ray-0001: local node provider list_of_node_ips O(N²) membership test during cluster reconciliation

Severity: MEDIUM CWE: CWE-407 (Algorithmic Complexity — linear membership test in hot loop) Speedup: ~Nx at N=200 node cluster (verified by unit test) Target: Ray (ray-project/ray) Files:

python/ray/autoscaler/_private/local/node_provider.py:79-83 — ClusterState.__init__: list_of_node_ips linear scan
python/ray/autoscaler/_private/local/node_provider.py:147-149 — OnPremCoordinatorState.__init__: same pattern

Description

In ClusterState.__init__, node IP reconciliation builds a list and then scans it linearly inside a loop over all tracked workers:

list_of_node_ips = list(provider_config["worker_ips"])   # line 79 — creates a list
list_of_node_ips.append(provider_config["head_ip"])
for worker_ip in list(workers):                           # O(N) outer loop
    if worker_ip not in list_of_node_ips:                 # O(N) linear scan — O(N²) total
        del workers[worker_ip]

list_of_node_ips is a Python list. The not in test on a list is O(N) via sequential comparison. With N cluster nodes, total cost is O(N²).

The same pattern appears in OnPremCoordinatorState.__init__ (line 147-149), where list_of_node_ips is a list parameter passed from the caller, and the same O(N²) scan occurs during coordinator state initialization.

Both ClusterState.__init__ and OnPremCoordinatorState.__init__ are called during every cluster state sync (create_or_update call path), meaning the O(N²) cost is incurred on the autoscaler's hot reconciliation loop.

Root Cause

list() was used to convert provider_config["worker_ips"] (which may be any iterable) into a concrete sequence. Python list was chosen without considering that membership tests would be performed against it inside a loop. Converting to a set at construction time gives O(1) average membership test with no semantic change.

Patch

See patch/ray-0001-local-node-provider-list-membership.patch

Complexity Before

worker_ip not in list_of_node_ips: O(N) Total across N workers: O(N²)

Complexity After

worker_ip not in node_ip_set (set): O(1) average Total: O(N)

Reproduction

cd defects/ray/unit && javac -d . RayTest.java && java -ea unit.RayTest

2.3 KiB Raw Blame History