Thread: index creation order?
Yet another question.. thanks to everyone responding to all these so far.. ;)
This one is basically.. given I have a big table already in COPY format,
about 28 million rows, all keys guaranteed to be unique, I'm trying to find
out which of the following will get the import finished the fastest:
a) CREATE TABLE with no indexes or keys. Run the COPY (fast, ~30min), then
CREATE INDEX on each column it's needed on, and ALTER TABLE for the pk and
each fk needed.
b) Same as above, but instead of ALTER TABLE -- ditch the FK, and CREATE
UNIQUE INDEX on the PK.
c) CREATE TABLE with the PK/FK's in the table structure, CREATE INDEX on
needed columns, then run the COPY.
d) .. is to c as b is to a .. Don't create PK/FK's, just CREATE UNIQUE
INDEX after table creation, then run the COPY.
My gut instinct tells me that in order, fastest to slowest, it's going to
be d,b,c,a; this is what I've experienced on other DBs such as MSSQL and
Oracle.
If there isn't a significant difference between all of them, performance
wise, I think something is dreadfully wrong here. Running "a", the ALTER
TABLE to add the PK ran for 17 hours and still wasn't finished.
The table without indexes or keys is:
CREATE TABLE foo (
id BIGINT NOT NULL DEFAULT nextval('foo_id_sequence'),
master_id BIGINT NOT NULL,
other_id INTEGER NOT NULL,
status INTEGER NOT NULL,
addtime TIMESTAMP WITH TIME ZONE DEFAULT now()
);
Details on machine and configuration are:
The machine is the same one I've mentioned before.. SMP AthlonMP 2800+
(2.1GHz), 4x18GB 15krpm SCSI RAID-0 with 256MB onboard cache on a
quad-channel ICP-Vortex controller, 2GB physical memory. Running FreeBSD
RELENG_4, relevant filesystems with softupdates enabled and mounted noatime.
kernel options are:
maxusers 0
options MAXDSIZ="(1536UL*1024*1024)" # maximum limit
options MAXSSIZ="(512UL*1024*1024)" # maximum stack
options DFLDSIZ="(512UL*1024*1024)" # default limit
options VM_BCACHE_SIZE_MAX="(384UL*1024*1024)" # cache size upped
from default 200MB
options SYSVSHM #SYSV-style shared memory
options SYSVMSG #SYSV-style message queues
options SYSVSEM #SYSV-style semaphores
options SHMMAXPGS=262144
options SHMALL=262144
options SHMSEG=256
options SEMMNI=384
options SEMMNS=768
options SEMMNU=384
options SEMMAP=384
postgresql.conf settings are:
shared_buffers = 30000
max_fsm_relations = 10000
max_fsm_pages = 2000000
max_locks_per_transaction = 64
wal_buffers = 128
sort_mem = 1310720 (1.2GB)
vacuum_mem = 262144 (256MB)
checkpoint_segments = 64
checkpoint_timeout = 1200
commit_delay = 20000
commit_siblings = 2
fsync=true
random_page_cost = 1.7
cpu_tuple_cost = 0.005
cpu_index_tuple_cost = 0.005
cpu_operator_cost = 0.0012
stats_start_collector = true
stats_command_string = true
stats_row_level = true
stats_block_level = true
If it is 7.4 beta 5 or later, I would definitely go with A.
Adding indexes after the fact seems to be much quicker. Foreign keys use
the same algorithm prior to beta 5 regardless of timing.
A primary key and unique index will have approx the same performance (a
check for NULL isn't very costly).
On Fri, 2003-10-31 at 11:02, Allen Landsidel wrote:
> Yet another question.. thanks to everyone responding to all these so far.. ;)
>
> This one is basically.. given I have a big table already in COPY format,
> about 28 million rows, all keys guaranteed to be unique, I'm trying to find
> out which of the following will get the import finished the fastest:
>
> a) CREATE TABLE with no indexes or keys. Run the COPY (fast, ~30min), then
> CREATE INDEX on each column it's needed on, and ALTER TABLE for the pk and
> each fk needed.
>
> b) Same as above, but instead of ALTER TABLE -- ditch the FK, and CREATE
> UNIQUE INDEX on the PK.
>
> c) CREATE TABLE with the PK/FK's in the table structure, CREATE INDEX on
> needed columns, then run the COPY.
>
> d) .. is to c as b is to a .. Don't create PK/FK's, just CREATE UNIQUE
> INDEX after table creation, then run the COPY.
>
> My gut instinct tells me that in order, fastest to slowest, it's going to
> be d,b,c,a; this is what I've experienced on other DBs such as MSSQL and
> Oracle.
>
> If there isn't a significant difference between all of them, performance
> wise, I think something is dreadfully wrong here. Running "a", the ALTER
> TABLE to add the PK ran for 17 hours and still wasn't finished.
>
> The table without indexes or keys is:
> CREATE TABLE foo (
> id BIGINT NOT NULL DEFAULT nextval('foo_id_sequence'),
> master_id BIGINT NOT NULL,
> other_id INTEGER NOT NULL,
> status INTEGER NOT NULL,
> addtime TIMESTAMP WITH TIME ZONE DEFAULT now()
> );
>
> Details on machine and configuration are:
>
> The machine is the same one I've mentioned before.. SMP AthlonMP 2800+
> (2.1GHz), 4x18GB 15krpm SCSI RAID-0 with 256MB onboard cache on a
> quad-channel ICP-Vortex controller, 2GB physical memory. Running FreeBSD
> RELENG_4, relevant filesystems with softupdates enabled and mounted noatime.
>
> kernel options are:
> maxusers 0
>
> options MAXDSIZ="(1536UL*1024*1024)" # maximum limit
> options MAXSSIZ="(512UL*1024*1024)" # maximum stack
> options DFLDSIZ="(512UL*1024*1024)" # default limit
> options VM_BCACHE_SIZE_MAX="(384UL*1024*1024)" # cache size upped
> from default 200MB
> options SYSVSHM #SYSV-style shared memory
> options SYSVMSG #SYSV-style message queues
> options SYSVSEM #SYSV-style semaphores
> options SHMMAXPGS=262144
> options SHMALL=262144
> options SHMSEG=256
> options SEMMNI=384
> options SEMMNS=768
> options SEMMNU=384
> options SEMMAP=384
>
> postgresql.conf settings are:
>
> shared_buffers = 30000
> max_fsm_relations = 10000
> max_fsm_pages = 2000000
> max_locks_per_transaction = 64
> wal_buffers = 128
> sort_mem = 1310720 (1.2GB)
> vacuum_mem = 262144 (256MB)
> checkpoint_segments = 64
> checkpoint_timeout = 1200
> commit_delay = 20000
> commit_siblings = 2
> fsync=true
> random_page_cost = 1.7
> cpu_tuple_cost = 0.005
> cpu_index_tuple_cost = 0.005
> cpu_operator_cost = 0.0012
>
> stats_start_collector = true
> stats_command_string = true
> stats_row_level = true
> stats_block_level = true
>
>
> ---------------------------(end of broadcast)---------------------------
> TIP 2: you can get off all lists at once with the unregister command
> (send "unregister YourEmailAddressHere" to majordomo@postgresql.org)
>
Attachment
Allen, > a) CREATE TABLE with no indexes or keys. Run the COPY (fast, ~30min), then > CREATE INDEX on each column it's needed on, and ALTER TABLE for the pk and > each fk needed. Did you ANALYZE after the copy? > If there isn't a significant difference between all of them, performance > wise, I think something is dreadfully wrong here. Running "a", the ALTER > TABLE to add the PK ran for 17 hours and still wasn't finished. Adding the *primary key* locked up? This seems unlikely; we have a known problem with *foreign* keys until the current beta. But I've added primary keys on 20Gb tables and had it complete in a couple of hours. Ignore this adivice and look for Stephan Szabo's FK patch instead if what you really meant was that the FK creation locked up. > shared_buffers = 30000 hmmm ... 236MB .... > max_fsm_pages = 2000000 2MB, fine ... > wal_buffers = 128 1MB, also fine ... > sort_mem = 1310720 (1.2GB) Problem here. As documented everywhere, sort_mem is allocated *per sort* not per query, user, or shared. This means that if the "add PK" operation involves 2 or more sorts (not sure, haven't tested it), then you're allocating .7GB RAM more than you acutally have. This may be the cause of your problem, particularly if *anything* is going on concurrent to the load. > checkpoint_segments = 64 IF you have the disk space (+ 2GB) I'd raise this to 150-300 during the load operation. > commit_delay = 20000 > commit_siblings = 2 These settings are for heavy multi-user update activity. They are not useful for a single-user load, and may even lower performance. > stats_start_collector = true > stats_command_string = true > stats_row_level = true > stats_block_level = true If you can do without stats collection during load, I would suggest that you do so. The above add both RAM and I/O overhead to your operation. -- Josh Berkus Aglio Database Solutions San Francisco
At 12:10 10/31/2003, Josh Berkus wrote: >Allen, > > > a) CREATE TABLE with no indexes or keys. Run the COPY (fast, ~30min), then > > CREATE INDEX on each column it's needed on, and ALTER TABLE for the pk and > > each fk needed. > >Did you ANALYZE after the copy? No, and this was my major mistake. I normally run analyze periodically from cron, anywhere from once an hour to ever 15 minutes depending on the db.. I had disabled that for this because I didn't want anything competing with this stuff for disk I/O. I followed your other suggestions as well, canceled the index that was running, analyzed the whole db, and ran the queries again. All of them are running in under 10 or so minutes after the analyze. I'll just be adding the PKs and the Indexes, I can add triggers/rules of my own for the RI, rather than worry about FK creation screwing up. I had no idea analyze was playing such a big role in this sense.. I really thought that other than saving space, it wasn't doing much for tables that don't have indexes on the. Thanks for the help. > > shared_buffers = 30000 >hmmm ... 236MB .... > > max_fsm_pages = 2000000 >2MB, fine ... > > wal_buffers = 128 >1MB, also fine ... > > sort_mem = 1310720 (1.2GB) >Problem here. As documented everywhere, sort_mem is allocated *per sort* >not >per query, user, or shared. This means that if the "add PK" operation >involves 2 or more sorts (not sure, haven't tested it), then you're >allocating .7GB RAM more than you acutally have. This may be the cause of >your problem, particularly if *anything* is going on concurrent to the load. I didn't know this was per-sort per-backend, I thought it was per-backend for all sorts running on that backend. I've dropped it down to 256MB. > > checkpoint_segments = 64 >IF you have the disk space (+ 2GB) I'd raise this to 150-300 during the load >operation. Done, at 128, which seems to be enough for now. I'll fiddle more with this later on. > > commit_delay = 20000 > > commit_siblings = 2 >These settings are for heavy multi-user update activity. They are not useful >for a single-user load, and may even lower performance. That's what's going on.. this database I'm working on isn't the only one in the system, and some things are using different schemas in the database I'm working on, so this isn't something I can afford to turn off. Most of the activity is heavy and transient.. many INSERT/UPDATE/DELETE cycles. Again, thanks for the help, I really do appreciate it. It's gratifying and depressing to know the last two or so days work could've been compressed into 3 hours if I'd just run that damn analyze. ;)
Nope, still 7.3.4 here.. I am very excited about 7.4 though.. almost as
excited as I am about FreeBSD 5.x going -STABLE.. it's a close race
between the two..
I'll keep this in mind for when I update though, thanks.
At 11:23 10/31/2003, Rod Taylor wrote:
>If it is 7.4 beta 5 or later, I would definitely go with A.
>
>Adding indexes after the fact seems to be much quicker. Foreign keys use
>the same algorithm prior to beta 5 regardless of timing.
>
>A primary key and unique index will have approx the same performance (a
>check for NULL isn't very costly).
>
>On Fri, 2003-10-31 at 11:02, Allen Landsidel wrote:
> > Yet another question.. thanks to everyone responding to all these so
> far.. ;)
> >
> > This one is basically.. given I have a big table already in COPY format,
> > about 28 million rows, all keys guaranteed to be unique, I'm trying to
> find
> > out which of the following will get the import finished the fastest:
> >
> > a) CREATE TABLE with no indexes or keys. Run the COPY (fast, ~30min),
> then
> > CREATE INDEX on each column it's needed on, and ALTER TABLE for the pk and
> > each fk needed.
> >
> > b) Same as above, but instead of ALTER TABLE -- ditch the FK, and CREATE
> > UNIQUE INDEX on the PK.
> >
> > c) CREATE TABLE with the PK/FK's in the table structure, CREATE INDEX on
> > needed columns, then run the COPY.
> >
> > d) .. is to c as b is to a .. Don't create PK/FK's, just CREATE UNIQUE
> > INDEX after table creation, then run the COPY.
> >
> > My gut instinct tells me that in order, fastest to slowest, it's going to
> > be d,b,c,a; this is what I've experienced on other DBs such as MSSQL and
> > Oracle.
> >
> > If there isn't a significant difference between all of them, performance
> > wise, I think something is dreadfully wrong here. Running "a", the ALTER
> > TABLE to add the PK ran for 17 hours and still wasn't finished.
> >
> > The table without indexes or keys is:
> > CREATE TABLE foo (
> > id BIGINT NOT NULL DEFAULT nextval('foo_id_sequence'),
> > master_id BIGINT NOT NULL,
> > other_id INTEGER NOT NULL,
> > status INTEGER NOT NULL,
> > addtime TIMESTAMP WITH TIME ZONE DEFAULT now()
> > );
> >
> > Details on machine and configuration are:
> >
> > The machine is the same one I've mentioned before.. SMP AthlonMP 2800+
> > (2.1GHz), 4x18GB 15krpm SCSI RAID-0 with 256MB onboard cache on a
> > quad-channel ICP-Vortex controller, 2GB physical memory. Running FreeBSD
> > RELENG_4, relevant filesystems with softupdates enabled and mounted
> noatime.
> >
> > kernel options are:
> > maxusers 0
> >
> > options MAXDSIZ="(1536UL*1024*1024)" # maximum limit
> > options MAXSSIZ="(512UL*1024*1024)" # maximum stack
> > options DFLDSIZ="(512UL*1024*1024)" # default limit
> > options VM_BCACHE_SIZE_MAX="(384UL*1024*1024)" # cache size upped
> > from default 200MB
> > options SYSVSHM #SYSV-style shared memory
> > options SYSVMSG #SYSV-style message queues
> > options SYSVSEM #SYSV-style semaphores
> > options SHMMAXPGS=262144
> > options SHMALL=262144
> > options SHMSEG=256
> > options SEMMNI=384
> > options SEMMNS=768
> > options SEMMNU=384
> > options SEMMAP=384
> >
> > postgresql.conf settings are:
> >
> > shared_buffers = 30000
> > max_fsm_relations = 10000
> > max_fsm_pages = 2000000
> > max_locks_per_transaction = 64
> > wal_buffers = 128
> > sort_mem = 1310720 (1.2GB)
> > vacuum_mem = 262144 (256MB)
> > checkpoint_segments = 64
> > checkpoint_timeout = 1200
> > commit_delay = 20000
> > commit_siblings = 2
> > fsync=true
> > random_page_cost = 1.7
> > cpu_tuple_cost = 0.005
> > cpu_index_tuple_cost = 0.005
> > cpu_operator_cost = 0.0012
> >
> > stats_start_collector = true
> > stats_command_string = true
> > stats_row_level = true
> > stats_block_level = true
> >
> >
> > ---------------------------(end of broadcast)---------------------------
> > TIP 2: you can get off all lists at once with the unregister command
> > (send "unregister YourEmailAddressHere" to majordomo@postgresql.org)
> >
On Fri, 2003-10-31 at 13:27, Allen Landsidel wrote: > I had no idea analyze was playing such a big role in this sense.. I really > thought that other than saving space, it wasn't doing much for tables that > don't have indexes on the. ANALYZE doesn't save any space at all -- VACUUM is probably what you're thinking of. -Neil
Allen, > I had no idea analyze was playing such a big role in this sense.. I really > thought that other than saving space, it wasn't doing much for tables that > don't have indexes on the. Among other things, ANALYZE tells postgres how many rows are in the table. So if you add a PK constraint after loading 10 million rows without ANALYZE, PostgreSQL is likely to think that there is only one row in the table ... and choose a nested loop or some other really inefficient method of checking for uniqueness. -- Josh Berkus Aglio Database Solutions San Francisco
At 13:40 10/31/2003, Neil Conway wrote: >On Fri, 2003-10-31 at 13:27, Allen Landsidel wrote: > > I had no idea analyze was playing such a big role in this sense.. I really > > thought that other than saving space, it wasn't doing much for tables that > > don't have indexes on the. > >ANALYZE doesn't save any space at all -- VACUUM is probably what you're >thinking of. Actually, I was thinking VACUUM ANALYZE.. which is what I ran after the COPY.. sorry for my lack of precision. I've yet to run straight-up ANALYZE AFAIK. -Allen
is there any way to update the stats inside a transaction? what i have is something like: select count(*) from foo; -> 0 begin; copy foo from '/tmp/foo'; -- about 100k rows -- run some queries on foo which perform horribly because the stats -- are way off (100k rows v. 0 rows) commit; it seems that you cannot run analyze inside a transaction: begin; analyze foo; ERROR: ANALYZE cannot run inside a BEGIN/END block i am using version 7.2.3. any work-a-rounds? should i try updating pg_statistic manually? On Fri, 31 Oct 2003, Josh Berkus wrote: > Among other things, ANALYZE tells postgres how many rows are in the table. So > if you add a PK constraint after loading 10 million rows without ANALYZE, > PostgreSQL is likely to think that there is only one row in the table ... and > choose a nested loop or some other really inefficient method of checking for > uniqueness.
> begin; > analyze foo; > ERROR: ANALYZE cannot run inside a BEGIN/END block > > i am using version 7.2.3. Time to upgrade. 7.3 / 7.4 allows this to happen.
Attachment
Chester Kustarz <chester@arbor.net> writes:
> it seems that you cannot run analyze inside a transaction:
You can in 7.3.* ...
regards, tom lane