issue_comments

i'll adopt this PR to make the changes @simonw suggested https://github.com/simonw/sqlite-utils/pull/203#issuecomment-753567932

Is there any progress elsewhere on the handling of compound / composite foreign keys, or is this PR still effectively open?

Thanks for looking at this - home schooling kids has prevented me from replying.

I'd struggled with how to adapt the API for the foreign keys too - I definitely tried the String/Tuple approach. I hadn't considered the breaking changes that would introduce though. I can take a look at this and try and make the change - see which of your options works best.

I've got a workaround for the use-case I was looking at this for, so it wouldn't be a problem for me if it was put on the back burner until a hypothetical v4.0 anyway.

Another option: expand the ForeignKey object to have .columns and .other_columns properties in addition to the existing .column and .other_column properties. These new plural properties would always return a tuple, which would be a one-item tuple for a non-compound-foreign-key.

The question then is what should .column and .other_column return for compound foreign keys?

I'd be inclined to say they should return None - which would trigger errors in code that encounters a compound foreign key for the first time, but those errors would at least be a strong indicator as to what had gone wrong.

We can label .column and .other_column as deprecated and then remove them in sqlite-utils 4.0.

Since this would still be a breaking change in some minor edge-cases I'm thinking maybe 4.0 needs to happen in order to land this feature. I'm not opposed to doing that, I was just hoping it might be avoidable.

One way that this could avoid a breaking change would be to have fk.column and fk.other_column remain as strings for non-compound-foreign-keys, but turn into tuples for a compound foreign key.

This is a bit of an ugly API design, and it could still break existing code that encounters a compound foreign key for the first time - but it would leave code working for the more common case of a non-compound-foreign-key.

Sorry for taking so long to review this!

This approach looks great to me - being able to optionally pass a tuple anywhere the API currently expects a column is smart, and it's consistent with how the pk= parameter works elsewhere.

There's just one problem I can see with this: the way it changes the ForeignKey(...) interface to always return a tuple for .column and .other_column, even if that tuple only contains a single item.

This represents a breaking change to the existing API - any code that expects ForeignKey.column to be a single string (which is any code that has been written against that) will break.

As such, I'd have to bump the major version of sqlite-utils to 4.0 in order to ship this.

Ideally I'd like to make this change in a way that doesn't represent an API compatibility break. I need to think a bit harder about how that might be achieved.

Sorry for not reviewing this yet! I'll try to carve out time to look at it in the next few days.