Alvaro Herrera <alvherre@commandprompt.com> writes:
> Steve Crawford wrote:
>> Am I reading that what it actually does is to thrash around keeping
>> indexes unnecessarily updated, bloating them in the process?
> Yes.
Just for the record, it's not "unnecessary". The point of that is to
not leave a corrupted table behind if VACUUM FULL fails midway through.
The algorithm is:
1. copy tuples to lower blocks, inserting index entries for them too
During this stage, if we fail then the copied tuples are invalid (since
they were inserted by a failed transaction) and so no corruption.
Meanwhile the original tuples are marked as "moved by this vacuum
transaction", but their validity is not affected by that.
2. mark the transaction committed
This atomically causes all the copied tuples to be GOOD and all the
originals to be INVALID according to the tuple validity rules.
3. remove the index entries for moved-off tuples
If we crash here, some of the invalid tuples will have index entries
and some won't, but that doesn't matter because they're invalid.
(The next vacuum will take care of finishing the cleanup.)
4. remove the moved-off tuples (which just requires truncating the
table)
I don't see a way to remove the old index entries before inserting new
ones without creating a window where the index and table will be
inconsistent if vacuum fails.
CLUSTER avoids all this thrashing by recopying the whole table, but
of course that has peak space requirements approximately twice the
table size (and is probably not a win anyway unless most of the table
rows need to be moved). You pays your money, you takes your choice.
regards, tom lane
