How to aggregate (min/max etc.) over Django JSONField data?

I'm using Django 1.9 with its built-in JSONField and Postgres 9.4. In my model's attrs json field I store objects with some values, including numbers. And I need to aggregate over them to find min/max values. Something like this:

Model.objects.aggregate(min=Min('attrs__my_key'))

Also, it would be useful to extract specific keys:

Model.objects.values_list('attrs__my_key', flat=True)

The above queries fail with

FieldError: "Cannot resolve keyword 'my_key' into field. Join on 'attrs' not permitted."

Is it possible somehow?

Notes:

I know how to make a plain Postgres query to do the job, but am searching specifically for an ORM solution to have the ability to filter etc.
I suppose this can be done with a (relatively) new query expressions/lookups API, but I haven't studied it yet.

Solution

For those who interested, I've found the solution (or workaround at least).

from django.db.models.expressions import RawSQL

Model.objects.annotate(
    val=RawSQL("((attrs->>%s)::numeric)", (json_field_key,))
).aggregate(min=Min('val'))

Note that attrs->>%s expression will become something like attrs->>'width' after processing (I mean single quotes). So if you hardcode this name you should remember to insert them or you will get error.

A little bit off-topic

And one more tricky issue not related to Django itself but that is needed to be handled somehow. As attrs is a JSON field and there are no restrictions on its keys and values, you can (depending on you application logic) get some non-numeric values, for example in a width key. In this case you will get DataError from Postgres as a result of executing the above query. NULL values will be ignored meanwhile so it's OK. If you can just catch the error then no problem, you're lucky. In my case I needed to ignore wrong values and the only way here is to write a custom Postgres function that will suppress casting errors.

create or replace function safe_cast_to_numeric(text) returns numeric as $$
begin
    return cast($1 as numeric);
exception
    when invalid_text_representation then
        return null;
end;
$$ language plpgsql immutable;

And then use it to cast text to numbers:

Model.objects.annotate(
    val=RawSQL("safe_cast_to_numeric(attrs->>%s)", (json_field_key,))
).aggregate(min=Min('val'))

Thus we get quite a solid solution for such a dynamic thing as JSON.