On Sat, Apr 17, 2021 at 12:01 PM Amit Kapila <amit.kapil...@gmail.com> wrote: > > On Fri, Apr 16, 2021 at 11:24 PM Andres Freund <and...@anarazel.de> wrote: > > > > > > > I think it is also important to *not* acquire any lock on relation > > > otherwise it can lead to some sort of deadlock or infinite wait in the > > > decoding process. Consider a case for operations like Truncate (or if > > > the user has acquired an exclusive lock on the relation in some other > > > way say via Lock command) which acquires an exclusive lock on > > > relation, it won't get replicated in synchronous mode (when > > > synchronous_standby_name is configured). The truncate operation will > > > wait for the transaction to be replicated to the subscriber and the > > > decoding process will wait for the Truncate operation to finish. > > > > However, this cannot be really relied upon for catalog tables. An output > > function might acquire locks or such. But for those we do not need to > > decode contents... > > > > I see that if we define a user_catalog_table (create table t1_cat(c1 > int) WITH(user_catalog_table = true);), we are able to decode > operations like (insert, truncate) on such a table. What do you mean > by "But for those we do not need to decode contents"? >
I think we are allowed to decode the operations on user catalog tables because we are using RelationIsLogicallyLogged() instead of RelationIsAccessibleInLogicalDecoding() in ReorderBufferProcessTXN(). Based on this discussion, I think we should not be allowing decoding of operations on user catalog tables, so we should use RelationIsAccessibleInLogicalDecoding to skip such ops in ReorderBufferProcessTXN(). Am, I missing something? Can you please clarify? -- With Regards, Amit Kapila.
