Thread: Help with slow table update

On 4/9/15 6:18 PM, Pawel Veselov wrote:

Hi.

I have a plpgsql procedure that updates a few similar tables.
for some reason, updates on one of the tables take a lot longer the
updates on the other ones. The difference is, say, 7 seconds vs. 80
milliseconds.

the procedure uses cursors and record variables to do the updates. For
example:

         update r_agrio_total set
           unserved = unserved + (agrow->>'unserved')::numeric(38),
           r_brkconn = mush_brk_conn(r_brkconn, q_item.r_brkconn),
         where
           tagid = _tagid and
           unitid = (akey->>'unitid')::numeric and
           placement = (akey->>'placement')::numeric and
           device_type = (akey->>'device_type')::numeric;

There is another table (xq_agr) that is read record by record, and for
each of those records, such update is executed.

I was trying to select analyze the updates to see where the time could
be spent.
There are only 24 row in the "bad" table, and 3,400 rows in "good"
table. So, for the "bad" table, most of the updates will be on the same
rows. The times were measured on processing 100 original records.

When I'm analyzing pure update statements, I don't see anything strange.

"bad" table: explain analyze update r_agrio_total set unconfirmed =
unconfirmed +0 where tagid = 1000 and unitid = 1000 and placement = 0
and device_type = 100;

RESULT:
  Update on r_agrio_total  (cost=0.42..4.46 rows=1 width=321) (actual
time=0.253..0.253 rows=0 loops=1)
    ->  Index Scan using tag_r_agrio_total on r_agrio_total
  (cost=0.42..4.46 rows=1 width=321) (actual time=0.037..0.041 rows=1
loops=1)
          Index Cond: (tagid = 1000::numeric)
          Filter: ((unitid = 1000::numeric) AND (placement = 0::numeric)
AND (device_type = 100::numeric))
          Rows Removed by Filter: 7
  Total runtime: 0.282 ms

"good" table: explain analyze update r_agrio_hourly set unconfirmed =
unconfirmed +0 where tagid = 1000 and unitid = 1000 and placement = 0
and device_type = 100 and rowdate = '2015-02-23T13';

RESULT:
  Update on r_agrio_hourly  (cost=0.42..17.36 rows=6 width=329) (actual
time=0.102..0.102 rows=0 loops=1)
    ->  Index Scan using u_r_agrio_hourly on r_agrio_hourly
  (cost=0.42..17.36 rows=6 width=329) (actual time=0.047..0.048 rows=1
loops=1)
          Index Cond: ((tagid = 1000::numeric) AND (unitid =
1000::numeric) AND ((rowdate)::text = '2015-02-23T13'::text) AND
(device_type = 100::numeric) AND (placement = 0::numeric))
  Total runtime: 0.135 ms

When I try doing it with WITH statement (really, to apply the actual
data that the plpgsql function uses), there is something strange in the
"bad" table.

explain analyze
with SRC as (select * from xq_agr where id = 914830)
         update r_agrio_total set
           unconfirmed = unconfirmed +
(SRC.r_agrio->>'unconfirmed')::numeric(38)
         from SRC
         where
           tagid = (SRC.r_agrio->'key'->>'tagid')::numeric and
           unitid = (SRC.r_agrio->'key'->>'unit')::numeric and
           placement = (SRC.r_agrio->'key'->>'placement')::numeric and
           device_type = (SRC.r_agrio->'key'->>'device_type')::numeric;

RESULT:
  Update on r_agrio_total  (cost=8.91..32777.51 rows=19331 width=409)
(actual time=0.107..0.107 rows=0 loops=1)
    CTE src
      ->  Index Scan using xq_agr_pkey on xq_agr  (cost=0.42..8.44
rows=1 width=379) (actual time=0.026..0.027 rows=1 loops=1)
            Index Cond: (id = 914830)
    ->  Nested Loop  (cost=0.46..32769.07 rows=19331 width=409) (actual
time=0.107..0.107 rows=0 loops=1)
          ->  CTE Scan on src  (cost=0.00..0.02 rows=1 width=88) (actual
time=0.032..0.033 rows=1 loops=1)
          ->  Index Scan using u_r_agrio_total on r_agrio_total

  (*cost=0.46..32285.78 rows=19331* width=321) (actual time=0.001..0.001

rows=0 loops=1)
                Index Cond: ((tagid = (((src.r_agrio -> 'key'::text) ->>
'tagid'::text))::numeric) AND (unitid = (((src.r_agrio -> 'key'::text)
->> 'unit'::text))::numeric) AND (device_type = (((src.r_agrio ->
'key'::text) ->> 'device_type'::text))::numeric) AND (placement =
(((src.r_agrio -> 'key'::text) ->> 'placement'::text))::numeric))
  Total runtime: 0.155 ms

explain analyze
with SRC as (select * from xq_agr where id = 914830)
         update r_agrio_hourly set
           unconfirmed = unconfirmed +
(SRC.r_agrio->>'unconfirmed')::numeric(38)
         from SRC
         where
           tagid = (SRC.r_agrio->'key'->>'tagid')::numeric and
           unitid = (SRC.r_agrio->'key'->>'unit')::numeric and
           placement = (SRC.r_agrio->'key'->>'placement')::numeric and
           device_type = (SRC.r_agrio->'key'->>'device_type')::numeric and
           rowdate = (SRC.r_agrio->'key'->>'rowdate');

RESULT:
  Update on r_agrio_hourly  (cost=8.91..52.91 rows=20 width=417) (actual
time=0.123..0.123 rows=0 loops=1)
    CTE src
      ->  Index Scan using xq_agr_pkey on xq_agr  (cost=0.42..8.44
rows=1 width=379) (actual time=0.023..0.024 rows=1 loops=1)
            Index Cond: (id = 914830)
    ->  Nested Loop  (cost=0.47..44.47 rows=20 width=417) (actual
time=0.121..0.121 rows=0 loops=1)
          ->  CTE Scan on src  (cost=0.00..0.02 rows=1 width=88) (actual
time=0.030..0.031 rows=1 loops=1)
          ->  Index Scan using u_r_agrio_hourly on r_agrio_hourly

  (*cost=0.47..43.95 rows=20* width=329) (actual time=0.000..0.000

rows=0 loops=1)
                Index Cond: ((tagid = (((src.r_agrio -> 'key'::text) ->>
'tagid'::text))::numeric) AND (unitid = (((src.r_agrio -> 'key'::text)
->> 'unit'::text))::numeric) AND ((rowdate)::text = ((src.r_agrio ->
'key'::text) ->> 'rowdate'::text)) AND (device_type = (((src.r_agrio ->
'key'::text) ->> 'device_type'::text))::numeric) AND (placement =
(((src.r_agrio -> 'key'::text) ->> 'placement'::text))::numeric))
  Total runtime: 0.176 ms

I've tried doing vacuum full analyze, rebuilding indexes. Vacuum did
help somewhat, but the time differential is still huge (that 7sec vs. 80ms)

The table structure (some fields thrown out just to save output length):

=> \d r_agrio_total
       Column       |     Type      |                         Modifiers
-------------------+---------------+------------------------------------------------------------
  id                | bigint        | not null default
nextval('r_agrio_total_id_seq'::regclass)
  tagid             | numeric(38,0) | not null
  unitid            | numeric(38,0) | not null
  device_type       | numeric(38,0) | not null
  placement         | numeric(38,0) | not null default 0
  unserved          | numeric(38,0) | not null default 0
  r_brkconn         | json          |
Indexes:
     "r_agrio_total_pkey" PRIMARY KEY, btree (id)
     "u_r_agrio_total" UNIQUE, btree (tagid, unitid, device_type, placement)
     "unit_r_agrio_total" btree (unitid)
     "tag_r_agrio_total" btree (tagid)

=> \d r_agrio_hourly
       Column       |         Type          |
  Modifiers
-------------------+-----------------------+-------------------------------------------------------------
  id                | bigint                | not null default
nextval('r_agrio_hourly_id_seq'::regclass)
  tagid             | numeric(38,0)         | not null
  unitid            | numeric(38,0)         | not null
  rowdate           | character varying(15) | not null
  device_type       | numeric(38,0)         | not null
  placement         | numeric(38,0)         | not null default 0
  unserved          | numeric(38,0)         | not null default 0
  unconfirmed       | numeric(38,0)         | not null default 0
  r_brkconn         | json                  |
Indexes:
     "r_agrio_hourly_pkey" PRIMARY KEY, btree (id)
     "u_r_agrio_hourly" UNIQUE, btree (tagid, unitid, rowdate,
device_type, placement)
     "unit_r_agrio_hourly" btree (unitid)
     "rowdate_r_agrio_hourly" btree (rowdate)
     "tag_r_agrio_hourly" btree (tagid)

Would appreciate any help on this, thank you!

What is the ->> operator? I'm not familiar with it.

numeric is FAR slower than int/bigint. Unless you *really* need 38 digits, use one of the other types.

Cursors tend to make things slow. Avoid them if you can.

As for your specific question, I suggest you modify the plpgsql function so that it's doing an EXPLAIN ANALYZE on the slow table. EXPLAIN ANALYZE actually returns a recordset the same way a SELECT would, with a single column of type text. So you just need to do something with that output. The easiest thing would be to replace this in your function:

UPDATE slow_table SET ...

to this (untested)

RETURN QUERY EXPLAIN ANALYZE UPDATE slow_table SET ...

and change the function so it returns SETOF text instead of whatever it returns now.

r_agrio_hourly - "good", r_agrio_total - "bad".

 Update on r_agrio_hourly  (cost=0.42..970.32 rows=250 width=329) (actual time=2.248..2.248 rows=0 loops=1)

   ->  Index Scan using u_r_agrio_hourly on r_agrio_hourly  (cost=0.42..970.32 rows=250 width=329) (actual time=0.968..1.207 rows=1 loops=1)

         Index Cond: ((tagid = 1002::numeric) AND (unitid = 1002::numeric) AND ((rowdate)::text = '2015-04-09T23'::text) AND (device_type = 3::numeric) AND (placement = 2::numeric))

 Total runtime: 2.281 ms

 Update on r_agrio_total  (cost=0.42..45052.56 rows=12068 width=321) (actual time=106.766..106.766 rows=0 loops=1)

   ->  Index Scan using u_r_agrio_total on r_agrio_total  (cost=0.42..45052.56 rows=12068 width=321) (actual time=0.936..32.626 rows=1 loops=1)

         Index Cond: ((tagid = 1002::numeric) AND (unitid = 1002::numeric) AND (device_type = 3::numeric) AND (placement = 2::numeric))

 Total runtime: 106.793 ms

On Mon, Apr 13, 2015 at 5:01 PM, Pawel Veselov <pawel.veselov@gmail.com> wrote:

r_agrio_hourly - "good", r_agrio_total - "bad".

 Update on r_agrio_hourly  (cost=0.42..970.32 rows=250 width=329) (actual time=2.248..2.248 rows=0 loops=1)

   ->  Index Scan using u_r_agrio_hourly on r_agrio_hourly  (cost=0.42..970.32 rows=250 width=329) (actual time=0.968..1.207 rows=1 loops=1)

         Index Cond: ((tagid = 1002::numeric) AND (unitid = 1002::numeric) AND ((rowdate)::text = '2015-04-09T23'::text) AND (device_type = 3::numeric) AND (placement = 2::numeric))

 Total runtime: 2.281 ms

 Update on r_agrio_total  (cost=0.42..45052.56 rows=12068 width=321) (actual time=106.766..106.766 rows=0 loops=1)

   ->  Index Scan using u_r_agrio_total on r_agrio_total  (cost=0.42..45052.56 rows=12068 width=321) (actual time=0.936..32.626 rows=1 loops=1)

         Index Cond: ((tagid = 1002::numeric) AND (unitid = 1002::numeric) AND (device_type = 3::numeric) AND (placement = 2::numeric))

 Total runtime: 106.793 ms

What it is you expect to see here?

​What are the results (count and times) for:

SELECT count(*) FROM r_agrio_total WHERE tagid = 1002 and unitid = 1002;

SELECT count(*) FROM r_agrio_hourly WHERE tagid = 1002 and unitid = 1002;

​More queries along this line might be needed.  The underlying question is how many index rows need to be skipped over on "total" to get the final result - or rather are the columns in the index in descending order of cardinality? 

Any chance you can perform a "REINDEX" - maybe there is some bloat present?  There are queries to help discern if that may be the case, I do not know then off the top of my head, but just doing it might be acceptable and is definitely quicker if so. 

​I'm still not really following your presentation but maybe my thoughts will spark something.​

On 4/13/15 7:01 PM, Pawel Veselov wrote:

    Cursors tend to make things slow. Avoid them if you can.


Is there an alternative to iterating over a number of rows, where a
direct update query is not an option?

I really doubt that either the actual processing logic, including use of
types has anything to do with my problem. This is based on the fact that
out of the tables that are being changed, only one is exhibiting the
problem. All of the involved tables have nearly the same structure, and
have the same logical operations performed on them. I thought may be the
"bad" table is slow because it was first in the list, and Postgres was
caching the functions results, but I moved things around, and pattern is
the same.

I'm guessing that you're essentially processing a queue. Take a look at http://www.postgresql.org/message-id/552C750F.2010401@BlueTreble.com for some ideas. Basically, not only do cursors have non-trivial overhead, doing a ton of single-row queries is going to have a non-trivial overhead itself.

    As for your specific question, I suggest you modify the plpgsql
    function so that it's doing an EXPLAIN ANALYZE on the slow table.
    EXPLAIN ANALYZE actually returns a recordset the same way a SELECT
    would, with a single column of type text. So you just need to do
    something with that output. The easiest thing would be to replace
    this in your function:

    UPDATE slow_table SET ...

    to this (untested)

    RETURN QUERY EXPLAIN ANALYZE UPDATE slow_table SET ...

    and change the function so it returns SETOF text instead of whatever
    it returns now.


Thank you, that made it a lot easier to see into what's really going on.
But the outcome is somewhat the same. The "bad" table analysis shows a
very high cost, and thousands of rows, where the table contains only 24
rows. This time, however, the actual run time is shown, and one can see
where the time is spent (I was using just a sum of clock_time()s around
the update statements to see where the problem is).

r_agrio_hourly - "good", r_agrio_total - "bad".

  Update on r_agrio_hourly  (cost=0.42..970.32 rows=250 width=329)
(actual time=2.248..2.248 rows=0 loops=1)
  ->  Index Scan using u_r_agrio_hourly on r_agrio_hourly
  (cost=0.42..970.32 rows=250 width=329) (actual time=0.968..1.207
rows=1 loops=1)
          Index Cond: ((tagid = 1002::numeric) AND (unitid =
1002::numeric) AND ((rowdate)::text = '2015-04-09T23'::text) AND
(device_type = 3::numeric) AND (placement = 2::numeric))
  Total runtime: 2.281 ms
  Update on r_agrio_total  (cost=0.42..45052.56 rows=12068 width=321)
(actual time=106.766..106.766 rows=0 loops=1)
  ->  Index Scan using u_r_agrio_total on r_agrio_total
  (cost=0.42..45052.56 rows=12068 width=321) (actual time=0.936..32.626
rows=1 loops=1)
          Index Cond: ((tagid = 1002::numeric) AND (unitid =
1002::numeric) AND (device_type = 3::numeric) AND (placement = 2::numeric))
  Total runtime: 106.793 ms

Keep in mind that the estimated cost is not terribly useful; it's the actual times that matter.

I suspect what's happening here is a combination of things. First, the hourly table is basically living in cache, but the total table is not. That means that when you go to find a row in the total table you're actually hitting the disk instead of pulling the data from memory.

Second, you may have a lot of dead rows in the total table. I suspect this because of the very large amount of time the index scan is taking. Unless you're running on an old 10MB MFM drive you'd be pretty hard pressed for even 2 IO operations (one for the index leaf page and one for the heap page) to take 32ms. I suspect the index scan is having to read many dead rows in before it finds a live one, and incurring multiple IOs. Swiching to EXPLAIN (analyze, buffers) would help confirm that.

Third, I think something odd is happening with the update itself. I'm pretty sure that the index scan itself is visiting the heap pages, so each page should be in shared buffers by the time each tuple hits the update node. That makes me wonder what on earth is taking 60ms to update the tuple. I suspect it's going into either finding a free buffer to put the new tuple on, or waiting to try and extend the relation. Selecting ctid from the freshly updated rows and comparing the first number to the total number of pages in the heap would show if the new tuples are all ending up at the end of the heap.

On 4/14/15 1:28 PM, Pawel Veselov wrote:

I wonder if what I need to do, considering that I update a lot of "the
same" rows as I process this queue, is to create a temp table, update
the rows there, and then update the actual tables once at the end...

That's what I'd do.

The other option would be to use a constraint trigger paired with a per-row trigger on the hourly table to drive the daily table, and on the daily table to drive the total table. The way that would work is the per-row table would simply keep track of all the unique records that were changed in a statement (presumably by putting them in a temp table). Once the statement is "done", the constraint trigger would fire; it would summarize all the changed data and do a much smaller number of updates to the table being summarized into.

BTW, you also made a comment about not having to hit the table if you look at something in an index. You can only do that if all the data you need is in the index, AND the page with the record is marked as being all-visible (google for Postgres Visibility Map). If that's not the case then you still have to pull the row in the table in, in order to determine visibility. The only case where you can still avoid hitting the table is something like a NOT EXISTS; if you can't find any entries in the index for something then they definitely won't be in the table.

But remember that if you update or delete a row, removing it from an index, the data will stay in that index until vacuum comes along.

Also, there's no point in doing a REINDEX after a VACUUM FULL; vacuum full rebuilds all the indexes for you.

On 4/14/15 4:44 PM, Pawel Veselov wrote:

On Tue, Apr 14, 2015 at 1:15 PM, Jim Nasby <Jim.Nasby@bluetreble.com

<mailto:Jim.Nasby@bluetreble.com>> wrote:

    On 4/14/15 1:28 PM, Pawel Veselov wrote:


        I wonder if what I need to do, considering that I update a lot
        of "the
        same" rows as I process this queue, is to create a temp table,
        update
        the rows there, and then update the actual tables once at the end...


    That's what I'd do.


Well, in short, I changed (repeat the body of loop for how many tables
are there)

LOOP (item)
   UPDATE table with item
   IF not found INSERT item INTO table; END IF;
END LOOP;

to:

CREATE TEMP TABLE xq_table (like table) on commit drop;
LOOP (item)
   LOOP
     UPDATE xq_table with item;
     exit when found;
     INSERT INTO xq_table select * from table for update;
     continue when found;
     INSERT item INTO xq_table;
     exit;
   END LOOP;
END LOOP;
UPDATE table a set (rows) = (xq.rows)
   FROM xq_table xq
   WHERE (a.keys) = (xq.keys)

That works significantly faster. The final update statement is very
fast. The process is somewhat slow in the beginning as it sucks in
records from "total" into "xq_total", but once all of that is moved into
the temp table, it rushes through the rest.

Databases like to think in sets. It will generally be more efficient to do set operations instead of a bunch of row-by-row stuff.

Since you're pulling all of this from some other table your best bet is probably something like:

CREATE TEMP TABLE raw AS DELETE FROM queue WHERE ... RETURNING *;

CREATE TEMP VIEW hourly_v AS SELECT ... FROM raw GROUP BY;
UPDATE ar_hourly SET ... FROM hourly_v JOIN ...;
INSERT INTO ar_hourly SELECT FROM hourly_v LEFT JOIN ar_hourly ON ...;

-- Same thing for daily
-- Same thing for total

    But remember that if you update or delete a row, removing it from an
    index, the data will stay in that index until vacuum comes along.

    Also, there's no point in doing a REINDEX after a VACUUM FULL;
    vacuum full rebuilds all the indexes for you.


I was being desperate :)

I still think there is something very wrong with this particular table.
First, I have production systems that employ this function on way larger
data set, and there is no problem (so far, but still). This machine is
part of a test deployment, there is no constant load, the only data that
is being written now is when I do these tests. Vacuuming should prune
all that dead stuff, and if it's absent, it's unclear where is the time
spent navigating/updating the table with 24 rows :)

I think you definitely have a problem with dead rows, as evidenced by the huge improvement VACUUM FULL made.

[skipped]

 

This is where using sets becomes really tedious, as Postgres severely lacks an upsert-like statement.

I don't think there are joins allowed in UPDATE statement, so I will need to use WITH query, right?

Also, I'm not sure how LEFT JOIN will help me isolate and insert "missed" entries...

 

[skipped]

 

Yes, you can do UPDATE with joins (http://www.postgresql.org/docs/9.4/static/sql-update.html) like this:

 

UPDATE table1 A SET col1 = B.col2

  FROM table2 B

  WHERE A.col3 = B.col4;

[skipped]

 

    But remember that if you update or delete a row, removing it from an
    index, the data will stay in that index until vacuum comes along.

    Also, there's no point in doing a REINDEX after a VACUUM FULL;
    vacuum full rebuilds all the indexes for you.


I was being desperate :)

I still think there is something very wrong with this particular table.
First, I have production systems that employ this function on way larger
data set, and there is no problem (so far, but still). This machine is
part of a test deployment, there is no constant load, the only data that
is being written now is when I do these tests. Vacuuming should prune
all that dead stuff, and if it's absent, it's unclear where is the time
spent navigating/updating the table with 24 rows :)

I think you definitely have a problem with dead rows, as evidenced by the huge improvement VACUUM FULL made.

But it's not clear why (and not reasonable, IMHO, that) it wouldn't improve past current point.

I found some dangling prepared transactions