On Thu, Nov 11, 2021 at 1:44 PM Greg Nancarrow <gregn4...@gmail.com> wrote: > > On Thu, Nov 11, 2021 at 5:52 PM houzj.f...@fujitsu.com > <houzj.f...@fujitsu.com> wrote: > > > > When looking into how to fix the second issue, I have a question: > > > > After changing publish_via_partition_root from false to true, the > > subcriber will fetch the partitioned table from publisher when refreshing. > > > > In subsriber side, If all the child tables of the partitioned table already > > subscribed, then we can just skip the table sync for the partitioned table. > > But > > if only some of the child tables(not all child tables) were already > > subscribed, > > should we skip the partitioned table's table sync ? I am not sure about the > > appropriate behavior here. > > > > What do you think ? > > > > I'm not sure you can skip the partitioned table's table sync as you > are suggesting, because on the subscriber side, the tables are mapped > by name, so what is a partitioned table on the publisher side might > not be a partitioned table on the subscriber side (e.g. might be an > ordinary table; and similarly for the partitions) or it might be > partitioned differently to that on the publisher side. >
Sure, we don't know about that, or at least there is no such mapping that is recorded. So, I think we should skip it even if any one of the child table is present. > (I might be > wrong here, and I don't have a good solution, but I can see the > potential for inconsistent data resulting in this case, unless say, > the subscriber "child tables" are first truncated on the refresh, if > they are in fact partitions of the root, and then the table sync > publishes the existing data via the root) > Do you want to say current behavior for this case where data is copied twice is okay? I think we need to find a way to handle this and document it to set the expectations right. -- With Regards, Amit Kapila.
