Single use forks

[SDQL]

Hi,

Is it possible to use a new fork for each client request?

My problem is that some client requests require a lot of RAM which is
not recovered after the request finishes.

Thanks in advance for any clues!


Joe

[Ben Darnell]

There's not a way to do this out of the box - forking per-request
isn't really compatible with the idea of a shared event loop. If you
want to use the full IOLoop stack you'll need to devise some way
(probably using the multiprocessing module) to farm work out to a
child process. If you don't need IOLoop you may be able to use
WSGIApplication in a fork-based WSGI container.

-Ben

[Samuel]

I don't know if my approach is the best but I choose to use 2 tornado
server instances. One for very light weight requests or ones that can
be easily async'ed. Then I have another process based tornado
server. For my lightweight instance for calls I deem long running I
make an async call to the process server and wait for the response.

Personally I strongly dislike maintaining two separate servers like
this, but honestly it works like a champ. I get the best of both
worlds even if it is a little bit more to manage.

Regards,
Samuel

[Shane Spencer]

I have found appreciation of inetd for this task.

[SDQL]

Tornado friends,

Thanks for the good thoughts.

My solution to the problem of RAM accumulation in long running process
is to kill forks that use more than (eg) 20% of their starting RAM
with:

def on_finish(self):
if resource.getrusage(resource.RUSAGE_SELF).ru_maxrss > 1.2 *
START_RU_MAXRSS:
sys.exit(1)

Where I define the global START_RU_MAXRSS on spin up.

Since I exit with a non zero signal, tornado is kind enough to restart
a fork for me.

In order to have this work past the default of 100 restarts, I added a
max_restarts parameter to netutils.start and modified one line in
process.fork_process to read:

if max_restarts and num_restarts > max_restarts:
raise RuntimeError("Too many child restarts, giving up")

So that when I start my server like this:

server.start(options.forks,max_restarts=None)

it spawns new forks forever.

It seems kind of bad to use sys.exit(1) like this but it works great
so far!

Cheers,

Joe

http://SportsDataBase.com

[jeremy]

Joe,

This is an interesting idea, but wouldn't it kill any jobs on the
ioloop prior to their callback and therefore prematurely close
connections?

Just curious. Am I missing something?

-jeremy

--
jeremy kelley - ceo, nerd, etc.
snoball - encouraging a lifestyle of giving.

e: jer...@snoball.com
o: 1.855.sno.fort
c: 979.820.8202
t: @nod

[SDQL]

Good thinking!

From earlier discussion here, this looks like a more graceful way to
restart a bloated tornado fork:

def on_finish(self):
if resource.getrusage(resource.RUSAGE_SELF).ru_maxrss > 1.2 *
START_RU_MAXRSS:

server.stop()
tornado.ioloop.IOLoop.instance().add_timeout( time.time()
+ SPINDOWN_TIME, lambda :sys.exit(1) )


Cheers,

Joe

http://SportsDataBase.com