Some of the ideas regarding vacuum improvements were discussed here:
A recent thread was started by Robert Haas, but I don't know if we logically
concluded that either.
This was once again brought up by Robert Haas in a discussion with Tom and
me during the PGCon and we agreed there are few things we can do make
vacuum more performant. One of the things that Tom mentioned is that the
vacuum today is not aware of the fact that its a periodic operation and
there might be ways to utilize that in some way.
The biggest gripe today is that vacuum needs two heap scans and each scan
dirties the buffer. While visibility map ensures that not-all blocks are
read and written during the scan, for a very large table, even a small
percentage of blocks can be significant. Further, post-HOT, the second scan
of the heap does not really reclaim any significant space, except for dead
line pointers. So there is a good reason to avoid that. I wanted to start a
discussion just about that. I am proposing one solution below, but I am not
married to the idea.
So the idea is to separate the index vacuum (removing index pointers to dead
tuples) from the heap vacuum. When we do heap vacuum (either by HOT-pruning
or using regular vacuum), we can spool the dead line pointers somewhere. To
avoid any hot-spots during normal processing, the spooling can be done
periodically like the stats collection. One obvious choice for spooling dead
line pointers is to use a relation fork. The index vacuum will be kicked off
periodically depending on the number of spooled deal line pointers. When
that happens, the index vacuum will remove all index pointers pointing to
those dead line pointers and forget the spooled line pointers.
The dead line pointers themselves will be removed whenever a heap page is
later vacuumed, either as part of HOT pruning or the next heap vacuum. We
would need some mechanism though to know that the index pointers to the
existing dead line pointers have been vacuumed and its safe to remove them
now. May be we can track the last operation that generated a dead line
pointer in the page using a LSN in the page header and also keep track of
the LSN of the last successful index vacuum. If the index vacuum LSN is
greater than the page header vacuum LSN, we can safely remove the existing
dead line pointers. I am deliberately not suggesting how to track the index
vacuum LSN since my last proposal to do something similar through a pg_class
column was shot down by Tom :-)
In nutshell, what I am suggesting is to do heap and index vacuuming
independently. The heap will be vacuumed either by HOT pruning or a periodic
heap vacuum and the dead line pointers will be collected. An index vacuum
will remove the index pointers to those dead line pointers. And at some
later point, the dead line pointers will be removed, either as part of
retail or complete heap vacuum. Its not clear if its useful, but a single
index vacuum can follow multiple heap vacuums or vice versa.
Another advantage of this technique would be that we can then support
start/stop heap vacuum or vacuuming a range of blocks at a time or even
vacuuming only those blocks which are already cached in the buffer cache.
Just a hand-waving at this point, but seems possible.
Suggestions/comments/criticism all welcome, but please don't shoot down the
idea on implementation details since I have really not spent time on that,
so it will be easy find holes and corner cases. That can be worked out if we
believe something like this will be useful.
pgsql-hackers by date
|Next:||From: Noah Misch||Date: 2011-05-24 09:07:34|
|Subject: Re: Reducing overhead of frequent table locks|
|Previous:||From: Vaibhav Kaushal||Date: 2011-05-24 06:19:27|
|Subject: Re: Foreign memory context read|