Fix logical replication's ideas about which type OIDs are built-in.
authorTom Lane <tgl@sss.pgh.pa.us>
Mon, 13 May 2019 21:23:00 +0000 (17:23 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Mon, 13 May 2019 21:23:00 +0000 (17:23 -0400)
Only hand-assigned type OIDs should be presumed to match across different
PG servers; those assigned during genbki.pl or during initdb are likely
to change due to addition or removal of unrelated objects.

This means that the cutoff should be FirstGenbkiObjectId (in HEAD)
or FirstBootstrapObjectId (before that), not FirstNormalObjectId.
Compare postgres_fdw's is_builtin() test.

It's likely that this error has no observable consequence in a
normally-functioning system, since ATM the only affected type OIDs are
system catalog rowtypes and information_schema types, which would not
typically be interesting for logical replication.  But you could
probably break it if you tried hard, so back-patch.

Discussion: https://postgr.es/m/15150.1557257111@sss.pgh.pa.us

src/backend/replication/logical/relation.c
src/backend/replication/pgoutput/pgoutput.c

index 1f20df5680e5b351b778f441186d614174033b98..905ca12fd80f5d3ea4f95ca570385297bbc76648 100644 (file)
@@ -425,7 +425,7 @@ logicalrep_typmap_gettypname(Oid remoteid)
    bool        found;
 
    /* Internal types are mapped directly. */
-   if (remoteid < FirstNormalObjectId)
+   if (remoteid < FirstBootstrapObjectId)
    {
        if (!get_typisdefined(remoteid))
        {
index 86e0951a700e1c3737c6098cc5bc5e143505db4d..dc07420fa10fb20485051a9dad837a40fd400f9d 100644 (file)
@@ -269,8 +269,12 @@ maybe_send_schema(LogicalDecodingContext *ctx,
        desc = RelationGetDescr(relation);
 
        /*
-        * Write out type info if needed. We do that only for user created
-        * types.
+        * Write out type info if needed.  We do that only for user-created
+        * types.  We use FirstBootstrapObjectId as the cutoff, so that we only
+        * consider objects with hand-assigned OIDs to be "built in", not for
+        * instance any function or type defined in the information_schema.
+        * This is important because only hand-assigned OIDs can be expected
+        * to remain stable across major versions.
         */
        for (i = 0; i < desc->natts; i++)
        {
@@ -279,7 +283,7 @@ maybe_send_schema(LogicalDecodingContext *ctx,
            if (att->attisdropped)
                continue;
 
-           if (att->atttypid < FirstNormalObjectId)
+           if (att->atttypid < FirstBootstrapObjectId)
                continue;
 
            OutputPluginPrepareWrite(ctx, false);