On 08/23/2016 01:00 PM, Craig James wrote:
> How do you create a foreign key that references a partitioned table?
>
> I'm splitting a large table "molecules" into 20 partitions, which also
> has an associated "molecular_properties" table. It looks something like
> this (pseudo-code):
>
> create table molecules(molecule_id integer primary key,
> molecule_data text,
> p integer);
>
> foreach $p (0..19) {
> create table molecules_$p (check(p = $p)) inherits (molecules);
> }
>
> create table molecular_properties(molprops_id integer primary key,
> molecule_id integer,
> molecular_weight numeric(8,3));
> alter table molecular_properties
> add constraint fk_molecular_properties
> foreign key(molecule_id)
> references molecules(molecule_id);
>
>
> (NB: There is no natural way to partition molecules, so the value for p
> is a random number. There is a good reason for partitioning that's not
> relevant to my question...)
>
> When I try to insert something into the molecular_properties table it fails:
>
> insert or update on table "molecular_properties" violates foreign
> key constraint "fk_molecular_properties"
> DETAIL: Key (molecule_id)=(83147) is not present in table "molecules".
>
>
> This surprised me. Obviously ID isn't in the "molecules" parent table,
> but I guessed that the foreign key would work anyway since the parent
> table is supposed to behave as though it includes all of the child tables.
I would say it is because of this:
https://www.postgresql.org/docs/9.5/static/sql-createtable.html
" Notes
...
Unique constraints and primary keys are not inherited in the current
implementation. This makes the combination of inheritance and unique
constraints rather dysfunctional.
...
"
>
> So how do you create a foreign key on a partitioned table?
>
> I suppose I could partition the molecular_properties table, but that
> would add unnecessary complication to the schema for no reason other
> than the "on delete cascade" feature.
>
> The only other thing I can think of is a delete trigger on each of the
> partition child tables. That would work, but it's a nuisance.
>
> Thanks,
> Craig
--
Adrian Klaver
adrian.klaver@aklaver.com
