Alternative implementation for JSONCodec.PlanScan

[felix-roehrich]

The fix from #2151 just pushed the original problem from #1470 and #1691 further along. This should be a comprehensive solution; also resolves #2229.

[felix-roehrich]

I will add a test to prevent regression from #2229 and look at failing tests tomorrow.

[felix-roehrich]

@jackc Looking at the following examples, I believe that the current scanning behavior is in inconsistent.

type Foo struct{}

var dst *Foo
err = conn.QueryRow(context.Background(), `select 'null'::jsonb`).Scan(&dst)
// err == nil
// dst == nil

var dst *Foo
err = conn.QueryRow(context.Background(), `select null::jsonb`).Scan(&dst)
// err == nil
// dst == nil

var dst Foo
err = conn.QueryRow(context.Background(), `select 'null'::jsonb`).Scan(&dst)
// err == nil
// dst == Foo{}

var dst Foo
err = conn.QueryRow(context.Background(), `select null::jsonb`).Scan(&dst)
// err == can't scan into dest[0] (col: jsonb): cannot scan NULL into *pgx_test.Foo
// dst == Foo{}

I believe either both cases of scanning Foo should error or none (or change scanning of *Foo). What is your opinion?

[felix-roehrich]

I decided to revert #2151 partially. This is because this should never be necessary, as such cases would be handled by the PointerPointerScanPlan. JSON requires special care to handle the scanning of 'null'::jsonb.

[jackc]

Thanks! sql.Scanner especially plus json has been such a pain. Hopefully, this finally resolves it.