Re: Regex matching where text is input and regex stored in column

Στις 15/9/23 13:52, ο/η Achilleas Mantzios - cloud έγραψε:
>
> Hello
>
> I have a table like :
>
> amantzio(at)sma/dynacom=# \d mail_vessel_addressbook
>                                    Table "public.mail_vessel_addressbook"
> Column  |  Type   | Collation | Nullable |
>                              Default
> ---------+---------+-----------+----------+--------------------------------------------------------------------
>
> id      | integer |           | not null |
> nextval(('public.mail_vessel_addressbook_id_seq'::text)::regclass)
> name    | text    |           | not null |
> address | text    |           | not null |
> vslid   | integer |           | not null |
> Indexes:
>
> .....
>
> where address stores basically email patterns, human / edited text
> with possible wild cards, meant to be input mainly by our admins.
> There is an query that checks a particular new address (this runs via
> an exim4 mail server) which has become quite slow :
>
> SELECT text('news3(at)newsaccess(dot)in') from mail_global_addressbook where
> text('news3(at)newsaccess(dot)in') ~*
> replace(replace(replace(address,'*','.*'),'+','.+'),'?','.?')  UNION SELE
> CT text('news3(at)newsaccess(dot)in') from mail_vessel_addressbook where
> text('news3(at)newsaccess(dot)in') ~*
> replace(replace(replace(address,'*','.*'),'+','.+'),'?','.?')  LIMIT 1;
>
> It basically seq scans both tables, and the 2nd table in particular is
> fairly big : 123966 rows. So the above query tends to take 3.6 secs on
> average to execute. A pg_trgm index would not help since there is no
> way (that I know of) to index a regex in pgsql. Most common use is
> about indexing values to be checked against a given regex, quite the
> reverse from what we need here.
>
> I got great improvement by materializing
> replace(replace(replace(address,'*','.*'),'+','.+'),'?','.?'), we are
> running pgsql 10.23 :( , no generated columns possible, so I used a
> trigger, and then adding a normal b-tree index on the column. This
> causes Index only scan which is much better than the seq scan. Also I
> had to re-write the query as :
>
> with
>
> qry1 as
>
>     (SELECT text('watchkeepers(at)ukmto(dot)org') from
> mail_vessel_addressbook where
>
>     text('watchkeepers(at)ukmto(dot)org') ~*
> replace(replace(replace(address,'*','.*'),'+','.+'),'?','.?')
>
>     LIMIT 1),
> qry2 as
>
>     (SELECT text('watchkeepers(at)ukmto(dot)org') from
> mail_global_addressbook where
>
>      text('watchkeepers(at)ukmto(dot)org') ~*
> replace(replace(replace(address,'*','.*'),'+','.+'),'?','.?')
>
>     LIMIT 1)
> select text('watchkeepers(at)ukmto(dot)org') from qry1 UNION select
> text('watchkeepers(at)ukmto(dot)org') from qry2 LIMIT 1;
>
> I was wondering if there is some more elegant and better solution,
> without the need for extra column and trigger.
>

Answering this myself, the way I dealt with it was just create an extra
normal index (btree) on the domain part, and include an extra equality
check for the domain. This boosted performance from 2.5 secs to less
than 4ms.

--
Achilleas Mantzios
IT DEV - HEAD
IT DEPT
Dynacom Tankers Mgmt