[Distutils] PyPI Migrated to New Infrastructure with some Breakage

[Chris Jerdonek]

On Sat, Jan 25, 2014 at 3:38 PM, Donald Stufft <donald at stufft.io> wrote:
> Today (Sat Jan 25, 2014) the Infrastructure team has migrated PyPI to new
> infrastructure.
>
> The old infrastructure was:
>
> - a single database server managed by OSUOSL
> - a pair of load balancers shared by all of the python.org services hosted on
>   OSUOSL
> - a single backend VM that served as everything else for PyPI
>
> The VM that was acting as the backend server from PyPI was partially hand
> configured and partially setup to be managed by chef. Additionally it had an
> issue that caused it to kernel panic every so often which had been the cause of
> a number of downtimes in the last few months. Because it was primarily
> configured and administered by hand and because the way it was set up it was
> not feasible to have any sort of failover or spare.
>
> The new infrastructure is:
>
> - 2 Web VMs
> - 2 Database servers in Master/Slave Configuration
> - 2 PgPool Servers pooling connections to the database servers and load
>   balancing reads across them.
> - 2 GlusterFS servers backed by Cloud Block Storage acting as the file storage
>   for package and package docs
> - 1 metrics server to handle updating the download counts as they come in from
>   Fastly
>
> All of the VMs are hosted on Rackspace’s Public Cloud and have their
> configuration completely controlled and managed using Salt. Going forward this

Can you say a little about the choice to use Salt instead of Chef?  I
don't really care either way, but am just curious.  Is it because Salt
is written in Python, or were there other reasons (functionality,
etc)?

--Chris


> will allow us to easily scale out as required or kill malfunctioning servers
> and spin up new ones easily. Additionally the setup has been setup so that
> where possible there is two servers performing the same role, ideally in an
> Active/Active configuration but at least in a Master/Slave configuration. This
> should allow PyPI to be far more stable moving forward and make downtimes much
> easier to recover from.
>
> The services are still fronted by Fasty’s CDN and in the new infrastructure
> we’ve removed our load balancer and have replaced it by having Fastly handle
> the load balancing for us. Additionally we’ve recently setup a static mirror of
> PyPI that is updated once every minute. This is hosted on Rackspace cloud as
> well but in a separate data center from the rest of PyPI. Fastly is configured
> to fall back to this static mirror in the case that neither of the two web
> heads are functioning. This should ensure that even in the event of a
> catastrophic failure of the PyPI service that the bulk of package installations
> should hopefully remain working.
>
> The bad news, (and the “Breakage” from the subject) is that while the new
> infrastructure was being planned out, built, and migrated to the “pypissh”
> package was forgotten. The pypissh package is an alternative way to upload
> packages to PyPI however it is very difficult, because of the way it works, to
> provide HA support for it as we’ve set up for everything else. We don’t have
> any numbers for how many people are actively using this package but looking
> at a roughly 2 week chunk of time in PyPI’s download history, the pypissh
> package was downloaded 7 times by a browser, and 7 times by pip. All other
> downloads were caused by the mirroring system.
>
> As of right now pypissh is non functional and due to the difficulty in HAing
> and monitoring the current setup and because it is apparently has a very
> small set of users we would like to effectively kill off this particular
> service. Additionally the benefits of pypissh have been reduced now that PyPI
> is available over a TLS connection with a well trusted certificate. My question
> to you is, is this something that distutils-sig is willing to have happen? If
> we are to re-enable pypissh we’ll need to write a new solution to doing it that
> can be properly HA’d and we’d prefer to put our efforts into improving things
> for a much larger set of people.
>
> So yea, PyPI should be loads more stable and more reliable now.
>
> -----------------
> Donald Stufft
> PGP: 0x6E3CBCE93372DCFA // 7C6B 7C5D 5E2B 6356 A926 F04F 6E3C BCE9 3372 DCFA
>
>
> _______________________________________________
> Distutils-SIG maillist  -  Distutils-SIG at python.org
> https://mail.python.org/mailman/listinfo/distutils-sig
>