YOU WILL REGRET THIS

YOU WILL REGRET THIS!My Other Mood is 'Even More Angry'

Friday, June 10, 2011

I'M GONNA TELL YOU A STORYIn which I tell you how to behave

- how dare I?


SOCKET.IO-ERLANG

•Developed by Yurii Rashkovskii, Omar Yasin, Fred Hebert (me)

• Used as a base demo app for this presentation

• Allows to write neat front-ends for conversational websites


BIZARRO SOCKET.IO-ERLANG

• 'Bad' socket.io-erlang used for the presentation

• Freely substitutes dependencies and actual code for whatever purpose I have


APP STRUCTURE


APP STRUCTUREBunch of supervisors



web server



web servertransport handlers



web servertransport handlers

Client code attaches here


WE USE A WEB SERVER

•We need a web server (the slide's title says just that (and also the previous slide))

•Oh, Carrying parameters and a connection around is annoying!

• Let's use parametrised modules!


PARAMETRISED MODULES

-‐module(test_module, [Param1]).some_method() -‐> Param1.

Equivalent to:

-‐module(test_non_pmod).some_method(Param1) -‐> Param1.



1> X = test_module:new(horror), X:some_method().horror

2> test_non_pmod:some_method(sane).sane



• They are great!

•We can use them as if they were objects!

• They let us carry everything as one large parameter!

• Let's do this!



• They are great!

•We can use them as if they were objects!

• They let us carry everything as one large parameter!

• Let's do this!


What?


I'LL TELL YOU WHAT

3> dbg:tp({X, some_method, 0}, [...]).** exception error: no case clause matching {test_module,horror}

4> dbg:tp({test_module, some_method, 0}, [...]).* never matches anything *

5> dbg:tp({test_module, some_method, 1}, [...]).* will actually match stuff *


What is this I don't even...


I'LL TELL YOU WHAT

• The parameters are global and they ghost values, become implicit function arguments. Errors out of nowhere.

• They're based on a old fun hack ({Mod, Fun}:(Args))

• They mess up the concept of arity

• They mess up tracing (previous slide!)

• They were added because some library accepts callback modules and nothing else but might still need state around. No other reason.


I'll use them anyway(in my server)


YOU PROBABLY SHOULDN'TBut you might not have a choice, so let's keep going


WHAT ELSE DO WE HAVE?

• A supervision tree, of course!

• Some of the supervisors are tricky and must interact with servers and dynamic children and ...

•Who starts who? By which API?

•We could probably just forget about supervisors, link stuff together and make it simpler

• Let's do this!


IN THIS VERY SLIDE I PRETEND I MAKE THE CHANGES


APP STRUCTURE


APP STRUCTUREtransport handlers

were moved and no longer need a supervisor.

They only use links.

The code is more straightforward!


APP STRUCTUREtransport handlers

were moved and no longer need a supervisor.

They only use links.

The code is more straightforward!


What?... my code is so much easier to read now!


I'LL TELL YOU WHAT

• You need to add ad-hoc start/restart policies and make sure orderly shutdowns work fine with just links. Care to add tests?

• You can no longer benefit from systool's application upgrades and downgrades as they depend on supervisors

• You will need to take your app down!


It's okay, I'll do rolling upgrades through all nodesthis should avoid downtime


NOT WITH LIVE SESSIONS!Nobody likes to be disconnected during a conversation!

Would you disconnect this guy?


NOT WITH LIVE SESSIONS!Nobody likes to be disconnected!

Stolen from Yurii's talk earlier today


NOT WITH LIVE SESSIONS!Nobody likes to be disconnected!

Stolen from Yurii's talk earlier today


You're killing me. I'll leave the change in. Code is meant to be

read first!


OK, AS LONG AS YOU LISTEN TO THE NEXT POINT


SURE, I'M HAPPY WITH MY APP

• It's all done in OTP

•Now Easier to read

• Uses parametrized modules, hell yes!

• It's time to start it!

• Let's do this!


THIS IS GONNA BE FUN

main(_) -> appmon:start(), sasl:start(normal, []), socketio:start(normal,[]), {ok, Pid} = socketio_listener:start([ {http_port, 7878}, {default_http_handler,?MODULE}]), EventMgr = socketio_listener:event_manager(Pid), ok = gen_event:add_handler(EventMgr, ?MODULE,[]), receive _ -> ok end.

And it works!(believe me)


THIS IS GONNA BE FUN

main(_) -> appmon:start(), sasl:start(normal, []), socketio:start(normal,[]), {ok, Pid} = socketio_listener:start([ {http_port, 7878}, {default_http_handler,?MODULE}]), EventMgr = socketio_listener:event_manager(Pid), ok = gen_event:add_handler(EventMgr, ?MODULE,[]), receive _ -> ok end.

And it works!(believe me)


YOU, AGAIN!?You are worse than clippy

What is it this time?


I'LL TELL YOU WHAT

• The VM starts all applications under the Application Controller

• Starting one outside of it means it is not supervised by it

•No app failure strategy (permanent, transient, temporary)

•No access to 'env' variables

•No respect of start phasesor dependencies


AND IT GETS WORSE•With bad default values (or overly defensive code) for env

variables, bad things happen...

• In this case, infinitely many web sockets opened for each client(before a patch to change that, because I forgot to listen to this hint)


DO THIS INSTEAD

main(_) -> appmon:start(), application:start(sasl), application:start(misultin), application:start(socketio), {ok, Pid} = socketio_listener:start([ {http_port, 7878}, {default_http_handler,?MODULE}]), EventMgr = socketio_listener:event_manager(Pid), ok = gen_event:add_handler(EventMgr, ?MODULE,[]), receive _ -> ok end.

And it works better!(believe me)


Oh...


RIGHTYour OTP code is not worth much without that


One more thing...


DON'T LISTEN TO MEIf your product might not ship because of this.Making it work is #1. Making it pretty is not.


HAPPY HACKINGHopefully you won't regret this


Technology

YOU WILL REGRET THIS