Re: Suspicious behaviour on applying XLOG_HEAP2_VISIBLE.

On 26 April 2016 at 04:48, Andres Freund <andres@anarazel.de> wrote:
> No, I think we got to do this in all branches. I was just wondering
> about how to fix vm_extend(). Which I do think we got to fix, even in
> the back-branches.
>

I think replacing CacheInvalidateSmgr() with CacheInvalidateRelcache()
in vm_extend() is probably the safer thing to do, and ought to be
relatively harmless.

It means that an index-only scan won't be notified of VM extension
until the end of the other transaction, which might lead to extra heap
fetches, but I think that's unlikely to have any performance impact
because it ought to be a fairly rare event, and if it was another
transaction adding tuples, they wouldn't be all visible before it was
committed anyway, so the extra heap fetches would be required in any
case.

Regards,
Dean