Обсуждение: Is there a way to make VACUUM run completely outside transaction
Hi As VACUUM is not something that can be rolled back, could we not make it run completely outside transactions. It already needs to be run outside a transaction block. I try to explain the problem more thoroughly below (I'm quite sleepy, so the explanation may be not too clear ;) My problem is as follows: I have a fairly write intensive database that has two kinds of tables - 1) some with a small working set (thousands of rows), but their get constant traffic of hundreds of inserts/updates/deletes per second. 2) and some with bigger tables (a few million rows) which get smaller load of updates. I keep the first type of tables filesize small/dead tuples count low by running vacuum on them in a tight loop with 15 sec sleeps in between vacuuming individual tables) And I keep the 2nd type balanced by running another loop of vacuums on them. It worked quite well for a while, but as the tables grow, the vacuums on the 2nd kind of tables run long enough for the first kind of tables to accumulate hundreds of thousands dead tuples, which can't be freed because the vacuums on 2nd kind run in their own long transactions, keeping the oldest active transaction id smaller than needed. And the new provisions of making VACUUM less intrusive by allowing delays in vacuuming make this problem worse, by kind of priority inverion - the less intrusive vacuum runs longer, thereby keeping its transaction open longer and thereby being *more* intrusive by not allowing old tuples to be deleted by another vacuum command. I can see two ways to solve this problem: 1) If possible, VACUUM command should close its transaction id early in the command processing, as it does not really need to be in transactions 2) the transaction of the VACUUM command could be assigned some really bug trx id, so it won't get in a way 3) VACUUM commits/or aborts its transaction after some predetermined interval (say 1 min) and starts a new one. 4) VACUUM itself could have some way to check if the "oldest" transaction is also VACUUM, and be able to free the tuples that are older than last *data altering* transaction. At least VACUUM, ANALYSE and TRUNCATE should not be considered *data altering* here. The problems caused by non-data-altering but long-running transactionsis are not unique to VACUUM, similar problems also affects Slony. So any of 1)-3) would be preferable to 4) Or perhaps it already solved in 8.0 ? My rant is about 7.4.6. -- Hannu Krosing <hannu@tm.ee>
Hannu Krosing <hannu@tm.ee> writes:
> As VACUUM is not something that can be rolled back, could we not make it
> run completely outside transactions.
No, because it has to be able to hold a table-level lock on the target
table. Besides, where did you get the idea that it can't be rolled back?
The VACUUM FULL case, at least, has to go through huge pushups to be
sure it is rollback-safe.
regards, tom lane
Ühel kenal päeval (esmaspäev, 7. veebruar 2005, 10:51-0500), kirjutas Tom Lane: > Hannu Krosing <hannu@tm.ee> writes: > > As VACUUM is not something that can be rolled back, could we not make it > > run completely outside transactions. > > No, because it has to be able to hold a table-level lock on the target > table. Does NON-FULL VACUUM release this lock when pausing when sleeping on vacuum_cost_delay ? If not, could it be made to release the lock and end current transaction when going to sleep and start a new transaction and aquire the lock again when waking up ? > Besides, where did you get the idea that it can't be rolled back? > The VACUUM FULL case, at least, has to go through huge pushups to be > sure it is rollback-safe. What are the semantics of ROLLBACKing a VACUUM FULL ? Will it restore the dead tuples to their original positions ? I assumed that it does its data-shuffling behind the scenes so that changes done by it are invisible to others at the *logical* level. If all the changes it does to internal data storage can be rolled back, then I can't see how VACUUM FULL can work at all without requiring 2x the filesize for the ROLLBACK. Also, why must it be run outside of transaction block if it can be rollbacked ? -- Hannu Krosing <hannu@tm.ee>
On Mon, Feb 07, 2005 at 07:16:41PM +0200, Hannu Krosing wrote: > If all the changes it does to internal data storage can be rolled back, > then I can't see how VACUUM FULL can work at all without requiring 2x > the filesize for the ROLLBACK. I think the point is that the table is still consistent if the system crashes while vacuum is running. For the internal details, see Xvac in the HeapTupleHeader struct. > Also, why must it be run outside of transaction block if it can be > rollbacked ? A vacuum actually uses several transactions, so it wouldn't work as the user would expect if run in a transaction. The first one is committed rather early and new ones are opened and closed. (One per table, IIRC.) -- Alvaro Herrera (<alvherre[@]dcc.uchile.cl>) "There was no reply" (Kernel Traffic)
Ühel kenal päeval (esmaspäev, 7. veebruar 2005, 19:01-0300), kirjutas Alvaro Herrera: > > Also, why must it be run outside of transaction block if it can be > > rollbacked ? > > A vacuum actually uses several transactions, so it wouldn't work as the > user would expect if run in a transaction. The first one is committed > rather early and new ones are opened and closed. (One per table, IIRC.) So I guess that making it commit and open new transaction at a regular interval (like each minute) during vacuuming single table would not alter its visible behaviour. That would solve my problem of long-running vacuums on large tables polluting unrelated small but heavily updated tables with dead tuples. I'll take a peak at code and try to come up with a naive proposal you can shoot down ;) -- Hannu Krosing <hannu@tm.ee>
On Tue, Feb 08, 2005 at 01:55:47PM +0200, Hannu Krosing wrote: > So I guess that making it commit and open new transaction at a regular > interval (like each minute) during vacuuming single table would not > alter its visible behaviour. That would solve my problem of long-running > vacuums on large tables polluting unrelated small but heavily updated > tables with dead tuples. Interesting. The problem is that a long running VACUUM on a single table will keep in PGPROC a TransactionId that will last very long, which will "pollute" every concurrent Snapshot; so smaller tables can't be cleaned up because the tuples are visible for the transaction running the vacuum -- except that that transaction cannot possibly want to look at them. -- Alvaro Herrera (<alvherre[@]dcc.uchile.cl>) "La tristeza es un muro entre dos jardines" (Khalil Gibran)
Ühel kenal päeval (teisipäev, 8. veebruar 2005, 13:39-0300), kirjutas Alvaro Herrera: > On Tue, Feb 08, 2005 at 01:55:47PM +0200, Hannu Krosing wrote: > > > So I guess that making it commit and open new transaction at a regular > > interval (like each minute) during vacuuming single table would not > > alter its visible behaviour. That would solve my problem of long-running > > vacuums on large tables polluting unrelated small but heavily updated > > tables with dead tuples. > > Interesting. The problem is that a long running VACUUM on a single > table will keep in PGPROC a TransactionId that will last very long, > which will "pollute" every concurrent Snapshot; so smaller tables can't > be cleaned up because the tuples are visible for the transaction running > the vacuum -- except that that transaction cannot possibly want to look > at them. Exactly. That's what I was trying to describe in my original post. -- Hannu Krosing <hannu@tm.ee>