Hi,
for performance reasons it's a good idea to use the binary protocol. But
doing so between two postgres installations is made unnecessarily hard
by the choice of embedding and verifying oids in composite and array
types. When using extensions, even commonly used ones like hstore, the
datatype oids will often not be the same between systems, even when
using exactly the same version on the same OS.
Specifically I'm talking about
Datum
array_recv(PG_FUNCTION_ARGS)
{
...element_type = pq_getmsgint(buf, sizeof(Oid));if (element_type != spec_element_type){ /* XXX Can we allow taking
theinput element type in any cases? */ ereport(ERROR, (errcode(ERRCODE_DATATYPE_MISMATCH),
errmsg("wrongelement type")));}
...
}
and
Datum
record_recv(PG_FUNCTION_ARGS)
{
.../* Process each column */for (i = 0; i < ncolumns; i++)
... /* Verify column datatype */ coltypoid = pq_getmsgint(buf, sizeof(Oid)); if (coltypoid != column_type)
ereport(ERROR, (errcode(ERRCODE_DATATYPE_MISMATCH), errmsg("wrong data type: %u,
expected%u", coltypoid, column_type)));
...
}
given that we're giving up quite some speed and adding complexity to
make send/recv functions portable, this seems like a shame.
I'm failing to see what these checks are buying us? I mean the text
representation of a composite doesn't include type information about
contained columns either, I don't see why the binary version has to?
I'm basically thinking that we should remove the checks on the receiving
side, but leave the sending of oids in place for backward compat.
Greetings,
Andres Freund
