Column name atttibutes lost in row query

[emarsden]

pgeltestdb=> SELECT 42 as z, 'bob' as bob;
 z  | ?column? 
----+----------
 42 | bob

The second column name should be bob instead of ?column?.

Using the :latest Docker image, which reports a server_version=15.17.

[zachmu]

This is a little bit tricky to resolve completely due to the subtlety of how postgres handles result column names. We're brainstorming to think if there could be an easy answer, but we aren't sure. Can you give us an idea of the severity of this issue on your end?

[emarsden]

No particular use case for this; I'm just reporting differences from PostgreSQL behaviour following the (perhaps slightly hubristic) response to #2545, rejecting the idea of signalling to clients that this is a PostgreSQL-compatible database rather than PostgreSQL.

[zachmu]

Thanks, appreciate the bug report and explanation.

I think we're open to signaling version info to clients to exactly the degree that it preserves compatibility with the existing constellation of postgres libraries. Our experience so far has been that any deviation in behavior can wreck compatibility with a particular framework or driver. But if you have some experience that suggests otherwise on this point in particular we're happy to listen. Our main goal is to be a drop-in replacement for any existing Postgres tool. If we can achieve that goal while vending more useful version / product info we are definitely interested.

[emarsden]

My perspective is as developer of an implementation of the PostgreSQL wire protocol (https://github.com/emarsden/pg-el) that supports around 20 variants of PostgreSQL (like CockroachDB, CrateDB, Yugabyte, Risingwave, Picodata). They are all detected using the content of the version() string, or specific ParameterStatus items sent during startup. For some variants, the connection initialization process has to be slightly different (e.g. you can't SET the DateStyle on Clickhouse).

As I mentioned in #2545, pretty much all variants report their brand, and mostly also their implementation version, in the version string, so you're unlikely to trigger bad behaviour from clients.

[zachmu]

That's very interesting context, thanks for that. I've reopened the original closed issue, so let's continue this conversation there. We'll leave this bug about result column naming open.