[erlang-questions] A couple of design questions

Steve Davis steven.charles.davis@REDACTED
Fri Mar 27 11:27:26 CET 2009


Mats,

Thanks for the assistance and guidance.

re: 1) process dictionary - I've done this refactor and it has improved 
the code clarity a great deal. My only rule with this was to make the 
process dictionary entries "write only" as advised by Mr Armstrong in 
his excellent book. The code looks a good deal more sane, and indeed is 
considerably safer because it's not using the public ets that was 
apparently necessary before.

re: 2) gen_server - I did notice that I had already begun implementing 
some remote calls for the more experimental features, so your answer 
here rings very, very true. I'm doing the gen_server refactor now rather 
than later.

Thanks again :)

Regards,
Steve


mats cronqvist wrote:
> Steve Davis <steven.charles.davis@REDACTED> writes:
> 
>> Hi all,
>>
>> I am working on a gs-style interface to wx, and was hoping someone may
>> have an opinion on a couple of design issues:
>>
>> 1) I've been maintaining an ets table to cope with naming/references
>> and command events. In order to keep the references unique across
>> multiple UI instances (processes), I've used a key format that is
>> basically {self(), name}/{reference, data}. I've had problems
>> convincing myself that these tables are safe as they must be marked
>> "public". However, I'm now thinking that these references would be
>> better off in the process dictionary (something I am usually cautious
>> of doing) as these stored references are, or should be (a) private to
>> the process, (b) "transient" with the process, and (c) irrelevant
>> outside the process. So my question here is: in this case, is it as
>> valid as it seems to me now to the use of the process dictionary to
>> store a fair amount of data, and if so what limitations should I watch
>> out for?
> 
>   Oldtimers avoid the process dict because it used to be slow. "Real"
>   functional programmers avoid both ets and the dict for religious
>   reasons.  AFAIK, there's no real reason to avoid the process dict in
>   favor of ets.
> 
>> 2) Starting the UI instance process merely uses spawn (and returns the
>> pid to the calling process). It's troubled me that perhaps the GUI
>> process should be more "OTP". However, given that this is a framework
>> that builds on wxe (with the graphics server already there), my
>> thinking is that the ownership (and supervision) of the UI process
>> would be better implemented by the calling application, since it can
>> link and supervise the UI. Making the UI run off a gen_server seems an
>> unnecessary complexity in the design of the framework. My question
>> here is: Simple though this is, is my thinking correct?
> 
>   In my experience, it pays off to use gen_server right from the
>   start. Otherwise, once the state machine starts to grow, you'll
>   either refactor to a gen_server or wish you had.
>  
>   gen_server has no noticeable performance penalty, but you do need to
>   write a bunch of boilerplate. You might want to check out this(*)
>   code for an example of how to abstract away the boilerplate.
> 
>   mats
> 
> (*) gen_serv implements a gen_server callback. It will delegate to its
> own callback if possible (it's some sort of inheritance). E.g. given
> this module;
> 
> -module('foo').
> -author('Mats Cronqvist').
> -export([handle_call/3]).
> 
> handle_call(Msg,From,State) -> 
>   io:fwrite("~p~n",[{Msg,From,State}]),
>   {ok,State}.
> 
>   you can do this;
>  1> gen_serv:start(foo).
>  2> gen_server:call(foo,bla).
> {bla,{<0.28519.2>,#Ref<0.0.12.74776>},undefined}
> 
>  http://code.google.com/p/eper/source/browse/trunk/src/gen_serv.erl
> 
> 




More information about the erlang-questions mailing list