Why doesn't Postgres apply limit on groups when retrieving N results per group using WHERE IN + ORDER BY

On Mon, Feb 5, 2024 at 7:23 AM Sean v <sean@vanmulligen.ca> wrote:

This is related to a question I asked on dbs.stackexchange.com:
https://dba.stackexchange.com/questions/335501/why-doesnt-postgres-apply-limit-on-groups-when-retrieving-n-results-per-group

But to reiterate - I have a query like this:

SELECT "orders".*

FROM "orders"

WHERE (user_id IN ?, ?, ?)

ORDER BY "orders"."created_at" LIMIT 50

[snip]

So my question is twofold:
- why doesn't Postgres use the composite index, and then retrieve only the
minimum necessary amount of rows (50 per user) using the query I posted
above?

But your query *does not* list the first 50 rows *per user*. It only
returns the first 50 rows of:

SELECT "orders".*

FROM "orders"

WHERE (user_id IN ?, ?, ?)

ORDER BY "orders"."created_at"

Who knows which users are going to be in that list???

On Mon, Feb 5, 2024 at 8:55 AM Ron Johnson <ronljohnsonjr@gmail.com> wrote:

Who knows which users are going to be in that list???

It doesn't matter. Worse case scenario there is only one user in the
result and so all 50 rows are their earliest 50 rows. The system will thus
never need more than the earliest 50 rows per user to answer this question.

David J.

Exactly. I'm really just trying to understand if there's some functional
limitation to it being able to do that with how it executes these types
of queries, or if its just an optimization that hasn't been built into
the query planner yet.

I know I can get it to do precisely this if I use a CROSS JOIN LATERAL:

|SELECT o.* FROM company_users cu CROSS JOIN LATERAL ( SELECT * FROM
orders o WHERE o.user_id = company_users.user_id ORDER BY created_at
DESC LIMIT 50 ) cu WHERE cu.company_id = ? ORDER BY created_at DESC
LIMIT 50 |

That makes sense to me, it forces a nested loop and executes for each
user. But doing a nested select like the query below doesn't use the
index or limit the results to 50 per user - even though it does a nested
loop just like the lateral join does:

|SELECT "orders".* FROM "orders" WHERE user_id IN (SELECT user_id FROM
company_users WHERE company_id = ?) ORDER BY "orders"."created_at" LIMIT 50 |

El mié, 7 feb 2024 8:07, Sean v <sean@vanmulligen.ca> escribió:

Exactly. I'm really just trying to understand if there's some functional
limitation to it being able to do that with how it executes these types of
queries, or if its just an optimization that hasn't been built into the
query planner yet.

I know I can get it to do precisely this if I use a CROSS JOIN LATERAL:

SELECT o.*FROM company_users cuCROSS JOIN LATERAL (
SELECT *
FROM orders o
WHERE o.user_id = company_users.user_id
ORDER BY created_at DESC LIMIT 50
) cuWHERE cu.company_id = ? ORDER BY created_at DESC LIMIT 50

That makes sense to me, it forces a nested loop and executes for each
user. But doing a nested select like the query below doesn't use the index
or limit the results to 50 per user - even though it does a nested loop
just like the lateral join does:

SELECT "orders".* FROM "orders" WHERE user_id IN (SELECT user_id FROM company_users WHERE company_id = ?)ORDER BY "orders"."created_at" LIMIT 50

Joins will generally query the whole tables, leading to long run times.
Have you tried to preselect the rows of interest with a "WITH ... SELECT
..." query to reduce the amount of data processed?

On 2024-02-05 7:58 a.m., David G. Johnston wrote:

On Mon, Feb 5, 2024 at 8:55 AM Ron Johnson <ronljohnsonjr@gmail.com>
wrote:

Who knows which users are going to be in that list???

It doesn't matter. Worse case scenario there is only one user in the
result and so all 50 rows are their earliest 50 rows. The system will thus
never need more than the earliest 50 rows per user to answer this question.

David J.

Cheers
Olivier

On Tue, 6 Feb 2024 at 01:23, Sean v <sean@vanmulligen.ca> wrote:

SELECT "orders".*
FROM "orders"
WHERE (user_id IN ?, ?, ?)
ORDER BY "orders"."created_at" LIMIT 50

I have two indexes - `(user_id)` and `(user_id, created_at)`. Only the first index is ever used with this query.

I imagined that it would be clever enough to determine that only 50 results are needed, and that it could use the `(user_id, created_at)` index to get 50 orders for each user. Then sort and filter those few hundred results in memory.

It's as simple as the planner currently does not consider fetching 50
rows per user and doing a final sort before applying an overall LIMIT
50.

I have found that I can speed this up significantly using CROSS JOIN LATERAL and it will use the composite index, but I'm struggling to understand WHY the CROSS JOIN LATERAL is needed here for it to use the index.

I'm afraid that's the best workaround until someone submits a patch to
have the planner consider doing this optimisation automatically.

I've tried tweaking costs, disabling bitmap scans, etc, so it seems like this is a functional limitation rather than something to do with cost/statistics.

No amount of that will get you the plan you want without the LATERAL JOIN query.

So my question is twofold:
- why doesn't Postgres use the composite index, and then retrieve only the minimum necessary amount of rows (50 per user) using the query I posted above?

- If it is a functional limitation, is it lack of implementation, or is there a deeper incompatibility with how the query planner works that would prevent it from being able to do this?

Likely it wouldn't be too difficult to make the planner consider this
optimisation. However, for it to be valid, the ORDER BY clause would
have to contain only columns from the column(s) on the left side of
the IN clause. I think likely this could be done by having the
planner consider performing a Nested Loop with an outer VALUES scan
and an inner parameterized index scan with a Limit node above it as a
path for scanning the base relation. The tricky part would be
adjusting the planner so it didn't needlessly leave the IN clause in
the WHERE clause when the chosen plan is the Nested Loop with the
values scan. The current planner data structures are not really
geared up for optional base quals right now. Something would need to
be done to make that work and off the top of my head, I don't know
what that would be.

David