Re: [HACKERS] Avoid MVCC using exclusive lock possible?
Andrew Sullivan a écrit : On Sun, Feb 29, 2004 at 10:43:34AM -0500, Tom Lane wrote: general I think our VACUUM-based approach is superior to the Oracle-style UNDO approach, because it pushes the maintenance overhead out of foreground transaction processing and into a schedulable background process. Certainly any Oracle DBA will tell you that huge I completely agree with this. If the recent work on lowering the overall cost ov VACUUM on loaded systems pays off, then I think there can be no argument that the work-now, vacuum-later strategy is the best approach, simply because it deals with the outlying and unexpected cases better than the alternatives. I know too many people who have been burned by running out of rollback segments when some use pattern emerged that they hadn't planned for. A I agree too. The VACUUM aproach is better as it reduces fragmentation and chained rows due to columns size change. ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
On Tuesday 02 March 2004 06:29, Paul Tillotson wrote:
> However, for this to be useful, your table must not have any indexes,
> views, foreign keys, sequences, triggers, etc., or else you must be
> prepared to re-create all of them using application level code.
Which isn't a big deal is it? You can write a single function to create entire
object and it's dependency. It is one time job but can save lots of time at
runtime.
> I imagine this would break lots of things, but it would be nice if
> instead of Shridhar's rename step (see below) one could do this:
>
> $table1node = query("SELECT relfilenode FROM pg_class WHERE relname =
> '$old_table';");
> $table2node = query("SELECT relfilenode FROM pg_class WHERE relname =
> '$new_table';");
> exec("UPDATE pg_class SET relfilenode = $table2node WHERE relname =
> '$old_table';");
> exec("UPDATE pg_class SET relfilenode = $table1node WHERE relname =
> '$new_table';");
>
> You would of course need to change the relfilenode for all of the
> toasted columns and indexes as well in the same atomic step, but it
> seems like this might be more compatible with postgresql's MVCC model
> than other ideas suggested.
I still don't understand what is not so good about rename. All the indexes
remain there. Views need updation, I agree. Missed that last time. But what
you are suggesting is a part of rename if not complete of it.
I would always prefer to let PG handle these kind of details. Not very
comfortable with mucking around catalogs especially if there exists an
alternative.
Shridhar
---(end of broadcast)---
TIP 8: explain analyze is your friend
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
Paul Tillotson <[EMAIL PROTECTED]> writes: > I use this type of approach when mirroring data from a foxpro database > (yuck) to a read-only postgres database. It is quicker and cleaner than > deleting all of the rows and inserting them again (TRUNCATE is not > transaction safe, which I need). Just FYI, TRUNCATE *is* transaction safe as of 7.4. regards, tom lane ---(end of broadcast)--- TIP 6: Have you searched our list archives? http://archives.postgresql.org
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
I use this type of approach when mirroring data from a foxpro database
(yuck) to a read-only postgres database. It is quicker and cleaner than
deleting all of the rows and inserting them again (TRUNCATE is not
transaction safe, which I need).
However, for this to be useful, your table must not have any indexes,
views, foreign keys, sequences, triggers, etc., or else you must be
prepared to re-create all of them using application level code.
I imagine this would break lots of things, but it would be nice if
instead of Shridhar's rename step (see below) one could do this:
$table1node = query("SELECT relfilenode FROM pg_class WHERE relname =
'$old_table';");
$table2node = query("SELECT relfilenode FROM pg_class WHERE relname =
'$new_table';");
exec("UPDATE pg_class SET relfilenode = $table2node WHERE relname =
'$old_table';");
exec("UPDATE pg_class SET relfilenode = $table1node WHERE relname =
'$new_table';");
You would of course need to change the relfilenode for all of the
toasted columns and indexes as well in the same atomic step, but it
seems like this might be more compatible with postgresql's MVCC model
than other ideas suggested.
Regards,
Paul Tillotson
Shridhar Daithankar wrote:
I am sure people have answered the approach you have suggested so let me
suggest a workaround for your problem.
You could run following in a transaction.
- begin
- Create another table with exact same structure
- write a procedure that reads from input table and updates the value in
between
- drop the original table
- rename new table to old one
- commit
- analyze new table
Except for increased disk space, this approach has all the good things
postgresql offers. Especially using transactable DDLs it is huge benefit. You
certainly do save on vacuum.
If the entire table is updated then you can almost certainly get things done
faster this way.
HTH
Shridhar
---(end of broadcast)---
TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
---(end of broadcast)---
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddressHere" to [EMAIL PROTECTED])
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
>Shridhar Daithankar > > Recently, I ran a huge update on an Integer column affecting 100 million > > rows in my database. What happened was my disk space increased in size > and > > my IO load was very high. It appears that MVCC wants to rewrite each row > > (each row was about 5kB due to a bytea column). In addition, VACUUM > needs > > to run to recover space eating up even more IO bandwidth. > > I am sure people have answered the approach you have suggested so let me > suggest a workaround for your problem. > > You could run following in a transaction. > > - begin > - Create another table with exact same structure > - write a procedure that reads from input table and updates the value in > between > - drop the original table > - rename new table to old one > - commit > - analyze new table > > Except for increased disk space, this approach has all the good things > postgresql offers. Especially using transactable DDLs it is huge benefit. > You > certainly do save on vacuum. > > If the entire table is updated then you can almost certainly get things > done > faster this way. I second this approach. Space usage similar, time faster. For a procedure, you can use an Insert Select with a Case statement instead of the UPDATE step. This type of approach is in use in many other DBMS. I would vote strongly (how is that possible?) against doing anything at all to MVCC. It works, it's brilliant and doesn't fall foul of anybody else's patent rights. It would be very hard to re-write anything as good and the time spent would double other requirements. MySQL has multiple storage managers and that must slow them down enormously trying to test or even fine tune things. Best Regards, Simon Riggs ---(end of broadcast)--- TIP 3: if posting/reading through Usenet, please send an appropriate subscribe-nomail command to [EMAIL PROTECTED] so that your message can get through to the mailing list cleanly
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
On Sun, Feb 29, 2004 at 10:43:34AM -0500, Tom Lane wrote: > general I think our VACUUM-based approach is superior to the > Oracle-style UNDO approach, because it pushes the maintenance overhead > out of foreground transaction processing and into a schedulable > background process. Certainly any Oracle DBA will tell you that huge I completely agree with this. If the recent work on lowering the overall cost ov VACUUM on loaded systems pays off, then I think there can be no argument that the work-now, vacuum-later strategy is the best approach, simply because it deals with the outlying and unexpected cases better than the alternatives. I know too many people who have been burned by running out of rollback segments when some use pattern emerged that they hadn't planned for. A -- Andrew Sullivan ---(end of broadcast)--- TIP 4: Don't 'kill -9' the postmaster
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
On Tuesday 24 February 2004 22:13, Stephen wrote: > Hi, > > Recently, I ran a huge update on an Integer column affecting 100 million > rows in my database. What happened was my disk space increased in size and > my IO load was very high. It appears that MVCC wants to rewrite each row > (each row was about 5kB due to a bytea column). In addition, VACUUM needs > to run to recover space eating up even more IO bandwidth. I am sure people have answered the approach you have suggested so let me suggest a workaround for your problem. You could run following in a transaction. - begin - Create another table with exact same structure - write a procedure that reads from input table and updates the value in between - drop the original table - rename new table to old one - commit - analyze new table Except for increased disk space, this approach has all the good things postgresql offers. Especially using transactable DDLs it is huge benefit. You certainly do save on vacuum. If the entire table is updated then you can almost certainly get things done faster this way. HTH Shridhar ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
Neil Conway <[EMAIL PROTECTED]> writes: > Jon Jensen wrote: >> How would you do a rollback if an error occurred mid-query? How would you >> keep your table from being a big pile of junk if a power failure happened >> during the query? > As most non-MVCC database do: by writing WAL records. > In theory, it seems to me that we could implement an overwriting > storage manager for PostgreSQL (i.e. we'd have both MVCC and non-MVCC, > and allow users to choose at runtime). It would require a *lot* of > work, but I can't see any fundamental reason why it wouldn't be possible. It would be possible, but what's the point? There's no evidence that this approach is superior to what we do, and a fair number of reasons to think it's inferior. In particular, with this approach you have to maintain the entire history of a transaction in WAL, so that you have the info available to roll back if it aborts. So the OP's concern about needing 2X the disk space to update his table still applies, it's just in a different place. It's true that it's easier to get the space back when no longer needed --- VACUUM FULL is expensive if it's moving lots of records. But in general I think our VACUUM-based approach is superior to the Oracle-style UNDO approach, because it pushes the maintenance overhead out of foreground transaction processing and into a schedulable background process. Certainly any Oracle DBA will tell you that huge transactions are a headache to handle in Oracle. I think what the OP actually wanted us to offer was non-transaction-safe overwrite-in-place updating. That might be technically feasible (not sure how we'd deal with changes in indexed columns though). But it's so contrary to the basic design goals of the project that I don't think anyone will take the suggestion seriously. regards, tom lane ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
Jon Jensen wrote: How would you do a rollback if an error occurred mid-query? How would you keep your table from being a big pile of junk if a power failure happened during the query? As most non-MVCC database do: by writing WAL records. In theory, it seems to me that we could implement an overwriting storage manager for PostgreSQL (i.e. we'd have both MVCC and non-MVCC, and allow users to choose at runtime). It would require a *lot* of work, but I can't see any fundamental reason why it wouldn't be possible. -Neil ---(end of broadcast)--- TIP 7: don't forget to increase your free space map settings
Re: [HACKERS] Avoid MVCC using exclusive lock possible?
On Tue, 24 Feb 2004, Stephen wrote: > It came to my mind that what if there could be a mechanism in place to allow > overwriting portions of the same row *whenever possible* instead of creating > a new row as MVCC would require. How would you do a rollback if an error occurred mid-query? How would you keep your table from being a big pile of junk if a power failure happened during the query? Jon ---(end of broadcast)--- TIP 4: Don't 'kill -9' the postmaster
