> > 3) What about the other DML operations (DELETE/UPDATE)? > > > > The other DML operations could probably benefit from the batching too. > > INSERT was good enough for a PoC, but having batching only for INSERT > > seems somewhat asmymetric. DELETE/UPDATE seem more complicated because > > of quals, but likely doable. > > Bulk INSERTs are more common in a sharded environment because of data > load in say OLAP systems. Bulk update/delete are rare, although not > that rare. So if an approach just supports bulk insert and not bulk > UPDATE/DELETE that will address a large number of usecases IMO. But if > we can make everything work together that would be good as well.
In most cases, I think the entire UPDATE/DELETE operations would be pushed down to the remote side by DirectModify. So, I'm not sure we really need the bulk UPDATE/DELETE.
That may not be true for a partitioned table whose partitions are foreign tables. Esp. given the work that Amit Langote is doing [1]. It really depends on the ability of postgres_fdw to detect that the DML modifying each of the partitions can be pushed down. That may not come easily.
> > 3) Should we do batching for COPY insteads? > > > > While looking at multi_insert, I've realized it's mostly exactly what > > the new "batching insert" API function would need to be. But it's only > > really used in COPY, so I wonder if we should just abandon the idea of > > batching INSERTs and do batching COPY for FDW tables.
> I think we have find out which performs > better COPY or batch INSERT.
Maybe I'm missing something, but I think the COPY patch [1] seems more promising to me, because 1) it would not get the remote side's planner and executor involved, and 2) the data would be loaded more efficiently by multi-insert on the demote side. (Yeah, COPY doesn't support RETURNING, but it's rare that RETURNING is needed in a bulk load, as you mentioned.)