Note: This was published in 2016. A update for 2021 is here.
The Python Package Index, or as most call it “PyPI” is a central part of the ecosystem of Python. It serves as a central registry of names, helping to prevent collision between different projects as well as the default repository that most Python users go to when looking for software. Though it has its issues (of which, there are many) it functions as a critical part of the user experience of Python and is a critical part of the development infrastructure for a veritable army of developers. It may not be the largest site in the world, but it’s certainly a significant one, having used 293.1TB of bandwidth serving 3 billion HTTP requests during the month of April 2016 alone.
For most, what powers this service is largely opaque to them — it’s (usually) there when they need it and who or what powers it is largely a mystery to them, but what and who really powers PyPI? 1
I suspect that for many people this is going to come as the largest surprise, although probably not for anyone who is familiar with how little investment F/OSS, and F/OSS infrastructure in particular tends to get, but when you’re talking about the people powering the current version of PyPI you’re largely talking about 3 people in total.
Donald Stufft (Me)
When it comes to PyPI the vast bulk of what happens falls squarely on my shoulders. This includes some of the good parts but also most of the bad parts as well. Most decisions that end up getting made are made or passionately advocated for by me and very little gets done without me touching it in some way. My responsibilities range from maintenance of the current code base (aka “Legacy PyPI”), to development of the new replacement (aka “Warehouse”), to operations and being one of the defacto on call persons 24/7 2.
Working on PyPI (along with pip and the general packaging ecosystem) is my full time job which means at a minimum I spend ~40 hours a week working on one of these pieces of the ecosystem (lately it’s been largely PyPI). That minimum of 40 hours is very rarely what I actually spend working on these projects, my total time very often looks more like 70-90 hours a week (or more!) 3 that I spend doing some sort of packaging related task 4.
When it comes to PyPI and ops, most of this has been done by Ernest. When PyPI migrated to it’s new infrastructure host Ernest more or less single handedly wrote the entire salt states and pillars (339 out of 410 commits). He is one of the people who are actually on call and has been a massive help in architecting the modern PyPI infrastructure to handle the load we’ve placed on it as well as be as reliable as we can get it given the state of the code that powers all of this 5.
The father of PyPI! Richard wrote PyPI all those years ago and has stood as one of the administrators since then. While he has stepped back from the day to day development and maintenance of PyPI he is still the main person dealing with support requests on the issue tracker.
Companies / Services
Given the skeleton crew that we run PyPI with, we decided a few years ago to try and push as much of the operations work as we can onto external services. This allows us to spend less time doing day to day operations and helps narrow it down to largely incident response (security updates, downtime, etc). If it wasn’t for the companies here 6 we simply would not be able to keep up with work load (to whatever extent we’re able to “keep up” today). Each of these companies donate their services to the running of PyPI, and PyPI is all the better for it.
I recently did the math, and omitting HPE and my salary from the equation and focusing on just the actual hosting costs, all together these companies are donating roughly $35,000 a month worth of services towards keeping PyPI up and running.
Hewlett Packard Enterprise
Technically Hewlett Packard Enterprise (HPE) is not actually donating a service, but what they are essentially donating is me. They employ me full time to work on all of this and give me free reign to work on the areas that I think most need it. Without HPE the time I have would be severely reduced.
Fastly is PyPI’s secret scaling sauce. The global CDN and Varnish caching provided by Fastly provides one of the single largest reduction in operations effort that we have. Every request to PyPI goes through Fastly and through their caching we’re able to prevent almost 90% of the incoming traffic from ever reaching the origin servers. This also allows us to take advantage of the cache (including stale cached items) so that whenever we do have downtime on the origin servers it affects fewer people as we’re able to serve it out of the cache rather than giving the end user an error.
The bulk of the infrastructure that we (the PSF) actually run is running on cloud servers provided by Rackspace. Without this we’d have nothing to actually run PyPI itself on and let me tell you, coming from a time when PyPI was ran on a single physical host the ability to spin up new machines at will and have actual support people making sure those machines keep running has decreased the MTTR (Mean-Time-To-Recovery) by orders of magnitude.
PyPI hosts files, and unsurprisingly it needs a place to store those files, which we do in the excellent object store by Amazon, S3. While there’s not a lot you can say about storage except that we stick our files in, and they make sure we don’t lose them and they’re available whenever someone tries to download a file.
Other than files, we also need a database for PyPI which is provided by the excellent Heroku Postgres service. Offloading our database to Heroku allows us to forget about dealing with HA and Master elections, fail over, and back ups.
The Elasticsearch cluster is running on Elastic Cloud, allowing us to free up VM resources from hosting ES to other more specialized tasks and also alleviating the need to keep Elasticsearch monitored and operational.
More than the above donate to the running of PyPI in ways that aren’t on the critical path of serving a web request, like:
- Dreamhost, providing long term log archival inside of DreamObjects.
- Statuspage.io, providing a managed status page hosted outside of our own infrastructure.
- Sentry, providing error reporting.
- Dyn, providing DNS hosting.
Due to a volunteer (or donated) work force and running on pretty much entirely donated infrastructure the actual real monetary cost of PyPI is very low. However, wherever we do need to spend money, that funding comes from the PSF so the funding comes from donations to the PSF as well as the income generated by events like PyCon.
This shows a snapshot of the people, services, and money that powers PyPI today in May of 2016. It doesn’t reflect efforts outside of the current PyPI such as Warehouse, pip, setuptools, the ongoing PEPs, etc. As we move forward this is bound to become outdated as our infrastructure changes.
Over the years many different people and services have participated in the maintenance and running of PyPI. I’m not going to attempt to create an exhaustive list of every person or service who has ever helped, but instead focus on those who are currently involved. This is not to lessen the impact of those who came before, but merely to avoid ballooning an already long post even more. Even limiting to just the present, I’m sure that I’ve missed someone and for that I am deeply sorry! ↩︎
I’m very rarely actually on call. We do have Pager Duty and Pingdom setup to monitor PyPI that will send notifications to a mailing list when things go down and also to Ernest Durbin and Noah Kantrowitz. However, I’m fairly publicly associated with PyPI and it’s rare that an issue occurs with it that doesn’t end up with people pinging me on IRC, Twitter, Email, etc, often times faster than Pingdom can notice. Realistically the only way for me to avoid notification whenever PyPI has downtime is to disconnect myself completely from any public method of contacting me. ↩︎
You want to know who the real heros of F/OSS are? It’s the spouses, families, and friends of those who work on it who put up with someone they know taking time they could be spending with them and choosing to help out a bunch of strangers on the internet. There are times when the only time I really interact with my wife is lunch and the 10 minutes between when I lay down at night and finally pass out. She often times calls herself a “Software Widow” with a varying level of amusement depending on long it’s been since I had a particularly overloaded week. ↩︎
Fun fact, the code that powers all of this? Yea it was a weekend hack for a proof of concept that was intended to get quickly replaced by the real code. Over a decade later? Yea that became critical infrastructure for most Python developers. This code predates modern Frameworks by years and it either predates or was a contemporary of WSGI. It’s had numerous things bolted onto the side of it over the years, turning it into
Frankenstein’sJones’ Monster. ↩︎
Much like with the people, there have been a variety of companies willing to provide services over the years, some of which we’ve either stopped using or have stopped offering it. While we thank them for the help they’ve given, this post is largely focusing on the present. I’m also sure that I’ve missed some, and for that I am sorry! ↩︎