I would suggest to try to use sklearn.cluster.HDBSCAN instead of DBSCAN and report if you see any improvements.

According to #26726 (comment), it may use a lot less memory.

For this specific use case, I can also down sample the dataset. But I'd like to make this decision automatically.

And I feel for general use case, it would be great to be able to recover from this memory error, or even predict the error, such that the user can adapt the algorithm.

So the problem is likely a low-level one. Somewhere in our Cython code our memory usage grows, and at one point the OS OOM killer kills the Python process.

I am not sure there is a straightforward way to surface the error in a user-friendly manner but maybe I am wrong and if someone finds a way to improve the situation, this would be more than welcome!

Just for reference, I did some analysis of the memory usage:

And what I observe is this step-like increase of the memory usage. So I guess there could be an opportunity to do some clean exception.
Doing several runs with Dimensions=3, and number of samples between [10'000-55'000] yields memory usage of between 10-15 Bytes, assuming scaling with the square of the number of samples O(n^2) as mentioned in #22531

For now, the hacky-workaround for me is to start DBSCAN in a separate multiprocessing.Process and check if succeeded.

And thanks @lesteve for pointing out HDBSCAN. It works quite well and good fore some usecases, but DBSCAN is for me my dataset often faster.

I just saw that this is closely related to known issues of high memory usage in #17650

As requested, here is a minimal snippet that runs .fit() multiple times and logs memory:

import numpy as np
from sklearn.datasets import make_blobs
from sklearn.cluster import DBSCAN
import psutil, os, time

def get_mem():
    return psutil.Process(os.getpid()).memory_info().rss / 1024**2

X, _ = make_blobs(n_samples=100_000, centers=3, n_features=10, random_state=42)

print("Initial Memory: %.2f MiB" % get_mem())

for i in range(12):
    model = DBSCAN(eps=0.5, min_samples=5, n_jobs=1)
    model.fit(X)
    del model
    time.sleep(0.1)
    print(f"Iteration {i+1}: Memory = {get_mem():.2f} MiB")

@Tahseen23 can you edit your previous comment and add the output of your snippet when you run it locally 🙏.

Do you see memory usage growing? What is your conclusion?

I have made my conclusion #31526 (comment)