Thoughts on stateful services in Guix

[Alex Sassmannshausen <alex.sassmannshausen@gmail.com>]

Hello,

As a result of FOSDEM conversations today I felt inspired to put some
thoughts on paper about an area where I think we currently run into
complications.

Not sure if it is appropriate to write this blog-post-ish contribution
here, but as I don't have a blog, and it's about Guix development, I
figured it might be OK.

Best wishes,

Alex

1 Introduction
══════════════

  Guix is amazing.  A large part of why it continues to be amazing is
  because it provides strong guarantees to the end-user and developer.
  As such it makes reasoning about packages and deployments relatively
  straight-forward.

  The functional paradigm cleary works fantastically for packages.
  Unfortunately it is not quite clear that it works just as well for
  services.  The reason for this is that too many useful services are
  inherently stateful.  Their statefulness means they have side-effects,
  which in turn cause issues when relying on Guix features such as
  roll-back or automated deployment & guaranteed reproducability.

  Disciplined developers can implement many services in such a way that
  their statefulness is delegated to other dedicated services.  In this
  way the problem of statefulness can be isolated.  However, many
  existing useful services /have not/ been implemented in such a way.

  These notes are an attempt to think through ways of formalising how we
  mitigate stateful services in Guix.

  The notes below are organised around the example of a popular PHP
  content management system, Drupal.  Similar problems will apply to
  many other end-user services.

  My intent here is to communicate my thinking in the hopes that others
  can point to obvious flaws in my reasoning — or to stimulate
  conversation about the topic.  I hope at the very least that it is an
  interesting read!


2 Learning by Doing: Packaging & Deploying Drupal
═════════════════════════════════════════════════

2.1 Problem (1): the Drupal tarball is a binary blob!
─────────────────────────────────────────────────────

  Released Drupal tarballs are shipped with a bunch of PHP dependencies,
  as well as compiled JS files.  A fully source-distributed installation
  of Drupal would:
  1) delete all shipped PHP dependencies
  2) independently build all PHP dependencies and make them available to
     Drupal (through it's vendor directory as symlinks?)
  3) delete all compiled JS files & libraries
  4) independently build the JS dependencies and make them avalable to
     Drupal.


2.2 Packages must always be stateless!
──────────────────────────────────────

  What does this mean in practice?  Let's look at Drupal again.  When
  you download Drupal, the resulting tarball contains the source code,
  and an empty sites/ folder.  The sites folder is intended to contain
  /state/ files.  The normal installation procedure is to simply drop
  the drupal distribution in your web root directory, and for state
  files to live underneath sites/ within your webroot.

  In Guix, Drupal is packaged so that it is installed in the store.

  The store is read-only and hence no /state/ files can live under the
  sites/ directory in the store.

  How do we get around this?
  1) either we patch drupal to expect the sites/ directory outside of
     its own folder tree.
  2) or we symlink /gnu/store/…drupal…/sites/ to a different location on
     the filesystem (e.g. /var/lib/drupal/sites/)

  The latter solution, while easy, means that a *successful*
  installation of the package Drupal in Guix results in an installation
  in the store with a *broken symlink* pointing outside the store.

  But the package itself has been rendered stateless!


2.3 Drupal: a stateful service
──────────────────────────────

  Services in Guix are rich and multidimensional entities.  At core they
  are promises of things that will have happened when a system is up and
  running.  These promises can be arbitrary, like generating a
  configuration file every boot; or they can be an extension of other,
  already existing services.

  A particularly popular service to extend is the shepherd service,
  which ensures that particular daemons are started as soon as possible
  after the system has started.

  What would a Drupal service look like?  Essentially the software is
  just a bunch of files in a webroot — so at it's heart it simply
  extends a web service (e.g. nginx) with new location directives.

  On the other hand it also requires that tha web server, a sql backend
  and php-fpm are running, so that drupal can actually function.

  These two requirements are easily met with the usual service
  infrastructure: simply extend nginx with a location definition
  pointing to the Drupal folder in the store as the webroot and extend
  shepherd to require mysql, nginx and php-fpm.


2.3.1 Enter the state dragon
╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌

  But hang-on… if Mysql service is a dependency, and we store state in a
  Mysql DB, what happens when we upgrade to a newer version of Drupal,
  with a different schema?  Would this service instantiation change the
  state of our system?  Could we still roll-back?

  And remember that broken symlink we introduced in the package?  That
  files/ subdirectory contains /at least/ a stateful configuration file
  — but also Drupal modules, themes, uploaded files and caches.  What
  happens when we upgrade Drupal?  Can we guarantee they won't be
  irrevocably changed as part of the upgrade?  What happens when we
  roll-back?

  The truth is that we have introduced state into our deployment and we
  hence lose a lot of the beautiful hard guarantees that Guix provides
  us with.


2.3.2 Fakin' it til we're makin' it
╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌

  Can we get those guarantees back?  I fear not.  But we can add some
  shims and mitigations that might take the edge off the sting.

  We've elaborated, above, the extend of the state problem in Drupal.
  What would a /state shim/ look like?

  Working backwards, to ensure we can roll-back we need to make sure
  that we have access to the state at that point in history that we want
  to roll-back to.  When we roll-back from version 2 of Drupal to
  version 1, we must make sure we still have the state of version 1.

  So therefore we need to simply make an addressable state dump just
  before we activate a new service instance, so that when we roll-back,
  we can flip back to that state dump.  In the case of our Drupal
  example, this means that we need to dump the sites/ directory and the
  mysql databases.

  This state dump needs to fulfill the following criteria:
  1) It needs to be reversible: a transaction into a new state needs to
     make a dump before transition and then carry out state changes; a
     transaction into a past state must make a dump before transition
     and then load the old state's dump as it's new state.
  2) It (optionally) needs to be secured: state will likely contain
     personal information, and this information needs to be protected.


2.4 The stateful service shimmy
───────────────────────────────

  Here's the proposed general shim then:

  1) We define a new type of shepherd service, a /stateful service/,
     which has additional actions: state-dump-shim and state-load-shim.
  2) Each stateful service (e.g. drupal) is responsible for, through
     provision of imperative code snippets in the state-dump-shim and
     state-load-shim fields, providing state dumpers/restorers.
  3) Whenever a new service revision is installed:
     1) before it is installed, state-dump-shim is called.  It should
        dump the service's state in a way that when state-load-shim is
        called with the new service revision's store hash, it can locate
        any state dumps that were stored for that particular hash.
     2) after it is installed, state-load-shim is called with the store
        hash of the new service revision.  It should then be able to
        locate and load any state associated with this hash.  If no
        state is associated (e.g. when we install a wholly new
        revision), nothing is done.

  What does this give us?

  When a migration to a new service revision causes a deployed bit of
  software to stop working as expected, we can do a `guix system
  –roll-back`.  This will then also restore the service's state to the
  previous version and thus *should* restore operations.

  *BEWARE*: any state changes that occured since the previous state-dump
  /will/ be lost.  We are restoring to a backup — so user generated
  state that was added after the backup will not be present after
  restoration!


3 A generalisation: the Stateful-Service Service
════════════════════════════════════════════════

  State dumping and restoration *should* be generalisable.

  It should normally consist of one or more operations of:
  • running a special program to dump data, encrypting the dump, and
    storing it, together with the service revision hash, in a known
    location.
  • tarring up a directory tree, encrypting the archive and storing it,
    together wtih the service revision hash, in a known location.

  These operations should be able to be provided by a daemon, managed by
  shepherd, which can be configured with a gpg key for encryption, a
  mechanism for decryption (interactive or programmatic), a location for
  storing/retrieving dumps (local or remote?), and a DSL for mapping
  data dump / data restoration program invocations or file-system
  locations to data dumps it already knows of.