Add spilling in SortMergeJoin

[viirya]

Is your feature request related to a problem or challenge?

In SortMergeJoin, it is possibly run of memory when it requires extra memory to hold polled buffer batches. We can consider adding spilling support there to make the operator resilient to the memory issue.

Describe the solution you'd like

Add spilling support in SortMergeJoin.

Describe alternatives you've considered

No response

Additional context

No response

[Omega359]

Possibly related: #8398, #9170, #5108, #1599

[comphead]

After reading some code and already opened issues on the same topic, probably its possible to summarize whats needed for POC at least:

• try to use external mem sorter to sort streamed and buffered batches for the sort phases
• MemoryReservation is already in SMJ impl and for merge phase it uses the try_grow on buffered, so naive approach is to spill buffered and read back lesser chunks from disk, we can play with that
• run test query and try to profile memory to see other places where spill can be useful

[comphead]

I'll try to start with creating a test query that fails on mem

[comphead]

Related to #9846

[comphead]

I think I managed to find the test query.
First observation is SMJ using external sorter, but the spill doesn't work, will dig into it.

Resources exhausted: Failed to allocate additional 1048576 bytes for ExternalSorterMerge[0] with 0 bytes already allocated - maximum available is 1043520

This gives me a suggestion ExternalSorter used but spill is not used

Playing with sort params I managed to make SMJ merge phase complain on mem exhaustion

./datafusion-cli/target/debug/datafusion-cli -m 1m
DataFusion CLI v37.0.0
❯ set datafusion.optimizer.prefer_hash_join = false;
0 row(s) fetched. 
Elapsed 0.003 seconds.

❯ set datafusion.execution.sort_in_place_threshold_bytes = 104857;
0 row(s) fetched. 
Elapsed 0.001 seconds.

❯ set datafusion.execution.sort_spill_reservation_bytes = 104857;
0 row(s) fetched. 
Elapsed 0.000 seconds.

❯ select * from (select unnest(range(0, 10000)) id) t inner join (select unnest(range(0, 100000)) id) t1 on t.id = t1.id;
Resources exhausted: Failed to allocate additional 165208 bytes for SMJStream[5] with 0 bytes already allocated - maximum available is 65456

[comphead]

My next steps

• figure out why ExternalSorterMerge doesn't use spill
• Do a spill in SMJ

[westonpace]

I can reproduce this pretty reliably in lance if it's any help. I run SortExec on a column of 100 million strings (each ~30 bytes long) and I have a 100MiB fair pool and it triggers in about 5 minutes. Let me know if there is anything I can do to assist you.

[milenkovicm]

FWIW, I believe I was looking at this problem some time ago, if I remember correctly the issue was that one of the memory consumer, presumably ExternalSorterMerge was returning all the memory it allocated just to ask for new allocation with similar size very soon after. MemoryConsumer would return all the memory when dropped.

My suspect is:

https://github.com/apache/arrow-datafusion/blob/a5cf0b8902ae55b81ac86b875c7e94cf1bdc205d/datafusion/physical-plan/src/sorts/sort.rs#L263

or

https://github.com/apache/arrow-datafusion/blob/a5cf0b8902ae55b81ac86b875c7e94cf1bdc205d/datafusion/physical-plan/src/sorts/sort.rs#L359

or somewhere around it.

line

Resources exhausted: Failed to allocate additional 165208 bytes for SMJStream[5] with 0 bytes already allocated - maximum available is 65456

may back my clain

I cant find more information looks like that branch is MIA

[comphead]

Thanks, I'm planning to check next week if spilling works or not for ExternalSorter used for SMJ and see how it is possible to make spilling work as well for the merge phase, the memory pool is injected and SMJ get new allocations through try_grow so lets see what we can do

[comphead]

Finally starting on it. The test doesn't work anymore