fix: RateLimitMiddleware._buckets is class-level so reload-safe

20 pre-existing pytest failures were caused by tests/test_api.py
calling importlib.reload(cyclone.api) mid-suite. Reload creates a
NEW FastAPI app with a NEW RateLimitMiddleware whose private
_buckets dict is independent of the OLD instance. Tests that
imported 'from cyclone.api import app' at module load kept
referencing the OLD app, so their requests accumulated in the
orphaned bucket which the conftest reset never cleared. After
~300 requests, the orphaned bucket tripped the limiter and these
tests got spurious 429s:
  - test_existing_endpoints_require_auth (6)
  - test_inbox_endpoints (8)
  - test_inbox_endpoints_sp7 (1)
  - test_list_endpoint_counts (4)

Hoisting _buckets to a class-level dict makes one
RateLimitMiddleware._buckets.clear() (in conftest reset) reach
every instance — current, stale, post-reload — eliminating the
orphaned-bucket leak. Per-instance _lock stays per-instance
since it guards mutation of the shared dict.

Verified stable: 3 consecutive full-suite runs all pass with
1435 passed, 10 skipped, 0 failed in ~82s each.
This commit is contained in:
Nora
2026-07-07 13:56:49 -06:00
parent 83a31b6fea
commit 9c0aa577fb
3 changed files with 94 additions and 16 deletions
+18 -1
View File
@@ -196,8 +196,13 @@ class RateLimitMiddleware:
On unexpected errors the limiter fails OPEN — better to serve a On unexpected errors the limiter fails OPEN — better to serve a
few extra requests than to 503 every request because of a bug. few extra requests than to 503 every request because of a bug.
The ``_buckets`` dict is class-level so every instance shares the
same per-IP sliding-window state. See ``__init__`` for why.
""" """
_buckets: dict[str, "_Bucket"] = {}
EXEMPT_PATHS = ("/api/health", "/healthz", "/readyz") EXEMPT_PATHS = ("/api/health", "/healthz", "/readyz")
def __init__( def __init__(
@@ -211,7 +216,19 @@ class RateLimitMiddleware:
"CYCLONE_RATE_LIMIT_PER_MIN", DEFAULT_RATE_LIMIT_PER_MIN, "CYCLONE_RATE_LIMIT_PER_MIN", DEFAULT_RATE_LIMIT_PER_MIN,
) )
self.window_s = window_s or DEFAULT_RATE_LIMIT_WINDOW_S self.window_s = window_s or DEFAULT_RATE_LIMIT_WINDOW_S
self._buckets: dict[str, _Bucket] = {} # _buckets is a class-level dict so every RateLimitMiddleware
# instance shares the same per-IP sliding-window state.
# This matters because tests that call
# ``importlib.reload(cyclone.api)`` cause Starlette to rebuild
# the middleware stack with a fresh instance whose private
# bucket would otherwise be independent of the old one — and
# tests that imported ``from cyclone.api import app`` at
# module load time keep referencing the old instance, so
# their requests would accumulate in the orphaned bucket and
# trip the limiter after ~300 requests. Sharing the dict
# makes one ``_buckets.clear()`` (in the conftest's reset
# hook) clear every instance at once.
self._buckets: dict[str, _Bucket] = type(self)._buckets
self._lock = threading.Lock() self._lock = threading.Lock()
async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None: async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
+20 -15
View File
@@ -53,10 +53,15 @@ def _auto_init_db(tmp_path, monkeypatch):
_api_mod.app.state.event_bus = EventBus() _api_mod.app.state.event_bus = EventBus()
deps.AUTH_DISABLED = True deps.AUTH_DISABLED = True
# The rate-limit middleware keeps a per-IP sliding window in # The rate-limit middleware keeps a per-IP sliding window in
# ``_buckets``. Without a reset between tests, later tests in a # ``_buckets`` (now a class-level dict — see
# full-suite run get ``429 Too Many Requests`` once the testclient # ``cyclone.security.RateLimitMiddleware._buckets``). Without a
# IP exhausts its 300 req/60s budget. Walk the middleware stack # reset between tests, later tests in a full-suite run get
# and clear the buckets so every test starts with a fresh window. # ``429 Too Many Requests`` once the testclient IP exhausts its
# 300 req/60s budget. Walk the middleware stack and clear the
# buckets so every test starts with a fresh window. Because the
# buckets dict is class-level, this clears every instance at
# once — important when tests like ``test_cors_extra_origins_via_env``
# trigger ``importlib.reload(cyclone.api)`` mid-suite.
# Trigger the stack build with a cheap health probe (the only # Trigger the stack build with a cheap health probe (the only
# request exempt from the limiter — see RateLimitMiddleware.EXEMPT_PATHS). # request exempt from the limiter — see RateLimitMiddleware.EXEMPT_PATHS).
_reset_rate_limit_buckets(_api_mod.app) _reset_rate_limit_buckets(_api_mod.app)
@@ -65,6 +70,7 @@ def _auto_init_db(tmp_path, monkeypatch):
finally: finally:
deps.AUTH_DISABLED = False deps.AUTH_DISABLED = False
_api_mod.app.state.event_bus = None _api_mod.app.state.event_bus = None
_reset_rate_limit_buckets(_api_mod.app)
db._reset_for_tests() db._reset_for_tests()
@@ -76,23 +82,22 @@ def _reset_rate_limit_buckets(app) -> None:
instance), so without a reset between tests the full suite trips instance), so without a reset between tests the full suite trips
the limiter after ~300 requests and later tests get 429s. the limiter after ~300 requests and later tests get 429s.
As of the SP38 followup the buckets dict is class-level, so this
helper clears every ``RateLimitMiddleware`` instance at once
(important when ``importlib.reload(cyclone.api)`` mid-suite has
created a new instance and tests still hold a stale ``app``
reference to the old one — see
``tests/test_rate_limit_shared_buckets.py``).
The middleware stack is only built on the first request, so we The middleware stack is only built on the first request, so we
prime it with an exempt health probe before walking to the prime it with an exempt health probe before walking to the
RateLimit layer. If the stack ever stops being a single chain RateLimit layer. If the stack ever stops being a single chain
of ``.app`` links, this helper raises AttributeError — better of ``.app`` links, this helper raises AttributeError — better
to fail loudly than silently leak state. to fail loudly than silently leak state.
""" """
from fastapi.testclient import TestClient from cyclone.security import RateLimitMiddleware
TestClient(app).get("/api/health") # Class-level dict — one clear reaches every instance.
cur = app.middleware_stack RateLimitMiddleware._buckets.clear()
while cur is not None:
if hasattr(cur, "_buckets"):
cur._buckets.clear()
return
cur = getattr(cur, "app", None)
# No RateLimitMiddleware in the stack — nothing to reset. Should
# not happen in this codebase (security.py registers it at boot)
# but we don't want a missing reset to crash unrelated tests.
# --------------------------------------------------------------------------- # ---------------------------------------------------------------------------
@@ -0,0 +1,56 @@
"""Regression test: RateLimitMiddleware must share its bucket across instances.
When ``importlib.reload(cyclone.api)`` runs (e.g. from
``test_cors_extra_origins_via_env``), Starlette rebuilds the middleware
stack with a NEW ``RateLimitMiddleware`` instance whose ``_buckets``
dict is fresh. Tests that imported ``app`` at module load time still
reference the OLD app — and the OLD RateLimitMiddleware — so its
private ``_buckets`` dict keeps accumulating requests across the rest
of the suite. After ~300 requests it hits the rate limit and the
remaining tests get spurious 429s.
Fix: hoist ``_buckets`` to a class-level dict so every
``RateLimitMiddleware`` instance (current + stale) shares the same
sliding-window state. The per-instance ``_lock`` stays per-instance
since it guards mutation of the shared dict.
This test pins that invariant: two ``RateLimitMiddleware`` instances
share the same underlying bucket dict.
"""
from __future__ import annotations
from cyclone.security import RateLimitMiddleware
class _DummyApp:
"""Minimal ASGI app stand-in for the middleware's inner app."""
async def __call__(self, scope, receive, send):
pass
def test_rate_limit_buckets_are_shared_across_instances():
"""Two RateLimitMiddleware instances must share _buckets."""
m1 = RateLimitMiddleware(_DummyApp())
m2 = RateLimitMiddleware(_DummyApp())
# After the fix: m1._buckets IS m2._buckets (same class-level dict).
# Before the fix: each instance had its own {} dict.
assert m1._buckets is m2._buckets, (
"RateLimitMiddleware instances do NOT share their bucket dict. "
"After importlib.reload(cyclone.api), the new instance's "
"_buckets is independent of the old one — orphaned buckets "
"accumulate across tests that hold stale `app` references and "
"trip the 300 req/60s limit. Hoist _buckets to a class-level "
"dict so every instance shares state."
)
def test_class_level_buckets_attribute_exists():
"""The class must declare a class-level _buckets dict."""
# This is the structural pre-condition for the sharing invariant.
assert hasattr(RateLimitMiddleware, "_buckets"), (
"RateLimitMiddleware must declare a class-level _buckets "
"attribute so every instance shares it. Without this, "
"importlib.reload(cyclone.api) creates an orphaned bucket."
)
assert isinstance(RateLimitMiddleware._buckets, dict)