[Perf] Source filters re-generate polydata on every request

[floryst]

High-level description

Context: #3280 (comment)

Should Source filters re-generate polydata on every request? What does VTK C++ do?

@jourdain @finetjul

Steps to reproduce

No repro, just a discussion

Detailed Current Behavior

N/A

Environment

N/A

[finetjul]

What does VTK C++ do?

VTK C++ reuse the same polydata. However, there is a new instance of vtkPoints.

[jourdain]

If the filter is not modified and none of if its input are modified (mtime>) then, the filter should not re-execute.

Each render call should re-use previous polydata.

[jourdain]

The method requestData should only execute once.

[finetjul]

The method requestData should only execute once.

The question was, when requestData() is called, should it generate a new polydata as output each time or reuse the same polydata.

[jourdain]

Should be a new polydata normally. But for performance reason we could have an option that reuse it in some special filter in case that make sense.

[finetjul]

Should be a new polydata normally

Why should it be a new polydata ? This is not the behavior of VTK C++

[jourdain]

It is. The executive give you a new instance of the dataset that the filter is suppose to fill.

Think about this way

cone_source = vtkConeSource()

cone_source.Update()
ds1 = cone_source.GetOutput()

cone_source.resolution = 24
cone_source.Update()
ds2 = cone_source.GetOutput()

assert ds1 != ds2

[finetjul]

In the example you shared, ds1 == ds2

[jourdain]

oh wow, I was not expecting that. I guess if I wanted to keep ds1 around, I would need to ShallowCopy it into another polydata.

[finetjul]

My point is that is should behave as in C++. And by doing so it would also improve performance in case of frequent filter updates. (limit the number of time-consuming newInstance() calls).
How does that sound ?

[jourdain]

Agree, it should behave like C++. Sounds good to me