On Thu, 19 Oct 2006 16:55:34 +0200, Richard Huxton <dev@xxxxxxxxxxxx>
wrote:
Jens Schipkowski wrote:
Hi,
we've got performance problems due to repeating SELECT, UPDATE,
DELETE, INSERT statements. This statements have to be executed every 10
seconds, but they run into a timeout.
To obviate problems regarding to our Java Software and JDBC drivers, we
put the repeating sequence of statements to a file more than 100k times
(half a million statements) and executed "psql ourDB -f ourFile.sql -o
/dev/null". To accelerate the occurence of the performance drop, we
started 6 instances of this command.
The performance drop occured after 10 minutes shifting the server to 0
percent idle and 85 - 95 percent user.
After 10 minutes of what?
start testing using the command above.
Did the half-million statements complete? If not, how many got
completed? Were they all in separate transactions or did you batch them?
How ofternwere you vacuuming here?
We wrote a sql batch file which simulates the repeating cycle of SELECT,
UPDATE, DELETE, INSERT. The INSERT is fired using another backend.
The half-million statements of this file will probably complete after all
SELECT and DELETE statements timed out.
We had 6 seperate transactions executing the batch file.
For tracing the statement which raised the load, we are using pg_locks,
pg_stat_activity with current_query enabled. The responsible statement
is the DELETE, it hangs until its canceled by timeout. The first run on
an vacuumed DB took 300 - 600ms.
In a second test we removed the DELETE statements to see wich
statements also needs longer time by increasing the amount of data.
After half an hour the SELECT statements timed out.
An additional VACUUM - every 1 minute - does extend the timeout
occurence by factor 5 - 6.
And running vacuum every 30 seconds does what?
Not yet fully tested. It seems to lower the slow down.
But minimizing the gain of slow down doesn't solve the problem. The
Problem is the decrease of execution speed of DELETE and SELECT statements
by a table row count between 150k - 200k. The table starts growing first
after DELETE statements fails during timeout.
It does not sound professional, but the database seems to be aging by
the increase of executed statements.
It sounds very likely if you aren't vacuuming enough or the tables are
growing rapidly.
vacuuming once a minute is not enough? We reach the execution of 3k
statements per minute (startup time of testing). 1/4 of them are INSERTs
and DELETEs. After 5 minutes a DELETE will took about 50 seconds -
compared to startup time about 300 - 600ms.
The Statements
---------------
// normal insert - does not seem to be the problem - runtime is ok
INSERT INTO
tbl_reg(idreg,idtype,attr1,...,attr6,loc1,...,loc3,register)
VALUES(nextval('tbl_reg_idreg_seq'),1,[attr],[loc],1);
// select finds out which one has not an twin
// a twin is defined as record with the same attr* values
// decreases speed over time until timeout by postgresql
SELECT *
FROM tbl_reg reg
WHERE register <> loc1 AND
idreg NOT IN
(
SELECT reg.idreg
FROM tbl_reg reg, tbl_reg regtwin
WHERE regtwin.register = 1 AND
regtwin.type <> 20 AND
reg.attr1 = regtwin.attr1 AND
reg.attr2 = regtwin.attr2 AND
reg.attr3 = regtwin.attr3 AND
reg.attr4 = regtwin.attr4 AND
reg.attr5 = regtwin.attr5 AND
reg.attr6 = regtwin.attr6 AND
reg.idreg <> regtwin.idreg AND
reg.register = 2
);
I tried to optimize the seslect statement but the group by having
count(*)
1 solution is half as fast as this statement - relating to the query
plan of EXPLAIN ANALYZE.
And what did EXPLAIN ANALYSE show here? I'm guessing you're getting time
increasing as the square of the number of rows in tbl_reg. So, if 50
rows takes 2500s then 100 rows will take 10000s. Now, if you had enough
RAM (or an index) I'd expect the planner to process the table in a
sorted manner so the query-time would increase linearly.
EXPLAIN ANALYZE shows at startup
QUERY PLAN
-------------------------------------------------------------------------------------------------
Seq Scan on tbl_reg (cost=25841.35..31433.17 rows=72424 width=6) (actual
time=673.877..673.877 rows=0 loops=1)
Filter: ((hashed subplan) OR ((loc1 = 2::smallint) AND (loc2 =
2::smallint)))
SubPlan
-> Merge Join (cost=22186.21..25646.57 rows=77913 width=8) (actual
time=285.624..285.624 rows=0 loops=1)
Merge Cond: (("outer".attr1 = "inner".attr1) AND
("outer"."?column8?" = "inner"."?column8?") AND ("outer"."?column9?" =
"inner"."?column9?") AND ("outer"."?column10?" = "inner"."?column10?") AND
("outer"."?column11?" = "inner"."?column11?") AND ("outer".attr6 =
"inner".attr6))
Join Filter: ("outer".idreg <> "inner".idreg)
-> Sort (cost=4967.06..4971.65 rows=1835 width=56) (actual
time=285.622..285.622 rows=0 loops=1)
Sort Key: reg.attr1, (reg.attr2)::text,
(reg.attr3)::text, (reg.attr4)::text, (reg.attr5)::text, reg.attr6
-> Seq Scan on tbl_reg reg (cost=0.00..4867.59
rows=1835 width=56) (actual time=285.551..285.551 rows=0 loops=1)
Filter: (register = 2)
-> Sort (cost=17219.15..17569.77 rows=140247 width=56) (never
executed)
Sort Key: regtwin.attr1, (regtwin.attr2)::text,
(regtwin.attr3)::text, (regtwin.attr4)::text, (regtwin.attr5)::text,
regtwin.attr6
-> Seq Scan on tbl_reg regtwin (cost=0.00..5229.70
rows=140247 width=56) (never executed)
Filter: ((register = 1) AND ("type" <> 20))
Total runtime: 604.410 ms
(15 rows)
EXPLAIN ANALYZE shows after 10 minutes load and 1x vacuuming
QUERY PLAN
----------------------------------------------------------------------------------------------------
Seq Scan on tbl_reg (cost=25841.35..31433.17 rows=72424 width=6) (actual
time=43261.910..43261.910 rows=0 loops=1)
Filter: ((hashed subplan) OR ((loc1 = 2::smallint) AND (loc2 =
2::smallint)))
SubPlan
-> Merge Join (cost=22186.21..25646.57 rows=77913 width=8) (actual
time=43132.063..43132.063 rows=0 loops=1)
Merge Cond: (("outer".attr1 = "inner".attr1) AND
("outer"."?column8?" = "inner"."?column8?") AND ("outer"."?column9?" =
"inner"."?column9?") AND ("outer"."?column10?" = "inner"."?column10?") AND
("outer"."?column11?" = "inner"."?column11?") AND ("outer".attr6 =
"inner".attr6))
Join Filter: ("outer".idreg <> "inner".idreg)
-> Sort (cost=4967.06..4971.65 rows=1835 width=56) (actual
time=387.071..387.872 rows=1552 loops=1)
Sort Key: reg.attr1, (reg.attr2)::text,
(reg.attr3)::text, (reg.attr4)::text, (reg.attr5)::text, reg.attr6
-> Seq Scan on tbl_reg reg (cost=0.00..4867.59
rows=1835 width=56) (actual time=303.966..325.614 rows=1552 loops=1)
Filter: (register = 2)
-> Sort (cost=17219.15..17569.77 rows=140247 width=56)
(actual time=42368.934..42530.986 rows=145324 loops=1)
Sort Key: regtwin.attr1, (regtwin.attr2)::text,
(regtwin.attr3)::text, (regtwin.attr4)::text, (regtwin.attr5)::text,
regtwin.attr6
-> Seq Scan on tbl_reg regtwin (cost=0.00..5229.70
rows=140247 width=56) (actual time=0.015..1159.515 rows=145453 loops=1)
Filter: ((register = 1) AND ("type" <> 20))
Total runtime: 44073.127 ms
(15 rows)
I know that the second query plan executes the sort, because it finds
matching data. So maybe indexing will help.
Oh, and you're doing one join more than you need to here (counting the
NOT IN as a join). You could get by with a LEFT JOIN and a test for
idreg being null on the right-hand table.
It sounds good. First tests doesn't improve runtime - it needs more
extensive testing.
// delete data without a twin
// drastically decreases speed over time until timeout by postgresql
[snip delete doing the same query as above]
The runtime of this statement increases until it will canceled by
PostgreSQL.
// the where clause of this update statement is normally build in java
UPDATE tbl_reg SET loc1=2 WHERE idreg IN ('...',...,'...');
The Table
---------------
Tested with: 20.000, 80.000, 500.000 records
CREATE TABLE tbl_reg
(
idreg bigserial NOT NULL,
idtype int8 DEFAULT 0,
You can have more than 4 billion "types"?
it seems so, or not?
attr1 int4,
attr2 int4,
attr3 varchar(20),
attr4 varchar(20),
attr5 int4,
attr6 varchar(140) DEFAULT ''::character varying,
loc1 int2 DEFAULT 0,
loc2 int2 DEFAULT 0,
loc3 int2 DEFAULT 0,
register int2 DEFAULT 1,
"timestamp" timestamp DEFAULT now(),
You probably want timestamp with time zone.
No, just the server time is important. This is short living data.
CONSTRAINT tbl_reg_pkey PRIMARY KEY (idreg)
)
WITHOUT OIDS;
The Hardware
----------------
Dual Xeon 3.2GHz Hyperthreading
SCSI harddrives
RAID and non-RAID tested
We have the problem, that we cannot see any potential to improve SQL
statements. Indexing the attr* columns seems not to be an solution,
because the data mustn't be unique (twins) and changes really often so
reindexing will took too long.
Eh? Why would an index force uniqueness? And are you sure that adding an
index makes updates too slow? What did your testing show as the
slow-down? I'd be tempted to put an index on attr1,attr2,attr5 (or
whichever combination provides the most selectivity) then make sure your
statistics are up to date (ANALYSE) and see if the plans change.
OK, I misunderstood the PostgreSQL INDEX. Will test it using an
multicolumn index.
Of course, that's assuming your postgresql.conf has some reasonable
performance-related settings.
postgresql.conf settings have been optimized. Searched the web for useful
information and got help from Mailing list by Tom Lane.
Oh, I'd also wonder whether, with "twin-ness" being such an important
concept it isn't its own thing and thus perhaps deserve its own table.
It's important due to software concept (conferencing groups).
Thank you for your suggestions. I will add indexes to the table and
overhaul the SELECT and DELETE statements. After testing I will post
results.
