Обсуждение: Why does it sort rows after a nested loop that uses already-sorted indexes?
Hi:
I've a question regarding nested loops and the order in which they return rows. Can you help me, please?
Suppose that I've two tables:
- Table [sales_order]
* Columns [id]
* Index [sales_order_pkey] on [id]
- Table [order_line]
* Columns [id], [sales_order_id]
* Index [order_line_ukey] on [sales_order_id], [id]
Then, I run the following query:
-----------------------------------------------------------
SELECT sales_order.id, order_line.id
FROM main.sales_order
JOIN main.order_line ON order_line.sales_order_id = sales_order.id
WHERE sales_order.customer_id = 2
ORDER BY sales_order.id, order_line.id;
-----------------------------------------------------------
The query planner decides to use the following nested loop:
-----------------------------------------------------------
Incremental Sort (cost=26.90..16020.06 rows=144955 width=8)
Sort Key: sales_order.id, order_line.id
Presorted Key: sales_order.id
-> Nested Loop (cost=0.70..4593.99 rows=144955 width=8)
-> Index Scan using sales_order_pkey on sales_order (cost=0.28..19.31 rows=79 width=4)
Filter: (customer_id = 2)
-> Index Only Scan using order_line_ukey on order_line (cost=0.42..39.22 rows=1869 width=8)
Index Cond: (sales_order_id = sales_order.id)
-----------------------------------------------------------
As you can see, the planner does detect that the outer loop returns the rows presorted by [sales_order.id]. However, it's unable to detect that the rows returned by the inner loop are also sorted by [sales_order.id] first, and then by [order_line.id].
Why is it? Is it because the planner is designed to always ignore the order of the inner loop, even although it could take advantage of it (for example, because the analysis time rarely is worth it)? Or is there something that I'm missing?
If I'm not mistaken, in this case both index scans seem to be done serially, in an N x M style, so I think the row order would be preserved, Right?
Thank you!
negora
negora <public@negora.com> writes:
> As you can see, the planner does detect that the outer loop returns the
> rows presorted by [sales_order.id]. However, it's unable to detect that
> the rows returned by the inner loop are also sorted by [sales_order.id]
> first, and then by [order_line.id].
That's a level of analysis that it doesn't do, and TBH I'm not even
entirely sure it's correct to assume that the output is sorted like
that. At minimum you'd need an additional assumption that the
outer side's join key is unique, which is a factor that we don't
currently track when reasoning about ordering.
regards, tom lane
> That's a level of analysis that it doesn't do... Great. I suspected that, but I needed a confirmation from a reliable source. Thank you! > ...and TBH I'm not even > entirely sure it's correct to assume that the output is sorted like > that. At minimum you'd need an additional assumption that the > outer side's join key is unique, which is a factor that we don't > currently track when reasoning about ordering. Ouch! I hadn't thought about that possibility! When I tried to mentally reproduce the nested loop, I always considered the values of the outer loop to be unique. I guess that was because, very often, I used unique indexes for my tests... But it doesn't have to be so, of course. Best regards. On 18/04/2024 16:53, Tom Lane wrote: > negora <public@negora.com> writes: >> As you can see, the planner does detect that the outer loop returns the >> rows presorted by [sales_order.id]. However, it's unable to detect that >> the rows returned by the inner loop are also sorted by [sales_order.id] >> first, and then by [order_line.id]. > > That's a level of analysis that it doesn't do, and TBH I'm not even > entirely sure it's correct to assume that the output is sorted like > that. At minimum you'd need an additional assumption that the > outer side's join key is unique, which is a factor that we don't > currently track when reasoning about ordering. > > regards, tom lane