@rth, @EemeliSaari please confirm whether this helps, or is a separate issue

So far this improves a little bit the issue. Running the script from #11051 (comment) still produces similar output. But calling multiple time radius_neighbors does not increase memory usage anymore.

Thanks @recamshak !

Running the script from #11051 (comment) still produces similar output. But calling multiple time radius_neighbors does not increase memory usage anymore.

For me, this doesn't seem to change the output of the benchmark script in both cases, but I still think it's an improvement over the current version.

Side note: I find https://stackoverflow.com/a/23873586/1791279 helpful to understand the general mechanism on a small example. Though when I tried to increase N = 500000000 in test() which results in a 400MB array, and measure the memory usage with memory_profiler and htop I don't even see the allocation which is very strange..