Postgres already has "COPY FROM STDIN" command, which makes database server use postgres connection for the raw data. However, since it uses the existing connection, it needs to be compatible with Postgres protocol, which means that there is an extra overhead in wrapping streaming data.

On the other hand, "COPY FROM COMMAND" has no wrapping overhead, as it opens direct pipe to/from command, so no Postgres protocol parsing is involved - as only short command string is sent via postgres connection, while bulk data goes over dedicated channel. This makes it faster, although I am not sure how much does this actually save.

amluto's point was that one can achieve no-wrapping performance of "COPY FROM COMMAND" if one could pass dedicated data descriptor to "COPY FROM STDIN". This could be done using SCM_RIGHTS (a AF_UNIX protocol feature that passes descriptors between processes) or pipes+slice (not 100% sure how those would help). But with SCM_RIGHTS, you'd have your psql client create pipe, exec process, and then pass the descriptor to the sql server. This would have exactly the same speed as "COPY FROM COMMAND" (no overhead, dedicated pipe) but would not mix security contexts and would execute any code under server's username - overall better solution.

Your point was "only on localhost", which I interpreted as "this approach would only work if psql runs on localhost (because SCM_RIGHTS only works on localhost); while "COPY FROM COMMAND" could even be executed remotely".

This is 100% correct, but as I said, I think an ability to execute commands remotely as a server user is a bad idea and should have never existed.