Archives of the TeradataForum
Message Posted: Wed, 02 Sep 2009 @ 18:59:35 GMT
Subj: | | Re: All Amps usage |
|
From: | | Victor Sokovin |
I think this is as scalable as it gets in this case. For a fixed receiver, the rows are distributed across AMPs according to the sender.
Assuming the sender can be anybody, you'll need the all-AMP retrieval.
I don't think the skew should be a problem in this case either.
A theoretical threat to the plan I can think of is when the optimizer estimates that the receiver part will return a significant percentage of
rows it may ignore the NUSI and go for the full scan instead. But the percentage must be really high, like > 15% (or is it 10%?). Must be a really
major account or a major problem with the stats. This is done because in most cases the full scan is then indeed a better option under the
circumstances. It won't be sub-second but then you can't expect sub-second performance when returning 15% of a large table.
The worst case is the largest receiver. It will give the most work for the AMPs. Can you predict how big it will be in the future? If so,
create the test case and run the query. This is pretty much the worst performing case. I don't think the size of the table matters very much for
the test. You don't need to future-size it unless the 10-15% limit is exceeded.
Victor
|