[FEA] Delete default constructor for device_buffer and device_uvector

[jrhemstad]

Is your feature request related to a problem? Please describe.

Similar to #418

The default constructor for device_buffer and device_uvector are problematic as they do not allow specifying a stream. This can lead to situations where the user isn't aware an operation is stream ordered:

device_buffer b{};
b.resize(100); // Defaults to using the default CUDA stream

Describe the solution you'd like

Delete the default ctor and force users to explicit specify a zero size and a stream.

[jakirkham]

I thought device_buffer used the default stream in this case. Or did that change?

[jrhemstad]

I thought device_buffer used the default stream in this case. Or did that change?

It does, and that's precisely the problem. Due to the widespread misuse of the defaulted stream arguments with device_buffer in libcudf and elsewhere, we've decided to make stream arguments explicit and force users to specify a stream to reduce the chance of doing the wrong thing.

[jakirkham]

I think this will become challenging in Python as we don't have a good way to handle stream objects today. It will also requires other Python libraries that don't really know what streams are in use/available to think about them, which will make things challenging from an ecosystem perspective.

[jrhemstad]

I think this will become challenging in Python as we don't have a good way to handle stream objects today. It will also requires other Python libraries that don't really know what streams are in use/available to think about them, which will make things challenging from an ecosystem perspective.

Nothing is stopping the Python side from still plugging in the default stream. The point is to make it explicit and force that behavior rather than defaulting the value.

[jakirkham]

Is this still an issue if we use PTDS ( #416 )?

[jrhemstad]

Is this still an issue if we use PTDS ( #416 )?

It's orthogonal to PTDS. The point is, to use a device_buffer you always need to be explicit about what stream it's being used on, even if it's the CUDA default stream.

[github-actions]

This issue has been marked rotten due to no recent activity in the past 90d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[github-actions]

This issue has been marked stale due to no recent activity in the past 30d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be marked rotten if there is no activity in the next 60d.

[harrism]

@jrhemstad still relevant?

[jrhemstad]

#424 (comment)

Closing.