Re: Add new protocol message to change GUCs for usage with future protocol-only GUCs

On Fri, 5 Apr 2024 at 18:30, Dave Cramer <davecramer(at)gmail(dot)com> wrote:
> On Fri, 5 Apr 2024 at 12:09, Jelte Fennema-Nio <me(at)jeltef(dot)nl> wrote:
>> I'll take a look at redesigning the protocol parameter stuff. To work
>> with dedicated functions instead.
>
> +1

It's been a while, but I now actually took the time to look into this.
And I ran into a problem that I'd like to get some feedback on before
continuing the implementation:

If we're not setting the protocol parameter in the StartupMessage,
there's currently no way for us to know if the protocol parameter is
supported by the server. If protocol parameters were unchangable then
that would be fine, but the whole point of introducing ParameterSet is
to make it possible to change protocol parameters on an existing
connection. Having the function SupportsProtocolCompression return
false, even though you can enable compression just fine, only because
we didn't ask for compression when connecting seems quite silly and
confusing.

I see five ways around this problem and would love some feedback on
which you think is best (or if you can think of any other/better
ones):
1. Have protocol parameters always be GUC_REPORT, so that the presence
of a ParameterStatus message during connection startup can be used as
a way of detecting support for the protocol parameter.
2. Make libpq always send each known protocol parameter in the
StartupMessage to check for their support, even if the connection
string does not contain the related parameters (set them to their
default value then). Then the non-presence of the parameter in the
NegotiateProtocolVersion message can be used reliably to determine
support for the feature. We could even disallow changing a protocol
parameter at the server side using ParameterSet if it was not
requested in the StartupMessage.
3. Very similar to 1, but require explicit user input in the
connection string to request the feature on connection startup by
having the user explicitly provide its default value. If it's not
requested on connection startup assume its unsupported and disallow
usage of the feature (even if the server might actually support it).
4. Make SupportsProtocolCompression return a tri-state, SUPPORTED,
UNSUPPORTED, UNKNOWN. If it's UNKNOWN people could send a ParameterSet
message themselves to check for feature support after connection
startup. We could even recognize this and change the state that
SupportProtocolCompression function to return SUPPORTED/UNSUPPORTED on
future calls according to the server response.
5. Basically the same as 4 but automatically send a ParameterSet
message internally when calling SupportsProtocolCompression and the
state is UNKNOWN, so we'd only ever return SUPPORTED or UNSUPPORTED.

The above options are listed in my order of preference, below some
reasoning why:

1 and 2 would increase the bandwidth used during connection handshake
slightly for each protocol parameter that we would add, but they have
the best user experience IMHO.

I slightly prefer 1 over 2 because there is another argument to be
made for always having protocol parameters be GUC_REPORT: these
parameters change what message types a client can send or receive. So
it makes sense to me to have the server tell the client what the
current value of such a parameter is. This might not be a strong
argument though, because this value would only ever change due to user
interaction. But still, one might imagine scenarios where the value
that the client sent is not exactly what the server would set the
parameter to on receiving that value from the client. e.g. for
protocol compression, maybe the client sends a list of prefered
compression methods and the server would send a ParameterStatus
containing only the specific compression method that it will use when
sending messages to the client.

3 seems also an acceptable option to me. While having slightly worse
user experience than 2, it allows the user of the client to make the
decision if the extra bandwidth during connection startup is worth it
to be able to enable the feature later.

4 assumes that we want people to be able to trigger sending
ParameterSet messages for every protocol parameter. I'm not sure we'd
want to give that ability in all cases.

5 would require SupportsProtocolCompression to also have a
non-blocking version, which bloats our API more than I'd like. Also as
a user you wouldn't be able to know if SupportsProtocolCompression
will do a network request or not.

PS. This is only a problem for feature detection for features relying
on protocol parameters, feature-support relying on protocol version
bumps are easy to detect based on the NegotiateProtocolVersion
message.