Обсуждение: Trouble managing planner for timestamptz columns

We have large tables that hold statistics based on time. They are of the
form.

CREATE TABLE stats (
    id serial primary key,
    logtime timestamptz,
    d1 int,
    s1 bigint
);

CREATE INDEX idx on stats(logtime);

Some of these tables have new data inserted at a rate of 500,000+ rows /
hour.  The entire table will grow to being 10's to 100's of millions of
rows in size.  (Yes, we are also paritioning these, it's the size of an
individual partition that we're talking about).

We tend to analyze these tables every day or so and this doesn't always
prove to be sufficient....

Our application is a reporting application and the end users typically
like to query the newest data the most. As such,  the queries of the
form...


select
 *
from stats
inner join dimension_d1 using (d1)
where logtime between X and Y and d1.something = value;

This usually results in a hash join (good thing) where the dimension
table is loaded into the hash table and it index scans stats using idx
index.

The trouble starts when both X and Y are times "after" the last analyze.
This restriction clause is outside the range of values in the historgram
created by the last analyze.  Postgres's estimate on the number of rows
returned here is usually very low and incorrect, as you'd expect...

Trouble can occur when the planner will "flip" its decision and decide
to hash join by loading the results of the index scan on idx into the
hash table instead of the dimension table....

Since the table is so large and the system is busy (disk not idle at
all), doing an analyze on this table in the production system can take
1/2 hour!  (statistics collector set to 100).  We can't "afford" to
analyze more often...

It certainly would be nice if postgres could understand somehow that
some columns are "dynamic" and that it's histogram could be stretched to
the maximal values or some other technique for estimating rows to the
right of the range of values in the histogram...

Or have some concept of error bars on it's planner decisions....

Suggestions? Comments?


Marc

"Marc Morin" <marc@sandvine.com> writes:
> We tend to analyze these tables every day or so and this doesn't always
> prove to be sufficient....

Seems to me you just stated your problem.  Instead of having the planner
make wild extrapolations, why not set up a cron job to analyze these
tables more often?  Or use autovacuum which will do it for you.

> Since the table is so large and the system is busy (disk not idle at
> all), doing an analyze on this table in the production system can take
> 1/2 hour!  (statistics collector set to 100).

I'd believe that for vacuum analyze, but analyze alone should be cheap.
Have you perhaps got some weird datatypes in the table?  Maybe you
should back off the stats target a bit?

We do support analyzing selected columns, so you might try something
like a cron job analyzing only the timestamp column, with a suitably low
stats target for that column.  This would yield numbers far more
reliable than any extrapolation the planner could do.

            regards, tom lane

Well this analyze just took 12 minutes...  Stats target of 100.

# time psql xxx xxx -c "analyze elem_trafficstats_1"
ANALYZE

real    12m1.070s
user    0m0.001s
sys     0m0.015s


A large table, but by far, not the largest...  Have about 1 dozen or so
tables like this, so analyzing them will take 3-4 hours of time...  No
weird datatypes, just bigints for facts, timestamptz and ints for
dimensions.

My problem is not the analyze itself, it's the fact that our db is
really busy doing stuff.... Analyze I/O is competing... I am random I/O
bound like crazy.

If I set the stats target to 10, I get

# time psql xxxx xxx -c "set session default_statistics_target to
10;analyze elem_trafficstats_1"
ANALYZE

real    2m15.733s
user    0m0.009s
sys     0m2.255s

Better, but not sure what side affect this would have.

> -----Original Message-----
> From: Tom Lane [mailto:tgl@sss.pgh.pa.us]
> Sent: Friday, March 10, 2006 1:31 PM
> To: Marc Morin
> Cc: pgsql-performance@postgresql.org
> Subject: Re: [PERFORM] Trouble managing planner for
> timestamptz columns
>
> "Marc Morin" <marc@sandvine.com> writes:
> > We tend to analyze these tables every day or so and this doesn't
> > always prove to be sufficient....
>
> Seems to me you just stated your problem.  Instead of having
> the planner make wild extrapolations, why not set up a cron
> job to analyze these tables more often?  Or use autovacuum
> which will do it for you.
>
> > Since the table is so large and the system is busy (disk
> not idle at
> > all), doing an analyze on this table in the production
> system can take
> > 1/2 hour!  (statistics collector set to 100).
>
> I'd believe that for vacuum analyze, but analyze alone should
> be cheap.
> Have you perhaps got some weird datatypes in the table?
> Maybe you should back off the stats target a bit?
>
> We do support analyzing selected columns, so you might try
> something like a cron job analyzing only the timestamp
> column, with a suitably low stats target for that column.
> This would yield numbers far more reliable than any
> extrapolation the planner could do.
>
>             regards, tom lane
>

"Marc Morin" <marc@sandvine.com> writes:
> Well this analyze just took 12 minutes...  Stats target of 100.
> # time psql xxx xxx -c "analyze elem_trafficstats_1"

Try analyzing just the one column, and try reducing its stats target to
10.  It does make a difference:

sorttest=# set default_statistics_target TO 100;
SET
Time: 0.382 ms
sorttest=# analyze verbose d10;
INFO:  analyzing "public.d10"
INFO:  "d10": scanned 30000 of 833334 pages, containing 3600000 live rows and 0 dead rows; 30000 rows in sample,
100000080estimated total rows 
ANALYZE
Time: 137186.347 ms
sorttest=# set default_statistics_target TO 10;
SET
Time: 0.418 ms
sorttest=# analyze verbose d10(col1);
INFO:  analyzing "public.d10"
INFO:  "d10": scanned 3000 of 833334 pages, containing 360000 live rows and 0 dead rows; 3000 rows in sample, 100000080
estimatedtotal rows 
ANALYZE
Time: 17206.018 ms
sorttest=#

            regards, tom lane