Sounds like the problem you are having is, the server is running out of temporary resources for the operation that users are trying to do. So according to Tom, on the postgres side, the operation cannot be optimized further.
I think you have few choices here,
- See if increasing the resources of the server will allow them to run the operation
- Ask users not to do that operation
- Use a extension like citus to scale horizontally
But I'm thinking why a massively inefficient join is needed in the first place. Shouldn't joins be for following keyed relationships. So ideally a unique indexed column, but at the very least an indexed column. Why is a join required on a dynamically calculated substring? Can it be made into a static computed value and indexed? Substring sounds like an op that should be in the filter stage.
Can you describe your data model? Maybe we can give some specific advice.
Regards,
Avin
On Mon, Jun 26, 2023 at 3:57 AM Marc Millas <marc.millas@xxxxxxxxxx> wrote:
On Sun, Jun 25, 2023 at 11:48 PM Tom Lane <tgl@xxxxxxxxxxxxx> wrote:David Rowley <dgrowleyml@xxxxxxxxx> writes:
> The problem is that out of the 3 methods PostgreSQL uses to join
> tables, only 1 of them supports join conditions with an OR clause.
> Merge Join cannot do this because results can only be ordered one way
> at a time. Hash Join technically could do this, but it would require
> that it built multiple hash tables. Currently, it only builds one
> table. That leaves Nested Loop as the join method to implement joins
> with OR clauses. Unfortunately, nested loops are quadratic and the
> join condition must be evaluated once per each cartesian product row.
We can do better than that if the OR'd conditions are each amenable
to an index scan on one of the tables: then it can be a nestloop with
a bitmap-OR'd inner index scan. I thought the upthread advice to
convert the substr() condition into something that could be indexed
was on-point.ok. but one of the tables within the join(s) tables is 10 billions rows, splitted in 120 partitions. Creating something like 20 more indexes to fulfill that condition do have its own problems.
> Tom Lane did start some work [1] to allow the planner to convert some
> queries to use UNION instead of evaluating OR clauses, but, if I
> remember correctly, it didn't handle ORs in join conditions, though
> perhaps having it do that would be a natural phase 2. I don't recall
> why the work stopped.
As I recall, I was having difficulty convincing myself that
de-duplication of results (for cases where the same row satisfies
more than one of the OR'd conditions) would work correctly.
You can't just blindly make it a UNION because that might remove
identical rows that *should* appear more than once in the result.I did rewrite the query using a cte and union(s). For that query, no dedup point.But my pb is that that DB will be used by a bunch of people writing raw SQL queries, and I cannot let them write queries that are going to go on for ages, and eventually crash over temp_file_limit after hours every now and then.So, my understanding of the above is that I must inform the users NOT to use OR clauses into joins.which maybe a pb by itself.regardsMarc
regards, tom laneMarc MILLAS
