QA: num_cpus/num_gpus cannot be set on the operator. And when num_cpus is less than 1, it seems to be ineffective?

[Jay-ju]

Describe the bug

@daft.udf(return_dtype=DataType.fixed_size_list(size=768, dtype=DataType.float32()), num_gpus=0.1, num_cpus=0.1, batch_size=BATCH_SIZE, concurrency=8)

Here I set num_cpus = 0.1, but the number of CPUs is still 1.

In addition, I also found that for the operators of project and scan, even if I change the num_cpus here, there is no setting for this part of resources during planning. It feels like python did not pass it to the rust object.

To Reproduce

No response

Expected behavior

No response

Component(s)

Ray Runner

Additional context

No response

[colin-ho]

See:

Daft/daft/runners/ray_runner.py

Line 455 in 02806c4

options["num_cpus"] = max(1, resource_request.num_cpus)

, we currently set the max to 1.0 to prevent worker thrashing

[Jay-ju]

@colin-ho Thank you very much. I later looked at the code and saw this too. I feel that if we have to handle it according to 1 here, we will often encounter problems with resource utilization.

[rchowell]

@colin-ho I think the ask is not about >1, but about using 0.1. Does it make sense to update to:

num_cpus if num_cpus <= 1 else 1

[codewider]

Hi, any updates on this?
Our tasks are not CPU-intensive, so we’d like to allocate fractional CPUs. Currently, it seems that map_batches doesn’t support the num_cpus option — is there any plan to add this feature to the roadmap?

[Jay-ju]

Hi, any updates on this? Our tasks are not CPU-intensive, so we’d like to allocate fractional CPUs. Currently, it seems that map_batches doesn’t support the num_cpus option — is there any plan to add this feature to the roadmap?

@codewider I have already fixed this. You can try it out. If there are any issues, you can bring them up again.