A complete solution for stable and safe sort with spill

[zhuqi-lucas]

Is your feature request related to a problem or challenge?

This is a follow-up for #14644:

See details in #14644 (comment) cc @alamb @Kontinuation

We need a complete solution about stable sort with spill, several issues until now:

1. Ensuring we can always spill data (now spilling will sometimes fail if we run out of memory to sort the batches in)
2. Ensuring that we can always merge the data that was spilled, even if it had a really wide fanout (like 1000 of spill files)

The problem 1 could be solved by potentially spilling unsorted batches (and then sorting them separately). This would be less efficient (read/write some tuples twice but would work.

Problem 2 could use the multi-pass merge:

We can implement a more complicated multi-stage merging phase which merges a small portion of streams each time, and perform multiple rounds of merges to produce the final merged stream.

Describe the solution you'd like

The problem 1 could be solved by potentially spilling unsorted batches (and then sorting them separately). This would be less efficient (read/write some tuples twice but would work.

Problem 2 could use the multi-pass merge:

We can implement a more complicated multi-stage merging phase which merges a small portion of streams each time, and perform multiple rounds of merges to produce the final merged stream.

Describe alternatives you've considered

No response

Additional context

No response

[alamb]

Thank you for filing this @zhuqi-lucas . I have also added it to our list here

• [Epic] A collection of items related to processing larger than memory datasets (via spilling, externalized algorithm, etc) #14077