Le 07/01/2014 00:59, Tomasz Sterna a écrit :
Dnia 2014-01-06, pon o godzinie 18:07 +0100, Sylvain Guglielmi pisze:
Hello everyone,
To use jabberd2 with my pre-existing contacts database, I started
writing a roster module with customisable MySQL requests (I mailed this
list a while back about it, but I just started actual work). It uses
prepared statements, and config file looks like :
This is really interesting. :-D
It could really accelerate "XMPP as the next social media protocol". :-)
Could you consider using libdbi instead of hardwiring to MySQL?
This would make your implementation more reusable.
Well, the code is written and I need to go forward right now. But I'll
look into it when I get the MySQL version running in production. It
should not be hard provided libdbi works with prepared statements.
- In order to make simpler custom databases, I wanted to remove the
"pkt->type == pkt_S10N_UN" / "item->ask == 2" mechanism. According to a
comment "there is no ask='unsubscribe' in RFC bis anymore ". Would
anyone advise against it ?
How would that make schema simpler?
You still need to store ask='subscribe', don't you?
Yes, but I assumed it is simpler to explain in the config file "the 4th
parameter of the statement is a boolean meaning that the user is
requiring subscription" instead of "a int meaning either requiring a
subscription or an unsubscription". Especially if it's not a standard
feature. It also cleans a few lines of code. But you're right : the
benefit isn't worth bothering. I won't remove it.
- something that seems weird to me : I was expecting item->ver to be
incremented each time the item is updated (for example in
_roster_save_item),
but couldn't find such code. It there something I don't grasp ?
Ahhh... Yes... There's a little trick I used there. ;-)
ver is coming from auto-incrementing field "object-sequence".
As storage never issues UPDATE but always DELETE first then INSERT,
every updated roster item automagically gets new ver value.
Ok, I get it now. My custom requests however will use Update... So I'll
add the increment when the item is modified. As I understand it, the
"object-sequence" don't need to be an "UNIQUE" field. Am I right on that ?
[...] I thought it would be good to do the same
thing in the regular roster, but I couldn't find a way to remove a
specific os_object_t. Is there a way to do that ?
storage_delete() with proper filter.
Thanks for the insight.
Anyway here's what the code could look like, with TODOs :
The clean separation between stanza parsing to roster-item and
roster-group structures which are then stored to DB, and in reverse
makes the code more comprehensible.
I doubt that trading readability for efficiency is worthy in case of
today's microprocessor speeds and RAM availability.
Well, I have broken that clean separation early on. For example I needed
to check if the max number of contacts in the roster is reached with a
MySQL request, etc... So this part of readability is already lost in my
code :S.
Also, with many more requests, I fear the jabberd2 load will be more
important when I switch our prod to this plugin, and the load is already
noticeable (15% daily peak of our server)... It seemed to me a good
tradeoff, I'll stick to that right now. Anyway it'll be easy to revert
that later on and test the change in performance..
Also, does your use-case covers users in many-many groups?
My experience shows that most users are in 0 groups, some in 1 group and
very few in 2+ groups. There's not many INSERTs to gain there.
Remember: "Premature optimization is the root of all evil."
"...until your profiling shows 1000 functions each taking 0.1% of the
time, then you regret not having been careful when writing each of
them." ^_^ (I lack insight on what is costly for the SessionManager, but
in my experience mallocs and frees can be long, so I try to minimize
these calls as a rule. I assumed the same goes for MySQL requests).
I would rather suggest adding 'dirty' flag, risen whenever group
membership actually changes. This would keep the separation still,
allowing for bulk drops of unneeded DELETE/INSERTs.
Thanks for the feedback. I'll keep everyone updated.
--
Sylvain "Gugli" Guglielmi
