Re: [Proposal] Fully WAL logged CREATE DATABASE - No Checkpoints

Further, the code changes in the patch looks good. I just have few comments:

+void
+LockRelationId(LockRelId *relid, LOCKMODE lockmode)
+{
+   LOCKTAG     tag;
+   LOCALLOCK  *locallock;
+   LockAcquireResult res;
+
+   SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);

Should there be an assertion statement here to ensure that relid->dbid and  relid->relid is valid?

--

    if (info == XLOG_DBASE_CREATE)
    {
        xl_dbase_create_rec *xlrec = (xl_dbase_create_rec *) XLogRecGetData(record);
-       char       *src_path;
-       char       *dst_path;
-       struct stat st;
-
-       src_path = GetDatabasePath(xlrec->src_db_id, xlrec->src_tablespace_id);
-       dst_path = GetDatabasePath(xlrec->db_id, xlrec->tablespace_id);
+       char       *dbpath;

-       /*
-        * Our theory for replaying a CREATE is to forcibly drop the target
-        * subdirectory if present, then re-copy the source data. This may be
-        * more work than needed, but it is simple to implement.
-        */
-       if (stat(dst_path, &st) == 0 && S_ISDIR(st.st_mode))
-       {
-           if (!rmtree(dst_path, true))
-               /* If this failed, copydir() below is going to error. */
-               ereport(WARNING,
-                       (errmsg("some useless files may be left behind in old database directory \"%s\"",
-                               dst_path)));
-       }

I think this is a significant change and probably needs some kind of explanation/comments as-in why we are just creating a dir and copying the version file when replaying create database operation. Earlier, this meant replaying the complete create database operation, that doesn't seem to be the case now.

--

Have you intentionally skipped pg_internal.init file from being copied to the target database?