PostgreSQL partial unique index and upsert

Question

I'm trying to do an upsert to a table that has partial unique indexes

create table test (
    p text not null,
    q text,
    r text,
    txt text,
    unique(p,q,r)
);

create unique index test_p_idx on test(p) where q is null and r is null;
create unique index test_pq_idx on test(p, q) where r IS NULL;
create unique index test_pr_idx on test(p, r) where q is NULL;

In plain terms, p is not null and only one of q or r can be null.

Duplicate inserts throw constraint violations as expected

insert into test(p,q,r,txt) values ('p',null,null,'a'); -- violates test_p_idx
insert into test(p,q,r,txt) values ('p','q',null,'b'); -- violates test_pq_idx
insert into test(p,q,r,txt) values ('p',null, 'r','c'); -- violates test_pr_idx

However, when I'm trying to use the unique constraint for an upsert

insert into test as u (p,q,r,txt) values ('p',null,'r','d') 
on conflict (p, q, r) do update 
set txt = excluded.txt

it still throws the constraint violation

ERROR: duplicate key value violates unique constraint "test_pr_idx" DETAIL: Key (p, r)=(p, r) already exists.

But I'd expect the on conflict clause to catch it and do the update.

What am I doing wrong? Should I be using an index_predicate?

index_predicate Used to allow inference of partial unique indexes. Any indexes that satisfy the predicate (which need not actually be partial indexes) can be inferred. Follows CREATE INDEX format. https://www.postgresql.org/docs/9.5/static/sql-insert.html

Answer 1

I don't think it's possible to use multiple partial indexes as a conflict target. You should try to achieve the desired behaviour using a single index. The only way I can see is to use a unique index on expressions:

drop table if exists test;
create table test (
    p text not null,
    q text,
    r text,
    txt text
);

create unique index test_unique_idx on test (p, coalesce(q, ''), coalesce(r, ''));

Now all three tests (executed twice) violate the same index:

insert into test(p,q,r,txt) values ('p',null,null,'a'); -- violates test_unique_idx
insert into test(p,q,r,txt) values ('p','q',null,'b');  -- violates test_unique_idx
insert into test(p,q,r,txt) values ('p',null, 'r','c'); -- violates test_unique_idx

In the insert command you should pass the expressions used in the index definition:

insert into test as u (p,q,r,txt) 
values ('p',null,'r','d') 
on conflict (p, coalesce(q, ''), coalesce(r, '')) do update 
set txt = excluded.txt;