|From:||Alvaro Herrera <alvherre(at)2ndquadrant(dot)com>|
|To:||Heikki Linnakangas <hlinnakangas(at)vmware(dot)com>|
|Cc:||Robert Haas <robertmhaas(at)gmail(dot)com>, Pg Hackers <pgsql-hackers(at)postgresql(dot)org>|
|Subject:||Re: Minmax indexes|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
Heikki Linnakangas wrote:
> On 06/23/2014 08:07 PM, Alvaro Herrera wrote:
> I feel that the below would nevertheless be simpler:
> >>I wonder if it would be simpler to just always store the revmap
> >>pages in the beginning of the index, before any other pages. Finding
> >>the revmap page would then be just as easy as with a separate fork.
> >>When the table/index is extended so that a new revmap page is
> >>needed, move the existing page at that block out of the way. Locking
> >>needs some consideration, but I think it would be feasible and
> >>simpler than you have now.
> >Moving index items around is not easy, because you'd have to adjust the
> >revmap to rewrite the item pointers.
> Hmm. Two alternative schemes come to mind:
> 1. Move each index tuple off the page individually, updating the
> revmap while you do it, until the page is empty. Updating the revmap
> for a single index tuple isn't difficult; you have to do it anyway
> when an index tuple is replaced. (MMTuples don't contain a heap
> block number ATM, but IMHO they should, see below)
> 2. Store the new block number of the page that you moved out of the
> way in the revmap page, and leave the revmap pointers unchanged. The
> revmap pointers can be updated later, lazily.
> Both of those seem pretty straightforward.
The trouble I have with moving blocks around to make space, is that it
would cause the index to have periodic hiccups to make room for the new
revmap pages. One nice property that these indexes are supposed to have
is that the effect into insertion times should be pretty minimal. That
would cease to be the case if we have to do your proposed block moves.
> >>ISTM that when the old tuple cannot be updated in-place, the new
> >>index tuple is inserted with mm_doinsert(), but the old tuple is
> >>never deleted.
> >It's deleted by the next vacuum.
> Ah I see. Vacuum reads the whole index, and builds an in-memory hash
> table that contains an ItemPointerData for every tuple in the index.
> Doesn't that require a lot of memory, for a large index? That might
> be acceptable - you ought to have plenty of RAM if you're pushing
> around multi-terabyte tables - but it would nevertheless be nice to
> not have a hard requirement for something as essential as vacuum.
I guess if you're expecting that pages_per_range=1 is a common case,
yeah it might become an issue eventually. One idea I just had is to
have a bit for each index tuple, which is set whenever the revmap no
longer points to it. That way, vacuuming is much easier: just scan the
index and delete all tuples having that bit set. No need for this hash
table stuff. I am still concerned with adding more overhead whenever a
page range is modified, so that insertions in the table continue to be
fast. If we're going to dirty the index every time, it might not be so
fast anymore. But then maybe I'm worrying about nothing; I will have to
measure how slower it is.
> Wouldn't it be simpler to remove the old tuple atomically with
> inserting the new tuple and updating the revmap? Or at least mark
> the old tuple as deletable, so that vacuum can just delete it,
> without building the large hash table to determine that it's
Yes, it might be simpler, but it'd require dirtying more pages on
insertions (and holding more page-level locks, for longer. Not good for
> I'm quite surprised by the use of LockTuple on the index tuples. I
> think the main reason for needing that is the fact that MMTuple
> doesn't store the heap (range) block number that the tuple points
> to: LockTuple is required to ensure that the tuple doesn't go away
> while a scan is following a pointer from the revmap to it. If the
> MMTuple contained the BlockNumber, a scan could check that and go
> back to the revmap if it doesn't match. Alternatively, you could
> keep the revmap page locked when you follow a pointer to the regular
> index page.
There's the intention that these accesses be kept as concurrent as
possible; this is why we don't want to block the whole page. Locking
individual TIDs is fine in this case (which is not in SELECT FOR UPDATE)
because we can only lock a single tuple in any one index scan, so
there's no unbounded growth of the lock table.
I prefer not to have BlockNumbers in index tuples, because that would
make them larger for not much gain. That data would mostly be
redundant, and would be necessary only for vacuuming.
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
|Next Message||Peter Geoghegan||2014-07-09 21:44:09||Re: Doing better at HINTing an appropriate column within errorMissingColumn()|
|Previous Message||Alvaro Herrera||2014-07-09 21:19:46||Re: Doing better at HINTing an appropriate column within errorMissingColumn()|