mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-16 19:04:08 +03:00
79 lines
3.4 KiB
Python
79 lines
3.4 KiB
Python
"""Concurrent writers of ``context_length_cache.yaml`` (race audit lostupdate-7).
|
|
|
|
Every agent (CLI, gateway sessions, cron jobs, subagents) persists probed context windows with a
|
|
load -> mutate -> write cycle; an unserialized pair rewrites a stale snapshot and drops the other
|
|
writer's entry, so the next session re-probes or overflows its context on the first call.
|
|
"""
|
|
|
|
import threading
|
|
|
|
import pytest
|
|
|
|
import agent.model_metadata as mm
|
|
from tests.lock_contention import watch_file_lock_contention
|
|
from tests.waiting import wait_until
|
|
|
|
|
|
@pytest.mark.parametrize("first_writer", ["save", "invalidate"])
|
|
def test_concurrent_cache_writers_keep_each_others_entries(monkeypatch, first_writer):
|
|
contended = watch_file_lock_contention(monkeypatch)
|
|
mm.save_context_length("model-a", "https://a.example/v1", 200_000)
|
|
real_load = mm._load_context_cache_document
|
|
loaded = threading.Event()
|
|
other_saved = threading.Event()
|
|
|
|
def parked_load():
|
|
document = real_load()
|
|
if threading.current_thread().name == "first-writer":
|
|
loaded.set()
|
|
# Unserialized writers let the other agent save here; serialized ones are seen
|
|
# blocking on the cache lock instead.
|
|
wait_until(lambda: other_saved.is_set() or contended.is_set(),
|
|
what="the other agent saved or blocked on the cache lock")
|
|
return document
|
|
|
|
monkeypatch.setattr(mm, "_load_context_cache_document", parked_load)
|
|
if first_writer == "save":
|
|
action = lambda: mm.save_context_length("model-a", "https://a.example/v1", 128_000)
|
|
expected_a = 128_000
|
|
else:
|
|
action = lambda: mm._invalidate_cached_context_length("model-a", "https://a.example/v1")
|
|
expected_a = None
|
|
worker = threading.Thread(target=action, name="first-writer")
|
|
worker.start()
|
|
assert loaded.wait(timeout=5)
|
|
mm.save_context_length("model-b", "https://b.example/v1", 64_000)
|
|
other_saved.set()
|
|
worker.join(timeout=10)
|
|
assert not worker.is_alive()
|
|
|
|
assert mm.get_cached_context_length("model-b", "https://b.example/v1") == 64_000
|
|
assert mm.get_cached_context_length("model-a", "https://a.example/v1") == expected_a
|
|
|
|
|
|
def test_invalidation_evicts_in_process_probe_caches_when_the_lock_is_unavailable(monkeypatch):
|
|
"""Only the file update needs the cross-process lock: a stale in-process probe result must be
|
|
dropped even when that lock cannot be taken, or the next resolution reuses it for its TTL."""
|
|
import contextlib
|
|
|
|
import hermes_cli.auth as auth_module
|
|
from hermes_constants import hermes_home_key
|
|
|
|
base_url = "https://a.example/v1"
|
|
mm._LOCAL_CTX_PROBE_CACHE[("model-a", base_url)] = (200_000, 0.0)
|
|
mm._LOCAL_CTX_PROBE_CACHE[("ollama_show", "model-a", base_url)] = (200_000, 0.0)
|
|
mm._BEDROCK_PROBE_FAILURE_CACHE[(hermes_home_key(), base_url, "model-a", "us-east-1")] = 0.0
|
|
|
|
@contextlib.contextmanager
|
|
def wedged_lock(*_args, **_kwargs):
|
|
raise TimeoutError("Timed out waiting for the context length cache lock")
|
|
yield
|
|
|
|
monkeypatch.setattr(auth_module, "_file_lock", wedged_lock)
|
|
|
|
mm._invalidate_cached_context_length("model-a", base_url)
|
|
|
|
assert ("model-a", base_url) not in mm._LOCAL_CTX_PROBE_CACHE
|
|
assert ("ollama_show", "model-a", base_url) not in mm._LOCAL_CTX_PROBE_CACHE
|
|
assert not [k for k in mm._BEDROCK_PROBE_FAILURE_CACHE if k[1:3] == (base_url, "model-a")]
|