BackendKeyData shouldn't be a required message during startup

[jordanlewis]

The py-postgresql library requires that a BackendKeyData or 'K' message be sent during connection initialization from Postgres, which leads to failure to connect to servers speaking the Postgres wire format that don't send such a message.

This message is not strictly required by the Postgres message flow rules, so this library should be able to tolerate not receiving such a message.

[jwp]

The message flow rules on that page are of questionable normative quality. That is, I'm not sure your interpretation is consistent with: "but in the normal case the backend will send some ParameterStatus messages, BackendKeyData, and finally ReadyForQuery." found in that page. However, if libpq is handling the absence of BackendKeyData without error, your interpretation may effectively be correct.

In any case, if you're seeking complete transparency, your server will need to send BackendKeyData in order to allow applications to perform query interrupts.

[jwp]

Finally got around to analyzing this further: https://github.com/postgres/postgres/blob/REL_13_1/src/interfaces/libpq/fe-protocol3.c#L265

It looks like libpq's treatment of BackendKeyData as optional might be an artifact of its implementation. Whether this behavior is intentional still seems ambiguous to me. There is no explicit branch during negotiation that allows us to clearly recognize that the message is intended to be optional. (Also, libpq's ability to eat subsequent 'K' messages looks suspicious to my eyes.)

Given that the canonical PQ server implementation unconditionally sends BackendKeyData, it seems reasonable that a client implementation might make the choice of expecting it unconditionally.

It is advisable to mimic the server's behavior in order to avoid possible incompatibilities with other implementations as well as the incompatibility with this driver.

I'm not closing this out as a statement that I'm unwilling to toss a branch in the negotiation code. I'm quite willing to, but I would like to see more evidence that it is intended to be optional as not sending BackendKeyData is a deviation from the behavior of the server.

[jordanlewis]

Fair enough. Thanks @jwp! cc @rafiss