On 12/30/2010 02:02 AM, Greg Smith wrote:
> Marko Tiikkaja wrote:
>> I have no idea why it worked in the past, but the patch was never
>> designed to work for UPSERT. This has been discussed in the past and
>> some people thought that that's not a huge deal.
>
> It takes an excessively large lock when doing UPSERT, which means its
> performance under a heavy concurrent load can't be good. The idea is
> that if the syntax and general implementation issues can get sorted
> out, fixing the locking can be a separate performance improvement to
> be implemented later. Using MERGE for UPSERT is the #1 use case for
> this feature by a gigantic margin. If that doesn't do what's
> expected, the whole implementation doesn't provide the community
> anything really worth talking about. That's why I keep hammering on
> this particular area in all my testing.
>
> One of the reflexive "I can't switch to PostgreSQL easily" stopping
> points for MySQL users is "I can't convert my ON DUPLICATE KEY UPDATE
> code". Every other use for MERGE is a helpful side-effect of adding
> the implementation in my mind, but not the primary driver of why this
> is important. My hints in this direction before didn't get adopted,
> so I'm saying it outright now: this patch must have an UPSERT
> implementation in its regression tests. And the first thing I'm going
> to do every time a new rev comes in is try and break it with the
> pgbench test I attached. If Boxuan can start doing that as part of
> his own testing, I think development here might start moving forward
> faster. I don't care so much about the rate at which concurrent
> UPSERT-style MERGE happens, so long as it doesn't crash. But that's
> where this has been stuck at for a while now.
I strongly agree. It *is* a huge deal.
cheers
andrew