Fix MotherDuck multi-catalog configs

[naoyak]

Creating a MotherDuck connection config with the catalogs argument (multiple attached databases / data sources) rather than specifying database (single db) was failing because the MotherDuck connection was never authenticated.

The reason is because the motherduck_token has to be sent in the initial connection string when instantiating the connection factory (or otherwise made available via env var) - TIL that ATTACH md:{database_name}?motherduck_token={token_string} is actually not valid, so the MD token can't be passed in on a per-catalog basis. In any case the same token (=MD account) should be shared throughout the gateway config.

[naoyak]

Currently, the ATTACH commands for the individual MotherDuck databases included in catalogs will raise an error since they have been automatically attached by the catch all md: in the connection string.

We could either:

1. Ignore the errors (this way)
2. Skip ATTACHing the MD catalogs and assume they exist
3. Skip ATTACHing the catalogs but verify they are already attached and available

[naoyak]

removing because it doesn't actually work

[izeigerman]

Suggested change

	# If a user tried to attach a MotherDuck database/share which has already by attached via
	# If a user tried to attach a MotherDuck database/share which has already been attached via

[izeigerman]

Why is this not an error? Why would anyone specify the same MD catalog more than once?

[naoyak]

So the way the implementation for MotherDuckConnectionConfig works right now, the db adapter uses md: as the connection string for duckdb, implicitly adding all the databases/shares available to the token account. if you then issue an ATTACH command to a specific db md:some_db, duckdb will emit an error.

per my comment (#4001) above, we could either skip the ATTACH trusting that the db actually exists on the account, or we could check that it's there.

[izeigerman]

I see. This makes sense, thank you

[izeigerman]

Why do we block accessing other databases here?

[naoyak]

so if the motherduck connection is created with the database argument instead of catalogs, that implies we are only using a single MotherDuck catalog/db and there shouldn't be any reason to allow attaching another MD database (see docs). Do you think that makes sense?

[izeigerman]

I see. Is there any reason to be this limiting? Alternatively, I might suggest to ALWAYS connect with attach_model=single in which case users need to manually ATTACH catalogs and so we no longer need to check that error message. Or am I thinking about this wrong?

[naoyak]

so i guess I was understanding this functionality wrong. this doesn't block you from attaching other databases but rather forces you to invoke ATTACH 'md:some_other_db' if you end up wanting to access another catalog.

Alternatively, I might suggest to ALWAYS connect with attach_model=single in which case users need to manually ATTACH catalogs and so we no longer need to check that error message.

attach_mode=single is only available in the connection string passed when launching duckdb - when the connection is figured with database, the user should specify a catalog name, whereas with the catalogs: option we invoke duckdb with the string 'md:' so we don't have a catalog name to pass to the attach_mode=single option.

We could possibly:

• arbitrarily pick the first MD catalog under catalogs and pass it to the connection string
• or alternatively, launch duckdb with a blank connection string (i.e. no MD connection until we ATTACH the listed catalogs). maybe this makes the most sense?

[izeigerman]

Apologies about the delay.

or alternatively, launch duckdb with a blank connection string

This makes sense to me