override previous O shared keys @ AllSelection & rowMode: 'array' processing @ PostgresDriver.

[igalklebanov]

Hey 👋

fixes #1094.

Running the following query:

const result = await ctx.db
  .selectFrom('person')
  .leftJoin('toy', 'person.id', 'toy.id')
  .selectAll('person')
  .selectAll('toy')
  .execute()

On the 4 core dialects in our tests,
yields the following values for the shared id column:

pg	mysql2	tedious	better-sqlite3
1	1	1	1
2	null	null	null
3	null	null	null

As you can see, pg's results take person.id values, and not toy.id values. Other dialects do the opposite.

To align pg results with the other dialects, this PR uses rowMode: 'array' in PostgresDriver to get the raw values and produce objects that take the last value of a column with the same name. This is a breaking change for PostgreSQL users who relied on the default behavior of pg.

pg	mysql2	tedious	better-sqlite3
1	1	1	1
null	null	null	null
null	null	null	null

This PR doesn't handle the .selectAll() AND .selectAll(['person', 'toy']) cases. Here we'll get a union of all possible values of a shared column which is not accurate, but can be $narrowType'd. .selectAll(['person', 'toy']) can be solved with compile-time array recursion (which afaik is not that performant - not worth it??). .selectAll() requires a lot of testing - probably affected by order of introduction of tables to query context - if so, will require a huge refactor to TB[] and basically most types.

We can recommend our users to use multiple single-table calls, as the tightest option.
We can deprecate the multi-table variant and even the empty variant of selectAll.

Should we allow users to switch back to pg's default row processing with a flag in the dialect config?

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
kysely	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Aug 5, 2024 1:25am