Re: Simulating sequences - Mailing list pgsql-general

Vilson farias wrote:

> Greetings,
>
>   I'm getting a big performance problem and I would like to ask you what
> would be the reason, but first I need to explain how it happens.
>
>   Let's suppose I can't use sequences (it seams impossible but my boss
> doesn't like specific database features like this one).

I don't see how using PL/pgSQL is any better with respect to specific
database feature, but that aside for a moment.

Your function not only misses the required FOR UPDATE when reading the
(possibly existing) current value, it also contains a general race
condition. Multiple concurrent transactions could try inserting the new
key and every but one would error out with a duplicate key error.

Sequence values are int8 and are by default safe against integer rollover.

Sequences do not rollback and therefore don't need to wait for
concurrent transactions to finish. Your table based replacement is a
major concurrency bottleneck. As soon as a transaction did an insert to
a table, it blocks out every other transaction from inserting into that
table until it either commits or rolls back.

Your VACUUM theory is only partial correct. A frequent VACUUM will
prevent the key table from growing. You'd have to do so very often since
the excess number of obsolete index entries pointing to dead tuples also
degrades your performance. Additionally if there is a very low number of
keys (sequences) in that table, an ANALYZE run might cause the planner
to go for a sequential scan and ignore the index on the table at which
point your function will actually cause "two" sequential scan over all
live and dead tuples of all sequences per call.

Sequences are specially designed to overcome all these issues.

If you cannot convice your boss to use sequences, he is a good example
for why people having difficulties understanding technical issues should
not assume leadership positions in IT projects.


Jan

>
>   For sequence simulation I had created a table called cnfg_key_generation
> and each tuple holds information for one of my tables (tablename,
> current_sequencial_number). Lets check :
>
> CREATE TABLE cnfg_key_generation (
>        department  integer     NOT NULL,
>        table_name  varchar(20) NOT NULL,
>        current_key integer NOT NULL,
>        CONSTRAINT XPKcnfg_key_generation PRIMARY KEY (department,
> table_name)
> );
>
>
> Per example, for a table called 'my_test' I would have the following values
> :
>   department  = 1
>   table_name  = 'my_test'
>   current_key = 1432
>
> Everytime I want a new key to use in my_test primary-key I just increment
> current_key value. For this job, I've created a simple stored procedure
> called key_generation
>
>
> CREATE FUNCTION key_generation(integer, varchar(20)) RETURNS integer AS'
> DECLARE
>   the_department ALIAS FOR $1;
>   the_table_name ALIAS FOR $2;
>   new_key_value  integer;
>   err_num        integer;
> BEGIN
>   new_value := 0;
>
>   LOCK TABLE cnfg_key_generation IN ROW EXCLUSIVE MODE;
>
>   SELECT current_value INTO new_value
>     FROM cnfg_key_generation
>     WHERE the_department = department AND the_table_name = table_name;
>
>   IF NOT FOUND THEN
>     new_key_value := 1;
>     INSERT INTO cnfg_key_generation VALUES (the_department,the_table_name,
> new_key_value);
>   ELSE
>     new_key_value := new_key_value + 1;
>
>     UPDATE cnfg_key_generation
>       SET current_key_value = new_key_value
>       WHERE department = the_department AND
>             table_name = the_table_name;
>   END IF;
>
>   RETURN new_key_value;
>
> END;
> '
> LANGUAGE 'plpgsql';
>
>
> Data insertion is done by the following way :
>
> INSERT INTO my_test VALUES (key_generation(1, 'my_test'), ...other
> fields...);
>
> Ok, this works, but has a huge performance problem. I read in pg docs that
> everytime I make an UPDATE, a new tuple is created and the old is marked as
> invalid. For a tuple that holds sequencial key generation, everytime a new
> key is generated, a new tuple is created inside cfg_key_generation. It means
> after 2million key generations for same table, performance will be completly
> degradated because there will be 2million of old versions of same tuple. For
> instance, I have a table called 'cham_chamada' that actually holds
> 1.5Million of tuples. The response time for key_generation execution for
> this table is more than 5seconds. In this same case if I execute
> key_generation for a table that has just few values (current key = 5 per
> example), response time is just some miliseconds (30 to 50ms).
>
> I tryied to fix this problem with a VACUUM and it was completly ineffective.
> After execution the problem was still there. Later, after execution of every
> kind of vacuum I knew (with and without ANALYZE, especific for that table,
> vacuumdb from shell ...) I gave up and tryied a COPY TO/DROP/CREATE/COPY
> FROM. At this time, the performance problem was fixed.
>
> What can I do to solve this problem without table reconstruction? Is there a
> performance degradation in PostgreSQL that can't be solved? If a have a huge
> table with millions of data being update constantly, what can I do to keep a
> good performance if vacuum isn't 100%?
>
> Does PostgreSQL sequences deal with these kinds performance questions? Is it
> clean, fast and secury to use sequences? Maybe I still can make up my boss
> mind about this subject if I get good arguments here to justify the use of
> sequences.
>
> Am I doing some stupid thing?
>
> Best regards,
>
> -----------------------------------------------------------------
> José Vilson de Mello de Farias
> Software Engineer
>
> Dígitro Tecnologia Ltda - www.digitro.com.br
> APC - Customer Oriented Applications
> E-mail: vilson.farias@digitro.com.br
> Tel.: +55 48 281 7158
> ICQ 11866179
>
>
> ---------------------------(end of broadcast)---------------------------
> TIP 8: explain analyze is your friend


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