Re: WIP patch for hint bit i/o mitigation

On Wed, Nov 7, 2012 at 6:01 AM, Amit Kapila <amit(dot)kapila(at)huawei(dot)com> wrote:
>> >> Following the sig is a first cut at a patch (written by Atri) that
>> >> attempts to mitigate hint bit i/o penalty when many pages worth of
>> >> tuples are sequentially written out with the same transaction id.
>> >> There have been other attempts to deal with this problem that fit
>> >> niche cases (especially those that create the table in the same
>> >> transaction as the one inserting) that work but don't really solve
>> the
>> >> problem generally.
>>
>> As we are now dealing with only the last xid(please refer to the details
>> of the patch attached), the invalidation issues are not significant any
>> more.
>
> I think you are right, but please check below the scenario I have in mind
> due to which I got confused:
>
> Session -1
> 1. let's say for tup-1 on page, cachedVisibilityXid is not set, and it go
> inside SetHinbits and set it to xid of tuple which is let's say = 708
> 2. now for all consecutive tuples which have same xmin (708), it can
> directly refer
> cached id and cached status, even if hint bit is not set.
>
> Other Sessions
> 3. now from other sessions, large operations happened on all other tables
> except this table.
> 4. The situation can reach where xid can wrap around.
>
> Session -1
> 5. It again tries to query the same table, at this point it will compare
> the xid in tuple with cachedVisibilityXid.
>
> Now if all tuple's xid's for that particular table are frozen in all cases
> then it seems to be okay, otherwise it might be problem.
> I am not fully aware about this wrap around and frozed xid concept, thats
> why I had doubted
> that it might create problem, on further thought, it appears that I was
> wrong.

Well there's that. But more to the point for the cache to fail you'd
have to have a scenario where a table didn't scan any records for 1
billion+ transactions. See note [3] above for reasoning why this is
implausible. Also we're already relying on this effect in transam.c.

merlin