[...] depends on what you value in a particular situation, latency or throughput. --DD
cursors are optimized for minimal cost of first row, queries are optimized for minimal cost of last row
That's a nice way to put it Pavel.
And to have it both ways, use COPY in binary protocol? That way the rows are streamed
to you in arbitrary chunks as soon as available (I hope), and the burden is on you the
client to decode and use those rows in parallel as they are "streamed" to you.
I've yet to test that (thus the 'i hope' above). I used COPY binary for INSERTs,
and COPY text/json for SELECTs, not yet COPY binary for SELECTs. I'm hoping
the latency of COPY will be small compared to a regular SELECT where I have to
wait for LIBPQ to assemble the whole ResultSet. Are my hopes unfounded? --DD
COPY is a different creature - it has no execution plan, and it is not interpreted by the executor.
Using COPY SELECT instead SELECT looks like premature optimization. The performance benefit will be minimal (maybe there can be exceptions depending on data, network properties or interface). Cursors, queries can use binary protocol, if the client can support it.
