Re: strange sql behavior

Thanks Bill.

Also, it's very important to include the headers with the queries!

Marco,

There is no top secret information that is requested, so please do not edit the output.

Sent via the Samsung Galaxy S® 6, an AT&T 4G LTE smartphone

-------- Original message --------

From: Bill Moran <wmoran@potentialtech.com>

Date: 2/1/2016 14:41 (GMT-05:00)

To: Yu Nie <nieyu04@gmail.com>

Cc: Melvin Davidson <melvin6925@gmail.com>, pgsql-general@postgresql.org

Subject: Re: [GENERAL] strange sql behavior

Came a little late to the thread, see many comments inline below:

On Mon, 1 Feb 2016 13:16:13 -0600
Yu Nie <nieyu04@gmail.com> wrote:

> Thanks  a lot for your reply. I ran the query you suggested and here are
> the results
>
> Large table: "public";"data2011_01";784236885;0;0;"";"";"";"2016-01-19
> 17:31:08-06";156847423
> Small table:  "public";"data2013_01w";300786444;0;0;"";"";"2016-02-01
> 08:57:24-06";"2016-02-01 04:01:04-06";60157336

You didn't do Melvin's query correctly. He specified 11 columns to
select, but you only show 10. Since you don't show the query you
actually ran, we don't know which of the numeric columns is missing.

More information inline below:

>
> On Mon, Feb 1, 2016 at 1:00 PM, Melvin Davidson <melvin6925@gmail.com>
> wrote:
>
> > One thing to look at is the last time both tables were vacuumed/analyzed.
> >
> > SELECT n.nspname,
> >        s.relname,
> >        c.reltuples::bigint,
> >        n_tup_ins,
> >        n_tup_upd,
> >        n_tup_del,
> >        date_trunc('second', last_vacuum) as last_vacuum,
> >        date_trunc('second', last_autovacuum) as last_autovacuum,
> >        date_trunc('second', last_analyze) as last_analyze,
> >        date_trunc('second', last_autoanalyze) as last_autoanalyze
> >        ,
> >        round( current_setting('autovacuum_vacuum_threshold')::integer +
> > current_setting('autovacuum_vacuum_scale_factor')::numeric * C.reltuples)
> > AS av_threshold
> >   FROM pg_stat_all_tables s
> >   JOIN pg_class c ON c.oid = s.relid
> >   JOIN pg_namespace n ON (n.oid = c.relnamespace)
> >  WHERE s.relname NOT LIKE 'pg_%'
> >    AND s.relname NOT LIKE 'sql_%'
> >    AND s.relname IN  ' "your_small_table", "your_large_table"'
> >  ORDER by 1, 2;
> >
> >
> > Also, please confirm the indexes for both tables are using the same method
> > (btree?).
> >
> > On Mon, Feb 1, 2016 at 1:35 PM, Yu Nie <nieyu04@gmail.com> wrote:
> >
> >> Hi there,
> >>
> >> Recently I am working with a large amount of taxis GIS data and had
> >> encountered some weird performance issues.  I am hoping someone in this
> >> community can help me figure it out.
> >>
> >> The taxi data were loaded in 5 minute block into a table.  I have two
> >> separate such tables, one stores a month of data with about 700 million
> >> rows, another stores about 10 days of data with about 300 million rows.
> >> The two tables have the exactly same schema and indexes. There are two
> >> indexes: one on taxiid (text), and the other on the time stamp (date
> >> time).  In order to process the data, I need to get all points for a single
> >> taxis; to do that, I use something like:
> >>  select * from table1 where taxiid = 'SZB00S41' order by time;
> >> What puzzled me greatly is that this query runs consistently much faster
> >> for the large table than for the small table, which seems to contradict
> >> with intuition.   At the end of message you may find explain (analyze
> >> buffer) results of two particular queries for the same taxiid (one for each
> >> table). You can see that it took much longer (more than 20 times) to get
> >> 20k rows from the small table than to get 44 k rows from the large table.
> >> Interestingly it seems that the planner does expect about 1/3 work for the
> >> small table query - yet for some reason, it took much longer to fetch the
> >> rows from the small table.   Why there is such a huge performance between
> >> the two seemingly identical queries executed on two different tables?
> >>
> >> Is is because the data on the second table is on some mysteriously
> >> "broken part" of the disk?  what else could explain such a bizarre
> >> behavior?  Your help is greatly appreciated.
> >>
> >> The above behavior is consistent through all queries.   Another issue I
> >> identified is that for the large table, the query can use the shared buffer
> >> more effectively.  For example, after you query one taxiid and immediately
> >> following that query run the same query for another taxi whose id ranks
> >> right behind the first id, then shared hit buffers would be quite high (and
> >> the query would run much faster); this however never works for the small
> >> table.
> >>
> >> Thanks   a lot!
> >>
> >> Best, Marco
> >>
> >>
> >> Results for the small table: it took 141 seconds to finish.  The planning
> >> time is 85256.31
> >>
> >> "Sort  (cost=85201.05..85256.31 rows=22101 width=55) (actual
> >> time=141419.499..141420.025 rows=20288 loops=1)"
> >> "  Sort Key: "time""
> >> "  Sort Method: quicksort  Memory: 3622kB"
> >> "  Buffers: shared hit=92 read=19816"
> >> "  ->  Bitmap Heap Scan on data2013_01w  (cost=515.86..83606.27
> >> rows=22101 width=55) (actual time=50.762..141374.777 rows=20288 loops=1)"
> >> "        Recheck Cond: ((taxiid)::text = 'SZB00S41'::text)"
> >> "        Heap Blocks: exact=19826"
> >> "        Buffers: shared hit=92 read=19816"

^^
Note that despite this table being smaller, Postgres had to read 19816
blocks from disk, which is 2.5x more than the larger table.

> >> "        ->  Bitmap Index Scan on data2013_01w_ixtaxiid
> >> (cost=0.00..510.33 rows=22101 width=0) (actual time=26.053..26.053
> >> rows=20288 loops=1)"
> >> "              Index Cond: ((taxiid)::text = 'SZB00S41'::text)"
> >> "              Buffers: shared hit=4 read=78"
> >> "Planning time: 0.144 ms"
> >> "Execution time: 141421.154 ms"
> >>
> >> Results for the large table: it took 5 seconds to finish.  The planning
> >> time is 252077.10
> >> "Sort  (cost=251913.32..252077.10 rows=65512 width=55) (actual
> >> time=5038.571..5039.765 rows=44204 loops=1)"
> >> "  Sort Key: "time""
> >> "  Sort Method: quicksort  Memory: 7753kB"
> >> "  Buffers: shared hit=2 read=7543"
> >> "  ->  Bitmap Heap Scan on data2011_01  (cost=1520.29..246672.53
> >> rows=65512 width=55) (actual time=36.935..5017.463 rows=44204 loops=1)"
> >> "        Recheck Cond: ((taxiid)::text = 'SZB00S41'::text)"
> >> "        Heap Blocks: exact=7372"
> >> "        Buffers: shared hit=2 read=7543"
> >> "        ->  Bitmap Index Scan on data2011_01_ixtaxiid
> >> (cost=0.00..1503.92 rows=65512 width=0) (actual time=35.792..35.792
> >> rows=44204 loops=1)"

^^
Note that the larger table took LONGER to do the index work than the
smaller table, which probably means there's nothing wrong with your
disks or anything ... that's the behavior I would expect.

> >> "              Index Cond: ((taxiid)::text = 'SZB00S41'::text)"
> >> "              Buffers: shared hit=2 read=171"
> >> "Planning time: 0.127 ms"
> >> "Execution time: 5042.134 ms"

So, what I'm seeing, is that Postgres is able to figure out _which_ rows
to fetch faster on the small table than the large table, which is what
you would expect, since a smaller index should be faster than a large one.

However, when it goes to actually fetch the row data, it takes
significantly longer on the small table, despite the fact that it's
only fetching 1/3 as many rows. It is, however, doing 2.5x as many
disk reads to get those rows: For the large table, it reads 61MB from
disk, but it reads 160MB to get all the data for the smaller table.

How the data was inserted into each table could lead to similar data
being clustered on common pages on the large table, while it's spread
across many more pages on the small table.

That still doesn't explain it all, though. 2.5x the disk
activity normally wouldn't equate to 28x the time required. Unless
you're disks are horrifically slow? Does this server have a lot of
other activity against the disks? I.e. are other people running
queries that you would have to contend with, or is the server a VM
sharing storage with other VMs, or even a combined use server that
has to share disk access with (for example) a web or mail server
as well? Is the performance difference consistently ~28x?

Other things: what is shared_buffers set to? The queries would seem
to indicate that this server has less than 1M of those two tables
cached in memory at the time you ran those queries, which seems to
suggest that either you've got shared_buffers set very low, or that
there are a lot of other tables that other queries are accessing at
the time you're running these. Perhaps installing pg_buffercache to
have a look at what's using your shared_buffers would be helpful.

--
Bill Moran