Though pg_shard correctly marks its metadata tables so they are included in the output of pg_dump, these dumps are unusable because the relation_id column of pg_shard's metadata tables is of type oid. Because there is no guarantee a given relation will have identical oids, the oids in the metadata tables will not match any relations, rendering the dump unusable.
I found an old thread that insinuates pg_dump will produce a textual representation of tables that works across a restore boundary if the column is of type regclass instead of oid. As far as lower-level code is concerned, these two types are identical (so we will not need to change any implementation). By switching to regclass we can give users the ability to rebuild a master in an identical fashion to restoring a normal PostgreSQL instance from backup.
Though
pg_shardcorrectly marks its metadata tables so they are included in the output ofpg_dump, these dumps are unusable because therelation_idcolumn ofpg_shard's metadata tables is of typeoid. Because there is no guarantee a given relation will have identicaloids, theoids in the metadata tables will not match any relations, rendering the dump unusable.I found an old thread that insinuates
pg_dumpwill produce a textual representation of tables that works across a restore boundary if the column is of typeregclassinstead ofoid. As far as lower-level code is concerned, these two types are identical (so we will not need to change any implementation). By switching toregclasswe can give users the ability to rebuild a master in an identical fashion to restoring a normal PostgreSQL instance from backup.