NumPy slice views keep deleted array's buffer alive — diagnosing retained memory

Answer

There is no supported way to detach a NumPy view from its underlying buffer without copying the data. The buffer stays alive as long as any view (including slices, transposes, or reshaped arrays) holds a reference to it. The writeable flag does not influence this reference‑counting behavior.

Confirmed facts

• NumPy views share the same data pointer as their base array; deleting the base variable only drops one reference.
• The buffer is freed only when the reference count of the underlying ndarray object reaches zero, which requires that all views no longer exist.
• view.writeable being False does not drop the reference; the buffer remains allocated regardless of the flag.

Steps to release the buffer (when a copy is acceptable)

1. Create a copy of the view: view_copy = view.copy().
2. Delete the original view and any other references: del view (and any other aliases).
3. Optionally delete the base array if it still exists: del base.
4. Run gc.collect() to reclaim the memory.

Steps to verify that the buffer is truly released (without copying)

1. Ensure no view references remain: set each view to None or let them go out of scope.
2. Delete the base array.
3. Call gc.collect().
4. Check that view.base is None (if you still have a reference to the view) or use np.shares_memory(view, base) when the base is still accessible; both should indicate no sharing.
5. Measure memory before and after with psutil.Process().memory_info().rss or tracemalloc. A genuine drop in RSS after gc.collect() confirms the buffer was freed.

Missing diagnostic detail

The exact NumPy version can affect older behavior of the base attribute. If you are using a release prior to 1.13, please confirm your numpy.__version__ so we can version‑specific advice.