Writing workers

[William Reade]

Hi all

I've noticed that there's a lot of confusion over how to write a useful
worker. Here follow some guidelines that you should be *very* certain of
yourself before breaking (and probably talk to me about anyway). If there's
any uncertainty about these, I'm more than happy to expand.

  * If you really just want to run a dumb function on its own goroutine,
use worker.NewSimpleWorker.

  * If you just want to do something every <period>, use
worker.NewPeriodicWorker.

  * If you want to react to watcher events, you should probably use
worker.NewNotifyWorker or worker.NewStringsWorker.

  * If your worker has any methods outside the Worker interface, DO NOT use
any of the above callback-style workers. Those methods, that need to
communicate with the main goroutine, *need* to know that goroutine's state,
so that they don't just hang forever.

  * To restate the previous point: basically *never* do a naked channel
send/receive. If you're building a structure that makes you think you need
them, you're most likely building the wrong structure.

  * If you're writing a custom worker, and not using a tomb.Tomb, you are
almost certainly doing it wrong. Read the blog post [0] or, hell, just read
the code [1] -- it's less than 200 lines and it's about 50% comments.

  * If you're letting tomb.ErrDying leak out of your workers to any
clients, you are definitely doing it wrong -- you risk stopping another
worker with that same error, which will quite rightly panic (because that
tomb is *not* yet dying).

  * If it's possible for your worker to call .tomb.Done() more than once,
or less than once, you are *definitely* doing it very very wrong indeed.

  * If you're using .tomb.Dead(), you are very probably doing it wrong --
the only reason (that I'm aware of) to select on that .Dead() rather than
on .Dying() is to leak inappropriate information to your clients. They
don't care if you're dying or dead; they care only that the component is no
longer functioning reliably and cannot fulfil their requests. Full stop.
Whatever started the component needs to know why it failed, but that parent
is usually not the same entity as the client that's calling methods.

  * If you're using worker/singular, you are quite likely to be doing it
wrong, because you've written a worker that breaks when distributed. Things
like provisioner and firewaller only work that way because we weren't smart
enough to write them better; but you should generally be writing workers
that collaborate correctly with themselves, and eschewing the temptation to
depend on the funky layer-breaking of singular.

  * If you're passing a *state.State into your worker, you are almost
certainly doing it wrong. The layers go worker->apiserver->state, and any
attempt to skip past the apiserver layer should be viewed with *extreme*
suspicion.

  * Don't try to make a worker into a singleton (this isn't particularly
related to workers, really, singleton is enough of an antipattern on its
own [2] [3] [4]). Singletons are basically the same as global variables,
except even worse, and if you try to make them responsible for goroutines
they become more horrible still.

Did I miss anything major? Probably. If so, please remind me.

Cheers
William


[0] http://blog.labix.org/2011/10/09/death-of-goroutines-under-control
[1] launchpad.net/tomb (apparently... we really ought to be using v2,
though)
[2] https://sites.google.com/site/steveyegge2/singleton-considered-stupid
[3]
http://jalf.dk/blog/2010/03/singletons-solving-problems-you-didnt-know-you-never-had-since-1995/
[4]
http://programmers.stackexchange.com/questions/40373/so-singletons-are-bad-then-what/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.ubuntu.com/archives/juju-dev/attachments/20150528/4296eb68/attachment.html>