Description of the problem
Fedora Infrastructure is composed of a number of services (koji, fedpkg, pkgdb, etc..) some of which are maintained outside the Fedora Project and some of which were built in-house by the infrastructure team. These are strung together in a pipeline. Think "how an upstream release becomes a package update", "How a new source distribution becomes a package."
At present, many of the steps in this process require the maintainer to wait and watch for a previous step to complete. For instance once a branch of a package is successfully built in koji, the maintainer must submit their update to bodhi (See the new package process for more details).
Other progressions in the pipeline are automated. For instance, AutoQA defines a set of watchers. Most watchers are run as a cron task. Each one looks for certain events and fires off tests when appropriate.
At LinuxFest Northwest (2009), jkeating gave a talk on the problem of complexity in the Fedora infrastructure and how this might be addressed with a message bus architecture. Each service in the infrastructure depends on many of the others. Some pieces directly poke others: git (fedpkg) currently pokes AutoQA from a post-update hook. Other pieces poll others' status: koji scrapes pkgdb for package-owner relationships and email aliases.
This dense coupling of services makes changing, adding, or replacing services more complicated: commits to one project require a spidering of code changes to all the others.
How messaging might address the problem
jkeating's talk on messaging in the Fedora Instructure proposed the adoption of a unified message bus to reduce the complexity of multiple interdependent services. Instead of a service interfacing with its dependencies' implementations, it could subscribe to a topic, provide a callback, and respond to events.
For instance, instead of having koji scrape pkgdb on an interval for changed
email addresses, pkgdb could emit messages to the
org.fedoraproject.service.pkgdb topic whenever an account's email address
changes. koji could subscribe to the same topic and provide a callback that
updates its local email aliases when invoked.
In another case, the git (fedpkg) post-update hook could publish messages to
org.fedoraproject.service.fedpkg.post-update topic. AutoQA could
subscribe to the same. Now if we wanted to enable another service to act when
updates are pushed to fedpkg, that service need only subscribe to the topic and
implement its own callback instead of appending its own call to fedpkg's
post-update hook (instead of coupling its own implementation with fedpkg's).
A message bus architecture, once complete, would dramatically reduce the work required to update and maintain services in the Fedora infrastructure.
By adopting a messaging strategy for Fedora Infrastructure we could gain:
- A stream of data which we can watch and from which we can garner statistics about infrastructure activity.
- The de-coupling of services from one another.
- libnotify notifications to developers' desktops.
- jquery.gritter.js notifications to web interfaces.
- this could be generalized to a
fedmsg.wsgimiddleware layer that injects a fedora messaging dashboard header into every page served by apps X, Y, and Z.
- An irc channel,
#fedora-fedmsgthat echoes every message on the bus.
- An identi.ca account, @fedora-firehose, that echoes every message on the bus.
AMQP, and 0mq
AMQP or "Broker? Damn near killed 'er!"
The following is recreated from J5's Publish/Subscribe Messaging Proposal as an example of how Fedora Infrastructure could be reorganized with AMQP and a set of federated AMQP brokers (qpid).
The gist is that each service in the Fedora Infrastructure would have the
address of a central message broker on hand. On startup, each service would
connect to that broker, ask the broker to establish its outgoing queues, and
begin publishing messages. Similarly, each service would ask the broker to
establish incoming queues for them. The broker would handle the routing of
messages based on
routing_keys (otherwise known as topics) from each
service to the others.
The downshot, in short, is that AMQP requires standing up a single central broker and thus a single-point-of-failure. In the author's work on narcissus I found that for even the most simple of AMQP configurations, my qpid brokers' queues would bloat over time until *pop*, the broker would fall over.
0mq or "Going for Broke(rless)"
0mq is developed by a team that had a hand in the original development of AMQP. It claims to be a number of things: an "intelligent transport layer", a "socket library that acts as a concurrency framework", and the sine qua non "Extra Spicy Sockets!"
- Recommended reading:
The following depicts an overview of a subset of Fedora Infrastructure organized with a decentralized 0mq bus parallel to the spirit of J5's recreated diagram in the AMQP section above.
No broker. The gist is that each service will open a port and begin publishing messages ("bind to" in zmq-language). Each other service will connect to that port to begin consuming messages. Without a central broker doing all the things, 0mq can afford a high throughput. For instance, in initial tests of a 0mq-enabled moksha hub, the Fedora Engineering Team achieved a 100-fold speedup over AMQP.
Shortly after you begin thinking over how to enable Fedora Infrastructure to pass messages over a fabric instead of to a broker, you arrive at the problem we'll call "service discovery".
In reality, (almost) every service both produces and consumes messages. For the sake of argument, we'll talk here just about a separate producing service and some consuming services.
Scenario: the producing service starts up a producing socket (with a hidden queue) and begins producing messages. Consuming services X, Y, and Z are interested in this and they would like to connect.
With AMQP, this is simplified. You have one central broker and each consuming service need only know it's one address. They connect and the match-making is handled for them. With 0mq, each consuming service needs to somehow discover its producer(s) address(es).
There are a number of ways to address this:
- Write our own broker; this would not be that difficult. We could (more simply) scale back the project and write our own directory lookup service that would match consumers with their providers. This could be done in surprisingly few lines of python. This issue is that we re-introduce the sticking point of AMQP, a single point of failure.
- Use DNS; There is a helpful blog post on how to do this with djbdns. DNS is always there anyways: if DNS goes down, we have bigger things to worry about than distributing updates to our messaging topology.
- Share a raw text file; This at first appears crude and cumbersome:
- Maintain a list of all fedmsg-enabled producers in a text file
- Make sure that file is accessible from every consuming service.
- Have each consuming service read in the file and connect to every (relevant) producer in the list
In my opinion, using DNS is generally speaking the most elegant solution. However, for Fedora Infrastructure in particular, pushing updates to DNS and pushing a raw text file to every server involves much-the-same workflow: puppet. Because much of the overhead of updating the text file falls in-line with the rest of Infrastructure work, it makes more sense to go with the third option. Better not to touch DNS when we don't have to.
In the above examples, the topic names are derived from the service names.
For instance, pkgdb publishes messages to
org.fedoraproject.service.pkgdb*, AutoQA presumably publishes messages
org.fedoraproject.service.autoqa*, and so on.
This convention, while clear-cut, has its limitations. Say we wanted to replace pkgdb whole-sale with our shiney new threebean-db (tm). Here, all other services are subscribed to topics that mention pkgdb explicitly. Rolling out threebean-db will require patching every other service; we find ourselves in a new flavor of the same complexity/co-dependency trap described in the first section.
The above service-oriented topic namespace is one option.
Consider an object-oriented topic namespace where the objects are things
like users, packages, builds, updates, tests, tickets, and composes. Having
bodhi subscribe to
org.fedoraproject.object.builds leaves us less tied down to the current
implementation of the rest of the infrastructure. We could replace bugzilla
with pivotal and bodhi would never know the difference - a ticket is a
That would be nice; but there are too many objects in Fedora Infrastructure that
would step on each other. For instance, Koji tags packages and Tagger
tags packages; these two are very different things. Koji and Tagger cannot
both emit events over
org.fedoraproject.package.tag.* without widespread
Consequently, our namespace follows a service-oriented pattern.
Event topics will follow the rule:
ENVis one of dev, stg, or production.
SERVICEis something like koji, bodhi, or fedoratagger
OBJECTis something like package, user, or tag
SUBOBJECTis something like owner or build (in the case where
OBJECTis package, for instance)
EVENTis a verb like update, create, or complete.
All 'fields' in a topic should:
- Be singular (Use package, not packages)
- Use existing fields as much as possible (since complete is already used by other topics, use that instead of using finished).