Re: [HACKERS] Another nasty cache problem

Bruce Momjian <pgman@candle.pha.pa.us> writes:
>> Perhaps the caches shouldn't store ctid?  Not sure.

> I am guilt of that.  There are a few place where I grab the tuple from
> the cache, then use that to update the heap.  I thought it was a nifty
> solution at the time.  I thought I used the CacheCopy calls for that,
> but I am not positive.  Even if I did, that doesn't help because the
> copy probably has an invalid tid at that point, thought I have opened
> the table. Maybe I have to make sure I open the table before geting the
> tid from the cache.

I believe we worked that out and fixed it a few months ago: it's safe
to use the cache to find a tuple you want to update, if you open and
lock the containing table *before* doing the cache lookup.  Then you
know VACUUM's not running on that table (since you have it locked)
and you have an up-to-date TID for the tuple (since the open+lock
would have processed any pending shared-inval messages).  I went
around and made sure that's true everywhere.

What I was thinking about was adding code to the caches that would
(a) maintain refcounts on cached tuples, (b) reread rather than
discard a tuple if it is invalidated while refcount > 0, and (c)
kick out an error if the reread shows that the tuple has in fact
changed.  It seems that we would need to ignore the TID when deciding
if a tuple has changed, however.
        regards, tom lane