Re: deadlock problems with foreign keys

Mario Weilguni wrote:
> I've a severe problem with deadlocks in postgres, when using referential integrity it's quite easy to trigger deadlocks. I think the may be a bug in ri_trigger.c (discussed later). Here's some short example:
>
> create table languages (
> id integer not null,
> name text not null,
> primary key(id)
> );
>
> create table entry (
> id integer not null,
> lang_id integer,
> sometext text,
> primary key (id),
> foreign key ( lang_id ) references languages (id)
> );
>
> insert into languages values (1, 'english');
> insert into languages values (2, 'german');
>
> insert into entry values (1, 1, 'text 1');
> insert into entry values (2, 1, 'text 2');
>
>
> transaction A: begin;
> transaction A: update entry set sometext='text 1.1' where id=1;
> transaction A: .... do more time-consuming processing here...
> meanwhile, B: begin;
> B: update entry set sometext='text 2.1' where id=2;
>
> -- both processes hang now

Cannot reproduce that problem in v7.2. Only B blocks until A
either commits or rolls back. So what exactly is your "more
time-consuming processing"?

>
> I think this is too much locking here, because the logfile show's something like this:
> 'select 1 from "languages" where id=$1 for update' (2 times).
>
> Now I've a lot of tables (around 30) and use referential integrity a lot on ~10 columns (language, country....) , and with more fields it's very easy to deadlock the whole system (it happens a lot in my web applicaiton with ~20 concorrent users).
>
> IMHO the "select ... for update" on languages is not necessary, since I do not want to update "lang_id", but I might be wrong. The other problem is, that this will make postgres in benchmarks very slow (with many concurrent connections), at least if the application is not trivial.
>
> IMO the problem is in ri_trigger.c around line 390:
> /* ----------
> * The query string built is
> * SELECT 1 FROM ONLY <pktable> WHERE pkatt1 = $1 [AND ...]
> * The type id's for the $ parameters are those of the
> * corresponding FK attributes. Thus, SPI_prepare could
> * eventually fail if the parser cannot identify some way
> * how to compare these two types by '='.
> * ----------
> */
>
> Any ideas if this is a bug or simply strict SQL standard?

It does a SELECT ... FOR UPDATE because we don't have a
SELECT ... AND PLEASE DO NOT REMOVE.

If we would only check if the PK is there now, another
concurrent transaction could delete the PK, it's own check
cannot see our uncommitted row yet and we end up with a
violation. And if you look at the comment a few lines up, it
explains why we cannot skip the check even if the key value
doesn't change.

Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#================================================== JanWieck(at)Yahoo(dot)com #

_________________________________________________________
Do You Yahoo!?
Get your free @yahoo.com address at http://mail.yahoo.com