Re: feature request - when_ready() hook

[Eric Wong <normalperson@yhbt.net>]

Suraj Kurapati <sunaku@gmail.com> wrote:
> On Wed, Nov 25, 2009 at 10:05 PM, Eric Wong <normalperson@yhbt.net> wrote:
> > Suraj Kurapati <sunaku@gmail.com> wrote:
> >> * before_fork() -- approx. 2 minute downtime
> >> * after_fork() -- approx. 2 minute downtime
> >> * storing the old-master-killing logic inside a lambda in after_fork()
> >> (for the last worker only) and later executing that lambda in Rails'
> >> config.after_initialize() hook -- approx. 20 second downtime
> >
> > I'm looking at those times and can't help but wonder if there's
> > something very weird/broken with your setup..  20 seconds is already an
> > eternity (even with preload_app=false), but 2 minutes?(!).
> 
> Yes, I am using preload_app=false.  These delays mainly come from
> establishing DB connections and loading XML datasets into the Rails
> app.  Our production DBs are pretty slow to give out new connections.
> The startup time is much faster in development, where I use SQLite.

Yikes.  Is there some sort of misconfiguration in the DBs?  Perhaps
they're trying to do reverse DNS on every client connection, that
can really ruin your day if your app servers don't have reverse DNS.

> Please note that the reported downtimes are shocking only because they
> were *visible* downtimes, where the last worker of the new Unicorn
> master killed the old Unicorn master too soon.  IMHO, it doesn't
> matter how long it takes for the Rails app to become ready, so long as
> the old Unicorn master + workers continue to exist & service requests
> until the new Unicorn master + workers take over.

<snip>

> > Are you using preload_app?  It should be faster if you do, but there
> > really appears to be something else wrong based on those times.
> 
> I was for a few weeks, but I stopped because the XML dataset loading
> (see above) kept increasing the master's (and the new set of workers')
> memory footprint by 1.5x every time Unicorn was restarted via SIGUSR2.

Side problem, but another thing that makes me go "Huh?"

Did the new master's footprint increase?  Are you mmap()-ing the XML
dataset?  Is RSS increasing or just VmSize?  Unicorn sets FD_CLOEXEC on
the first 1024 (non-listener) file descriptors, so combined with exec(),
that should give the new master (and subsequent workers) a clean memory
footprint.

> >> As you can see, the more I delayed the execution of that "killing the
> >> old master" logic, the closer I got to zero downtime deploys.  In this
> >> manner, I request the addition of a when_ready() hook which is
> >> executed just after Unicorn prints "worker=# ready!" to its error log
> >> inside Unicorn::HttpServer#worker_loop().
> >
> > At this stage, maybe even implementing something as middleware and
> > making it hook into request processing (that way you really know the
> > worker is really responding to requests) is the way to go...
> 
> Hmm, but that would incur a penalty on each request (check if I've
> already killed the old master and do it if necessary).  I'm pretty
> confident that killing the old master in the when_ready() hook will be
> Good Enough for my setup (at most I expect to see 1-2 second
> "down"time).  Let me try this out and I'll tell you the results &
> submit a patch.

I don't think a runtime condition would be any more expensive than all
the routing/filters/checks that any Rails app already does and you can
cache the result into a global variable.

As you may have noticed, I'm quite hesitant to add new features,
especially for uncommon/rare cases.  Things like supporting the
"working_directory" directive and user-switching took *months* of
debating with myself before they were finally added.

> >> my unicorn setup does not run very close to the memory limit of
> >> its host; instead, the amount of free memory is more than double of
> >> the current unicorn memory footprint, so I can safely spawn a second
> >> set of Unicorn master + workers (via SIGUSR2) without worrying about
> >> the SIGTTOU before_fork() strategy shown in the Unicorn configuration
> >> example.)
> >
> > Given your memory availability, I wouldn't even worry about the
> > automated killing of the old workers.
> >
> > Automatically killing old workers means you need a redeploy to roll back
> > changes, whereas if you SIGWINCH the old set away, you can HUP the old
> > master to bring them back in case the new set is having problems.
> 
> Wow this is cool.  Perhaps this strategy could be mentioned in the
> documentation?

It's already at the bottom of the SIGNALS
(http://unicorn.bogomips.org/SIGNALS.html) document

> Thanks for your consideration.

No problem, let us know if it's the DB doing reverse DNS because
I've seen that to be a problem in a lot of cases.

-- 
Eric Wong