pyodide: memory leak when accessing a.b. a is a Python instance & PyProxy in JS, b is either np.ndarray/bytes and is PyProxyBuffer in JS

Test environment:

• OS: macOS 11.6
• Chrome: 93.0.4577.82
• Hardware: MacBook Air (M1, 2020)
• Pyodide: 0.18.0

User scenario: I’m trying to make a medical image viewer which needs frequent (10~30 fps) rerendering in some cases. I define a Python class and plan to reuse it on JS side, this idea is from #1426 and a part of the code of the flow is like the part in https://github.com/pyodide/pyodide/discussions/1833#discussioncomment-1329114.

Python class (in an external Python file):

@dataclass
Class Dicom:
    ndarray_data: Optional[np.ndarray] = None # PyProxyBuffer in JS
    pixel_bytes: Optional[bytes] = None. #  # PyProxyBuffer in JS

When the mouse button is pressed and a mouse move event is triggered, JS will

• execute dicomObj.rerender(). dicomObj is a PyProxy.
  • some ndarray/bytes internally are re-calculated and renewed to ndarray_data/pixel_bytes attributes on Python side.
• trying to access the resulted ndarray_data or pixel_bytes.

The accessing result is

• ndarray_data
  • as long as dicomObj.rgba_1d_ndarray is accessed on JS. The memory leak is over 100 MB/s (fps is over 30fps), regardless I tried rgba_1d_ndarray.destroy() and commenting the following getBuffer usage.
  • I tried @property to access Python internal ndarray_data but got the same memory leak.
  • Use a method to access internal ndarray data and the memory will not leak. Use dicomObj.get_ndarray_data() to access. On python:

  ## workaround solution
  def get_ndarray_data(self):
      return self.ndarray_data
  

• pixel_bytes:
  • it will have memory leak but very slow (in similar flow), 1 MB/s (fps is over 30fps). And direct accessing the attribute, using property, and even using a method will all leak. Comparing to ndarray_data, this part is not very harmful.