On Wed, 20 Feb 2019 at 10:49, Tom Lane <tgl@sss.pgh.pa.us> wrote: > What if we dropped that idea, and instead defined the plan tree as > returning only the columns that are updated by SET, plus the row > identity? It would then be the ModifyTable node's job to fetch the > original tuple using the row identity (which it must do anyway) and > form the new tuple by combining the updated columns from the plan > output with the non-updated columns from the original tuple. > > DELETE would be even simpler, since it only needs the row identity > and nothing else.
While I didn't look at the patch in great detail, I think this is how Pavan must have made MERGE work for partitioned targets. I recall seeing the tableoid being added to the target list and a lookup of the ResultRelInfo by tableoid.
Maybe Pavan can provide more useful details than I can.
Yes, that's the approach I took in MERGE, primarily because of the hurdles I faced in handling partitioned tables, which take entirely different route for UPDATE/DELETE vs INSERT and in MERGE we had to do all three together. But the approach also showed significant performance improvements. UPDATE/DELETE via MERGE is far quicker as compared to regular UPDATE/DELETE when there are non-trivial number of partitions. That's also a reason why I recommended doing the same for regular UPDATE/DELETE, but that got lost in the MERGE discussions. So +1 for the approach.
We will need to consider how this affects EvalPlanQual which currently doesn't have to do anything special for partitioned tables. I solved that via tracking the expanded-at-the-bottom child in a separate mergeTargetRelation, but that approach has been criticised. May be Tom's idea doesn't have the same problem or most likely he will have a far better approach to address that.