `easy_install` works. This section synthesizes it and gives a few relevant links, plus a small part about `User-Agent`. The mirroring protocol :::::::::::::::::::::: Mirrors must reduce the amount of data transferred between the central server and the mirror. To achieve that, they MUST use the changelog() PyPI XML-RPC call, and only refetch the packages that have been changed since the last time. For each package P, they MUST copy documents /simple/P/ and /serversig/P. If a package is deleted on the central server, they MUST delete the package and all associated files. To detect modification of package files, they MAY cache the file's ETag, and MAY request skipping it using the If-none-match header. Each mirroring tool MUST identify itself using a descripte User-agent header. The pep381client package [#pep381client]_ provides an application that respects this protocol to browse PyPI. User-agent request header ::::::::::::::::::::::::: In order to be able to differentiate actions taken by clients over PyPI, a specific user agent name should be provided by all mirroring software. This is also true for all clients like: - zc.buildout [#zc.buildout]_. - setuptools [#setuptools]_. - pip [#pip]_. XXX user agent registering mechanism at PyPI ? How a client can use PyPI and its mirrors ::::::::::::::::::::::::::::::::::::::::: Clients that are browsing PyPI should be able to use alternative mirrors, by getting the list of the mirrors using `last.pypi.python.org`. Code example:: >>> import socket >>> socket.gethostbyname_ex('last.pypi.python.org')[0] 'h.pypi.python.org' The clients so far that could use this mechanism: - setuptools - zc.buildout (through setuptools) - pip Fail-over mechanism ::::::::::::::::::: Clients that are browsing PyPI should be able to use a fail-over mechanism when PyPI or the used mirror is not responding. It is up to the client to decide wich mirror should be used, maybe by looking at its geographical location and its responsivness. This PEP does not describe how this fail-over mechanism should work, but it is strongly encouraged that the clients try to use the nearest mirror. The clients so far that could use this mechanism: - setuptools - zc.buildout (through setuptools) - pip Extra package indexes ::::::::::::::::::::: It is obvious that some packages will not be uploaded to PyPI, whether because they are private or whether because the project maintainer runs his own server where people might get the project package. However, it is strongly encouraged that a public package index follows PyPI and Distutils protocols. In other words, the `register` and `upload` command should be compatible with any package index server out there. Software that are compatible with PyPI and Distutils so far: - PloneSoftwareCenter [#psc]_ wich is used to run plone.org products section. - EggBasket [#eggbasket]_. **An extra package index is not a mirror of PyPI, but can have some mirrors itself.** Merging several indexes ::::::::::::::::::::::: When a client needs to get some packages from several distinct indexes, it should be able to use each one of them as a potential source of packages. Different indexes should be defined as a sorted list for the client to look for a package. Each independent index can of course provide a list of its mirrors. XXX define how to get the hostname for the mirrors of an arbitrary index. That permits all combinations at client level, for a reliable packaging system with all levels of privacy. It is up the client to deal with the merging. References ========== .. [#pep305] http://www.python.org/dev/peps/pep-0305/#id19 .. [#pep381client] http://pypi.python.org/pypi/pep381client .. [#iso8601] http://en.wikipedia.org/wiki/ISO_8601 .. [#zc.buildout] http://pypi.python.org/pypi/zc.buildout .. [#setuptools] http://pypi.python.org/pypi/setuptools .. [#pip] http://pypi.python.org/pypi/pip .. [#psc] http://plone.org/products/plonesoftwarecenter .. [#eggbasket] http://www.chrisarndt.de/projects/eggbasket Acknowledgments =============== Georg Brandl. Copyright ========= This document has been placed in the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End: