Citus compatibility: replace jsonb_build_object with immutable functions when immutable_expr_error? is true

[stevebrambilla]

The previous PR (#633) added the IMMUTABLE ash_raise_error_immutable function and made raising errors immutable as long as the error payload was pre-encoded in the app.

If the payload contains any expression values, it takes a different code path that uses jsonb_build_object, which is again not-IMMUTABLE and incompatible with Citus.

This PR changes the way the json error payload is built (only when the repo opts-in to using the immutable_expr_error? callback).

It adds a new IMMUTABLE function, ash_to_jsonb_immutable, which pins GUCs that affect json encoding and wraps to_jsonb. Then uses jsonb_set (already IMMUTABLE) calls to add expression values to the error payload.

This is the last step of the ash_raise_error/citus compatibility stuff I've been working on.

I added this extension to the TestRepo but it only opts-in for the new tests added here.

Contributor checklist

Leave anything that you believe does not apply unchecked.

• I accept the AI Policy, or AI was not used in the creation of this PR.
• Bug fixes include regression tests
• Chores
• Documentation changes
• Features include unit/acceptance tests
• Refactoring
• Update dependencies

[zachdaniel]

This is interesting 🤔 we can't use something like jsonb_build_object? We have to build a big nested "set" expression?

[stevebrambilla]

Yeah... this was the cleanest solution I found that still kept everything immutable. It's only the expression values that add to the nesting, and in Ash core it basically maxes out at 2, so in practice I don't think the nesting should be too deep.

Ideally I would have been able to write a GUC-pinned wrapper around jsonb_build_object, but postgres doesn't allow creating sql functions that accept a VARIADIC "any" argument, which is what's needed to forward to jsonb_build_object, it's only allowed for C extensions.

Open to other ideas if you have any, though!

[zachdaniel]

I see, thanks for the explanation 🙇

[zachdaniel]

The coalesce here seems strange to me. This would make this differ from to_jsonb. Can this function not return NULL if to_jsonb(value) is NULL?

[stevebrambilla]

This is because jsonb_set is different than jsonb_build_object in that you must pass in a valid jsonb value, otherwise the whole expression resolves to NULL. jsonb_build_object is a bit nicer here in that it accepts any and converts database NULLs to json nulls for you.

Maybe I should rename the function to remove that to_jsonb association though, could change it to ash_jsonb_build_value_immutable?

[zachdaniel]

🤔 I think it's fine. ash_..._immutable makes it pretty clear we're doing a specific thing here, and the longer name doesn't really make it any clearer IMO.

[zachdaniel]

🚀 Thank you for your contribution! 🚀