Two Machines, One Recursion Limit

This DFS runs over a 200,000-node path graph. Both machines are 64-bit with an 8 MiB thread stack, and the input is identical.

import sys
sys.setrecursionlimit(300000)

def dfs(u):
    seen[u] = True
    for v in adj[u]:
        if not seen[v]:
            dfs(v)

On CPython 3.10 the process dies with a segmentation fault. On CPython 3.13 it finishes. What accounts for the difference?

On 3.13 setrecursionlimit also raises RLIMIT_STACK to fit the requested depth; on 3.10 it moved the interpreter counter alone.

Before 3.11 every Python call re-entered a C evaluation function, so pure Python recursion was capped by the C stack. From 3.11 the interpreter sets up the new frame and jumps to it inside its own loop, so a Python-to-Python call consumes no C stack and the depth is bounded only by the counter setrecursionlimit moves.

3.13 raised the default recursion limit above 200,000, so the setrecursionlimit call is what causes the crash on 3.10 and is redundant on 3.13.

3.13 sees that dfs calls only itself and rewrites the recursion as a loop, so no frames accumulate.