Re: Fwd: Clarification about HOT

On 11/5/07, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>
> "Gokulakannan Somasundaram" <gokul007(at)gmail(dot)com> writes:
> > Currently, during Vacuum, we goto the Index and mark it as dead and
> > reclaim the space. For doing this, we are acquiring a Super-Exclusive
> lock.
> > After this implementation, we would update the index tuple instead of
> > marking it for cleanup. What can be foreseen as a locking overhead here?
>
> There are three operations involved: marking the child tuple not-HOT,
> updating the index entry(s), and removing the parent tuple. I doubt you
> can do them safely as independent atomic operations; I think you'll need
> to have more than one page locked at a time.

I think, in my suggestion, there are only two steps, redirect the index
tuple tid to point to a new tuple and mark the parent heap tuple vacuumed.
This is no different from the existing Vacuum process, except that we are
updating the index tuple, instead of marking it for Vacuum . I think you are
thinking about the current implementation. We need not mark any tuple as HOT
in the proposed one. Instead our code will work under the assumption that
normal indexes won't get updated, unless the indexing info changes. i don't
foresee a requirement for HOT classification. Please explain me, if it is
required in any case.

The locking problem also
> applies to trying to collapse out a dead HOT tuple that's in the middle
> of the chain:

Currently, we are maintaining a list/array in the tid_reaped or
lazy_tid_Reaped function. So this will contain all the tids that can be
Vacuumed. In our case,as already explained, all the heap tuples, which are
updated, will contain an extra tid, which is taken from it t_ctid stored in
t_data. So after we update the tid in the index-tuple, we can call
tid_reaped again to check whether the replaced tid has to get replaced / can
be Vacuumed

if it's the end of a cross-page link then you need two
> pages super-exclusive-locked in order to do that.

If the chain ends, the heap tuple would have got marked deleted. it won't
point to any tuple in the t_data and would have got marked HEAP_XMAX_VALID.
So it is the same case as Vacuuming a normal tuple. Have i correctly
answered your question??

There's also the little problem that a redirect line pointer doesn't
> have room for a cross-page link,

i don't know, what is a re-direct line pointer. The tid part of index tuple
can be made to point to any place. So i don't understand what you mean over
here.

and the big problem that having to
> chase across multiple pages for an index lookup would almost certainly
> negate any performance gains you might get. (In particular it'd
> completely destroy locality of access for bitmap indexscans...)

Yes, i think bitmap index scans will get affected, but only for the select
on heavily updated tuples. That's the cost we need to pay for extracting the
performance out of updates. As i already said the normal index would be the
index for heavily updated tables and thick index would be the index for
tables with heavy selects.

We can also look at optimizing the index tuple structure to reduce its
footprint.

Please get back, if i am not clear in any.
--
Thanks,
Gokul.
CertoSQL Project,
Allied Solution Group.
(www.alliedgroups.com)