Unify Copy-On-Write and Spilling #15436
Conversation
madsbk
added
improvement
| def __init__( | ||
| self, | ||
| ptr: int, | ||
| size: int, | ||
| owner: object, | ||
| exposed: bool, | ||
| ): | ||
| """Create a new buffer owner. | ||
|
|
||
| Do not use this directly, instead use `_from_device_memory` or | ||
| `_from_host_memory`. | ||
|
|
||
| Raises | ||
| ------ | ||
| ValueError | ||
| If size is negative | ||
| """ | ||
| if size < 0: | ||
| raise ValueError("size cannot be negative") | ||
|
|
||
| self._ptr = ptr | ||
| self._size = size | ||
| self._owner = owner | ||
| self._exposed = exposed | ||
| self._slices = weakref.WeakSet() |
There was a problem hiding this comment.
Notice, I have removed the direct-use-of-Buffer guard. I think the four mandatory arguments are enough to prevent accidental use.
Co-authored-by: Bradley Dice <[email protected]>
There was a problem hiding this comment.
I'm excited to have this happen! Thanks for pushing through Mads.
We should also not forget about these two bits of work proposed from the last round:
- Refactoring of Buffers (last step towards unifying COW and Spilling) #13801 (comment)
- Refactoring of Buffers (last step towards unifying COW and Spilling) #13801 (comment)
I also wonder if there is a sensible way to unify the SpillableBufferOwner with the BufferOwner (current names, which will be changed in the future). Not in scope for this PR, but some of my prior concerns about the Liskov substitutability of different Buffer types still apply there (also CC @wence-).
| @property | ||
| def exposed(self) -> bool: | ||
| return self._owner.exposed | ||
| self.owner._slices.add(self) |
There was a problem hiding this comment.
question: should we assert isinstance(owner, BufferOwner) here? Or is it not necessary because we implicitly require it via the _slices attribute?
There was a problem hiding this comment.
I guess, we could add assert isinstance(owner, BufferOwner) in Buffer.__init__ but I am not sure if that is pythonic?
There was a problem hiding this comment.
An assertion would basically be there to protect against developer error. There's nothing unpythonic about it, that's definitely what the functionality is there for. The main question I would have is, how would we end up with something other than a BufferOwner here, and how bad would the failure mode be? Unless the assertion makes it markedly easier to debug the code it doesn't add a whole lot of value IMHO.
There was a problem hiding this comment.
I think it is arguable a little unpythonic to prevent duck-typing by using isinstance to force a specify type (as opposed to a general ABC/protocol type).
Anyways, in this case as_buffer() is the only one creating ExposureTrackedBuffer and it explicitly sets the owner to SpillableBufferOwner | SpillableBuffer | BufferOwner so I don't think a type error here is likely.
Co-authored-by: Lawrence Mitchell <[email protected]>
|
Regarding renaming: I think we should do it in a follow up PR, it will touch a lot of files throughout cudf! Regarding #13801 (comment):
I agree, if dask or ucx should handle this, we need to publish a buffer API that includes this |
|
Happy to address one or both of those in a follow-up, especially since the second one apparently requires further discussion (I haven't given it any more thought yet). For the first, we should try to turn that around quickly to minimize conflicts since as you say it will touch a lot of files. Once you have that PR ready (after this one merges) feel free to ping me aggressively for reviews. |
|
/merge |
Description
This is the final step to unify COW and spilling. Now,
SpillableBufferinherits fromExposureTrackedBufferso the final class hierarchy becomes:Additionally, spill-on-demand is now set globally using
set_spill_on_demand_globally()instead of in theSpillManagerconstructor.Checklist