Обсуждение: VACUUM and transactions in different databases
Hi all If I have a running transaction in database1 and try to vacuum database2 but the dead tuples in database2 cannot be removed. INFO: vacuuming "public.dummy1" INFO: "dummy1": found 0 removable, 140000 nonremovable row versions in 1341 pages DETAIL: 135000 dead row versions cannot be removed yet. How can I achieve that database2 is vacuumed while a transaction in database1 is not yet commited? Regards conni
Cornelia Boenigk <c@cornelia-boenigk.de> wrote: > > Hi all > > If I have a running transaction in database1 and try to vacuum database2 > but the dead tuples in database2 cannot be removed. > > INFO: vacuuming "public.dummy1" > INFO: "dummy1": found 0 removable, 140000 nonremovable row versions in > 1341 pages > DETAIL: 135000 dead row versions cannot be removed yet. > > How can I achieve that database2 is vacuumed while a transaction in > database1 is not yet commited? I don't believe that's the reason. AFAIK, activity in one database will never block activity in another. I would suspect that you haven't vacuumed this database in a long time, and an ordinary vacuum can't reclaim that space. Can you run a "vacuum full", and does it reclaim the space? If you do regular vacuum often enough, you should never end up with so much unused space, unless your usage pattern is very drastic, in which case you should look at other methods of managing that table -- perhaps CLUSTER. -Bill
Hi Bill > I don't believe that's the reason. AFAIK, activity in one database > will never block activity in another. This way I read the documentation. > I would suspect that you haven't vacuumed this database in a long time, I created both databases one hour ago for just testing this behaviour. I started with two identical tables, each with 5000 rows in both databases. In db1 I opened a transaction, updated the table and left the transaction open. In db2 I updated, inserted and deleted a lot and then tried to vacuum. > Can you run a "vacuum > full", and does it reclaim the space? I tried but it hangs. [root@conni ~]# ps axw|grep postgres 1746 ? S 0:00 postgres: writer process 1747 ? S 0:00 postgres: stats buffer process 1748 ? S 0:00 postgres: stats collector process 2106 pts/1 S 0:00 su postgres 2120 pts/1 S+ 0:00 psql postgres 2188 ? S 0:04 postgres: postgres dummy1 [local] VACUUM waiting 2200 pts/3 S 0:00 su postgres 2215 ? S 0:00 postgres: postgres dummy2 [local] idle in transaction 2717 pts/2 R+ 0:00 grep postgres Regards Conni
Hi as soon as I committed the open transaction the hangig vacuum full completed and the table was vacuumed: regards Conni
Cornelia Boenigk <c@cornelia-boenigk.de> wrote:
>
> Hi Bill
>
> > I don't believe that's the reason. AFAIK, activity in one database
> > will never block activity in another.
>
> This way I read the documentation.
psql -U pgsql db2
Welcome to psql 8.2.0, the PostgreSQL interactive terminal.
Type: \copyright for distribution terms
\h for help with SQL commands
\? for help with psql commands
\g or terminate with semicolon to execute query
\q to quit
psql -U pgsql db1
Welcome to psql 8.2.0, the PostgreSQL interactive terminal.
Type: \copyright for distribution terms
\h for help with SQL commands
\? for help with psql commands
\g or terminate with semicolon to execute query
\q to quit
db1=# begin;
BEGIN
db1=# insert into t1 values (44, 'text string');
INSERT 0 1
db1=#
[1]+ Stopped psql -U pgsql db1
[wmoran@working ~]$ psql -U pgsql db2
Welcome to psql 8.2.0, the PostgreSQL interactive terminal.
Type: \copyright for distribution terms
\h for help with SQL commands
\? for help with psql commands
\g or terminate with semicolon to execute query
\q to quit
db2=# vacuum full;
VACUUM
db2=# \q
[wmoran@working ~]$ fg
psql -U pgsql db1
rollback;
ROLLBACK
Works that way for me ...
> > I would suspect that you haven't vacuumed this database in a long time,
>
> I created both databases one hour ago for just testing this behaviour. I
> started with two identical tables, each with 5000 rows in both
> databases. In db1 I opened a transaction, updated the table and left the
> transaction open.
>
> In db2 I updated, inserted and deleted a lot and then tried to vacuum.
So, long time then. My definition of "long time" is equal to your
definition of "a lot" :)
> > Can you run a "vacuum
> > full", and does it reclaim the space?
>
> I tried but it hangs.
>
> [root@conni ~]# ps axw|grep postgres
> 1746 ? S 0:00 postgres: writer process
> 1747 ? S 0:00 postgres: stats buffer process
> 1748 ? S 0:00 postgres: stats collector process
> 2106 pts/1 S 0:00 su postgres
> 2120 pts/1 S+ 0:00 psql postgres
> 2188 ? S 0:04 postgres: postgres dummy1 [local] VACUUM waiting
> 2200 pts/3 S 0:00 su postgres
> 2215 ? S 0:00 postgres: postgres dummy2 [local] idle in
> transaction
> 2717 pts/2 R+ 0:00 grep postgres
You might want to provide some more details on what you're doing.
Obviously, the simplified version of your problem doesn't exist (as
demonstrated by the fact that I can't reproduce it). Perhaps your
transaction is doing something different that what you expect.
-Bill
Oops! c@cornelia-boenigk.de (Cornelia Boenigk) was seen spray-painting on a wall: > Hi all > > If I have a running transaction in database1 and try to vacuum > database2 but the dead tuples in database2 cannot be removed. > > INFO: vacuuming "public.dummy1" > INFO: "dummy1": found 0 removable, 140000 nonremovable row versions > in 1341 pages > DETAIL: 135000 dead row versions cannot be removed yet. > > How can I achieve that database2 is vacuumed while a transaction in > database1 is not yet commited? You can't, unless you're on 8.1, and the not-yet-committed transaction is VACUUM. You have discovered a known factor, that a transaction left open on one database may have adverse effects on another database. -- (format nil "~S@~S" "cbbrowne" "acm.org") http://linuxdatabases.info/info/rdbms.html Signs of a Klingon Programmer #4: "A TRUE Klingon Warrior does not comment his code!"
In response to Christopher Browne <cbbrowne@acm.org>: > Oops! c@cornelia-boenigk.de (Cornelia Boenigk) was seen spray-painting on a wall: > > Hi all > > > > If I have a running transaction in database1 and try to vacuum > > database2 but the dead tuples in database2 cannot be removed. > > > > INFO: vacuuming "public.dummy1" > > INFO: "dummy1": found 0 removable, 140000 nonremovable row versions > > in 1341 pages > > DETAIL: 135000 dead row versions cannot be removed yet. > > > > How can I achieve that database2 is vacuumed while a transaction in > > database1 is not yet commited? > > You can't, unless you're on 8.1, and the not-yet-committed transaction > is VACUUM. I'm a little confused. First off, it would seem as if this is completely eliminated in 8.2, as I tested a scenario involving an idle transaction in one database, and both vacuum and vacuum full were able to complete in another database without completing the first transaction. Are you saying that in 8.1, there is a single exception to this, which is that if db1 (for example) is in the process of running vacuum, it won't block db2 from vacuuming? But that any other type of transaction can block operations in other databases? -- Bill Moran Collaborative Fusion Inc.
Bill Moran wrote: > In response to Christopher Browne <cbbrowne@acm.org>: > > Oops! c@cornelia-boenigk.de (Cornelia Boenigk) was seen spray-painting on a wall: > > > Hi all > > > > > > If I have a running transaction in database1 and try to vacuum > > > database2 but the dead tuples in database2 cannot be removed. > > > > > > INFO: vacuuming "public.dummy1" > > > INFO: "dummy1": found 0 removable, 140000 nonremovable row versions > > > in 1341 pages > > > DETAIL: 135000 dead row versions cannot be removed yet. > > > > > > How can I achieve that database2 is vacuumed while a transaction in > > > database1 is not yet commited? > > > > You can't, unless you're on 8.1, and the not-yet-committed transaction > > is VACUUM. > > I'm a little confused. > > First off, it would seem as if this is completely eliminated in 8.2, as > I tested a scenario involving an idle transaction in one database, and > both vacuum and vacuum full were able to complete in another database > without completing the first transaction. Of course they are able to complete, but the point is that they would not remove the tuples that would be visible to that idle open transaction. > Are you saying that in 8.1, there is a single exception to this, which > is that if db1 (for example) is in the process of running vacuum, it > won't block db2 from vacuuming? But that any other type of transaction > can block operations in other databases? In 8.2, a process running lazy vacuum (but not vacuum full) will not interfere with another process running vacuum, i.e., the second vacuum will be able to remove the tuples even if they would be seen by the transaction doing the first vacuum -- regardless of the database to which any of them is connected (i.e., it may be the same database or different databases). I don't remember if this was in 8.1 or was introduced in 8.2. -- Alvaro Herrera http://www.CommandPrompt.com/ PostgreSQL Replication, Consulting, Custom Development, 24x7 support
In response to Alvaro Herrera <alvherre@commandprompt.com>: > Bill Moran wrote: > > In response to Christopher Browne <cbbrowne@acm.org>: > > > Oops! c@cornelia-boenigk.de (Cornelia Boenigk) was seen spray-painting on a wall: > > > > Hi all > > > > > > > > If I have a running transaction in database1 and try to vacuum > > > > database2 but the dead tuples in database2 cannot be removed. > > > > > > > > INFO: vacuuming "public.dummy1" > > > > INFO: "dummy1": found 0 removable, 140000 nonremovable row versions > > > > in 1341 pages > > > > DETAIL: 135000 dead row versions cannot be removed yet. > > > > > > > > How can I achieve that database2 is vacuumed while a transaction in > > > > database1 is not yet commited? > > > > > > You can't, unless you're on 8.1, and the not-yet-committed transaction > > > is VACUUM. > > > > I'm a little confused. > > > > First off, it would seem as if this is completely eliminated in 8.2, as > > I tested a scenario involving an idle transaction in one database, and > > both vacuum and vacuum full were able to complete in another database > > without completing the first transaction. > > Of course they are able to complete, but the point is that they would > not remove the tuples that would be visible to that idle open > transaction. I would expect that, but the OP claimed that vacuum full waited until the other transaction was finished. > > Are you saying that in 8.1, there is a single exception to this, which > > is that if db1 (for example) is in the process of running vacuum, it > > won't block db2 from vacuuming? But that any other type of transaction > > can block operations in other databases? > > In 8.2, a process running lazy vacuum (but not vacuum full) will not > interfere with another process running vacuum, i.e., the second vacuum > will be able to remove the tuples even if they would be seen by the > transaction doing the first vacuum -- regardless of the database to > which any of them is connected (i.e., it may be the same database or > different databases). I don't remember if this was in 8.1 or was > introduced in 8.2. So lazy vacuum never waits on transactions. Apparently (based on the OP) vacuum full _does_ wait on transactions in versions prior to 8.2, but based on my experiment, in 8.2 vacuum full no longer does. Of course, in any version, vacuum can't clean up tuples held by open transactions. At least, that's what it's looking like to me. -- Bill Moran Collaborative Fusion Inc. wmoran@collaborativefusion.com Phone: 412-422-3463x4023 **************************************************************** IMPORTANT: This message contains confidential information and is intended only for the individual named. If the reader of this message is not an intended recipient (or the individual responsible for the delivery of this message to an intended recipient), please be advised that any re-use, dissemination, distribution or copying of this message is prohibited. Please notify the sender immediately by e-mail if you have received this e-mail by mistake and delete this e-mail from your system. E-mail transmission cannot be guaranteed to be secure or error-free as information could be intercepted, corrupted, lost, destroyed, arrive late or incomplete, or contain viruses. The sender therefore does not accept liability for any errors or omissions in the contents of this message, which arise as a result of e-mail transmission. ****************************************************************
Bill Moran <wmoran@collaborativefusion.com> writes:
> In response to Alvaro Herrera <alvherre@commandprompt.com>:
>> Of course they are able to complete, but the point is that they would
>> not remove the tuples that would be visible to that idle open
>> transaction.
> I would expect that, but the OP claimed that vacuum full waited until
> the other transaction was finished.
No, she didn't claim that. As far as I see she was just complaining
about the failure to remove dead tuples:
> > > > If I have a running transaction in database1 and try to vacuum
> > > > database2 but the dead tuples in database2 cannot be removed.
regards, tom lane
On fim, 2006-12-07 at 11:28 -0500, Tom Lane wrote: > Bill Moran <wmoran@collaborativefusion.com> writes: > > In response to Alvaro Herrera <alvherre@commandprompt.com>: > >> Of course they are able to complete, but the point is that they would > >> not remove the tuples that would be visible to that idle open > >> transaction. > > > I would expect that, but the OP claimed that vacuum full waited until > > the other transaction was finished. > > No, she didn't claim that. As far as I see she was just complaining > about the failure to remove dead tuples: > > > > > > If I have a running transaction in database1 and try to vacuum > > > > > database2 but the dead tuples in database2 cannot be removed. well actually, there was also this: On fim, 2006-12-07 at 00:57 +0100, Cornelia Boenigk wrote: > Hi Bill > > > Can you run a "vacuum > > full", and does it reclaim the space? > > I tried but it hangs. and also this: On fim, 2006-12-07 at 01:03 +0100, Cornelia Boenigk wrote: > > as soon as I committed the open transaction the hangig vacuum full > completed and the table was vacuumed: gnari
In response to Tom Lane <tgl@sss.pgh.pa.us>: > Bill Moran <wmoran@collaborativefusion.com> writes: > > In response to Alvaro Herrera <alvherre@commandprompt.com>: > >> Of course they are able to complete, but the point is that they would > >> not remove the tuples that would be visible to that idle open > >> transaction. > > > I would expect that, but the OP claimed that vacuum full waited until > > the other transaction was finished. > > No, she didn't claim that. As far as I see she was just complaining > about the failure to remove dead tuples: > > > > > > If I have a running transaction in database1 and try to vacuum > > > > > database2 but the dead tuples in database2 cannot be removed. Yes, but a later mail in the thread read: > > Can you run a "vacuum > > full", and does it reclaim the space? > > I tried but it hangs. > > [root@conni ~]# ps axw|grep postgres > 1746 ? S 0:00 postgres: writer process > 1747 ? S 0:00 postgres: stats buffer process > 1748 ? S 0:00 postgres: stats collector process > 2106 pts/1 S 0:00 su postgres > 2120 pts/1 S+ 0:00 psql postgres > 2188 ? S 0:04 postgres: postgres dummy1 [local] VACUUM waiting > 2200 pts/3 S 0:00 su postgres > 2215 ? S 0:00 postgres: postgres dummy2 [local] idle in transaction > 2717 pts/2 R+ 0:00 grep postgres Admittedly, I had the (incorrect) idea that she might need a vacuum full to reclaim space that lazy vacuum couldn't. And, admittedly, this wasn't the point of the original post. -- Bill Moran Collaborative Fusion Inc.
Bill Moran <wmoran@collaborativefusion.com> writes:
>> [root@conni ~]# ps axw|grep postgres
>> 1746 ? S 0:00 postgres: writer process
>> 1747 ? S 0:00 postgres: stats buffer process
>> 1748 ? S 0:00 postgres: stats collector process
>> 2106 pts/1 S 0:00 su postgres
>> 2120 pts/1 S+ 0:00 psql postgres
>> 2188 ? S 0:04 postgres: postgres dummy1 [local] VACUUM waiting
>> 2200 pts/3 S 0:00 su postgres
>> 2215 ? S 0:00 postgres: postgres dummy2 [local] idle in transaction
>> 2717 pts/2 R+ 0:00 grep postgres
Too bad this wasn't accompanied by a dump of pg_locks ... but if that's
the only other open transaction, the only way I can see for it to block
the vacuum is if the vacuum was database-wide, and had gotten to the
point of trying to vacuum one of the shared catalogs (eg, pg_database),
and the other transaction had some type of lock on that shared catalog.
regards, tom lane
Hi
Sorry, i was out
The first try was:
create database dummy1;
create table dummy ... and filled with 500 records
create database dummy2;
create table dummy ... and filled with 500 records
connecting to dummy1, opening a transaction and issued an update
begin;
update dummy set f1='achterbahn';
then opened a second console and connected to dummy2:
dummy2=# select count(*) from dummy;
count
-------
5000
(1 row)
dummy2=# SELECT relpages, reltuples FROM pg_class WHERE relname ='dummy';
relpages | reltuples
----------+-----------
160 | 5000
(1 row)
updated the table several times - to generate dead tuples:
dummy2=# SELECT relpages, reltuples FROM pg_class WHERE relname ='dummy';
relpages | reltuples
----------+-----------
326 | 30000
(1 row)
dummy2=# vacuum;
VACUUM
dummy2=# SELECT relpages, reltuples FROM pg_class WHERE relname ='dummy';
relpages | reltuples
----------+-----------
326 | 30000
(1 row)
dummy2=# select count(*) from dummy;
count
-------
5000
(1 row)
dummy2=# vacuum full;
--------------------------------
vacuum was in waiting state as long the transaction in dummy1 was
opened. After committing the transaction the vacuum full was carried out.
-------------------------------
VACUUM
dummy2=# SELECT relpages, reltuples FROM pg_class WHERE relname ='dummy';
relpages | reltuples
----------+-----------
326 | 30000
(1 row)
running on pg 8.1.4 on Fedora 5
Thanks
Conni
On fim, 2006-12-07 at 20:04 +0100, Cornelia Boenigk wrote: > Sorry, i was out > > [ snip demonstration of blocked vacuum full] > running on pg 8.1.4 on Fedora 5 could not duplicate this. can you show us the contents of pg_locks and pg_stat_activity while the VACUUM is blocked? gnari
Hi Ragnar > could not duplicate this. I also cannot reproduce the hanging VACUUM FULL. The problem remains thet the dead tuples cannot be vemoved. dummy1=# vacuum full; VACUUM dummy1=# SELECT relpages, reltuples FROM pg_class WHERE relname ='dummy1'; -[ RECORD 1 ]----- relpages | 997 reltuples | 100000 dummy1=# analyze verbose; ... INFO: analyzing "public.dummy1" INFO: "dummy1": scanned 997 of 997 pages, containing 5000 live rows and 95000 dead rows; 3000 rows in sample, 5000 estimatedtotal rows ... dummy1=# select * from pg_stat_activity; -[ RECORD 1 ]-+------------------------------ datid | 21529 datname | dummy1 procpid | 2065 usesysid | 10 usename | postgres current_query | <command string not enabled> query_start | backend_start | 2006-12-07 21:03:54.877779+01 client_addr | client_port | -1 -[ RECORD 2 ]-+------------------------------ datid | 21530 datname | dummy2 procpid | 2152 usesysid | 10 usename | postgres current_query | <command string not enabled> query_start | backend_start | 2006-12-07 21:07:59.973477+01 client_addr | client_port | -1 the transaction in db dummy2 performed an update and select count(*) and is still running. dummy1=# select * from pg_locks; -[ RECORD 1 ]-+----------------- locktype | relation database | 21530 relation | 21540 page | tuple | transactionid | classid | objid | objsubid | transaction | 85385 pid | 2152 mode | AccessShareLock granted | t -[ RECORD 2 ]-+----------------- locktype | relation database | 21530 relation | 21540 page | tuple | transactionid | classid | objid | objsubid | transaction | 85385 pid | 2152 mode | RowExclusiveLock granted | t -[ RECORD 3 ]-+----------------- locktype | relation database | 21529 relation | 10342 page | tuple | transactionid | classid | objid | objsubid | transaction | 85925 pid | 2065 mode | AccessShareLock granted | t -[ RECORD 4 ]-+----------------- locktype | transactionid database | relation | page | tuple | transactionid | 85925 classid | objid | objsubid | transaction | 85925 pid | 2065 mode | ExclusiveLock granted | t -[ RECORD 5 ]-+----------------- locktype | transactionid database | relation | page | tuple | transactionid | 85385 classid | objid | objsubid | transaction | 85385 pid | 2152 mode | ExclusiveLock granted | t Thanks Conni
Cornelia Boenigk wrote: > Hi Ragnar > >> could not duplicate this. > > I also cannot reproduce the hanging VACUUM FULL. The problem remains > thet the dead tuples cannot be vemoved. > [snip] I am interested in this. As one database cannot talk to another database in a transactional way a long running transaction in one database should not effect the vacuuming of another database. From my limited understanding VACUUM takes the lowest open transaction number and only cleans up transactions with TID's lower than that. The reason I believe that it has to use cluster wide is because the shared catalogs might be effected. Do shared catalogs follow MVCC or ACID strictly? I don't know, but I assume they don't follow both given my reading of the list. So if shared catalogs are the problem, what happens if you just vacuum the relevant table public.dummy1 and not the whole database, does the vacuum remove all the tuples that are dead? Is it possible to add logic for lazy vacuum that takes the lowest TID in our database when not vacuuming shared catalogs? This may already be the case, I don't know. Just putting forward suggestions. Russell Smith