Hi Dan,
On Wed, 10 Aug 2005, Dan Harris wrote:
> I have a web page for my customers that shows them count of records
> and some min/max date ranges in each table of a database, as this is
> how we bill them for service. They can log in and check the counts
> at any time. I'd like for the counts to be as fresh as possible by
> keeping this dynamic, but I will use a periodic 'snapshot'/cron job
> if that is the only option to speed this up. I have thought about
> using the table statistics, but the estimate error is probably
> unacceptable because of the billing purposes.
>
> For some reason, the SQL Server we migrated the app from can return
> count(*) in a split second on multi-million row tables, even though
> it is a MUCH slower box hardware-wise, but it's now taking many
> seconds to run. I have read in the archives the problems MVCC brings
> into the count(*) dilemma forcing Pg to run a seq scan to get
> counts. Does SQLServer not use MVCC or have they found another
SQL Server probably jumps through a lot of hoops to do fast count(*)s. I'm
sure we could do something similar -- it's just a question of complexity,
resources, desirability, etc. The are other solutions, which makes the
idea of doing it less attractive still.
> approach for arriving at this number? Compounding all the min/max
> and counts from other tables and all those queries take about a
> minute to run. The tables will contain anywhere from 1 million to 40
> million rows.
>
> Also, I am using "select ... group by ... order by .. limit 1" to get
> the min/max since I have already been bit by the issue of min() max()
> being slower.
I generally pre generate the results. There are two ways to do this: the
'snapshot'/cronjon you mentioned or using rules and triggers to maintain
'count' tables. The idea is that if data is added, modified or removed
from your table, you modify counters in these other tables.
Alternatively, feel free to post your schema and sample queries with
explain analyze results to this list. Alternatively, jump on irc at
irc.freenode.net #postgresql and someone will be more than happy to look
through the problem in more detail.
Thanks,
Gavin
