stroblmeandClaude Fable 5
3f5bb12ed0
A training loop has numbers worth keeping thousands of steps before it has a result. Node code now imports fluksio and calls log_metric/progress, which sends a line back without ending the call; the engine writes those to run_metric in batches from the run's own driver rather than folding them off the event bus, which drops what it cannot keep up with. Two things fall out. Each event resets the worker deadline, so a node's timeout measures silence rather than duration — which is what lets a two-hour training keep a liveness contract instead of racing it. And the worker pool's _running is now keyed by (run, node), so cancelling one config of a sweep kills that training and leaves the rest alone. Fixes a latent framing bug: read_line returned whatever a read had taken, which was fine while a worker only ever sent one line per request and unparseable as soon as it sent several. It now keeps the remainder. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AD8SfVhzXBG2nAfFcVh3iD
283 lines
9.0 KiB
Python
283 lines
9.0 KiB
Python
"""Python nodes run in a worker process, and stay there when things go wrong."""
|
|
|
|
import sys
|
|
import threading
|
|
import time
|
|
from collections.abc import Iterator
|
|
|
|
import pytest
|
|
|
|
from app.flow.workers import NodeCancelled, NodeTimeout, PythonWorkerPool
|
|
|
|
|
|
@pytest.fixture
|
|
def pool() -> Iterator[PythonWorkerPool]:
|
|
# One worker: a respawn is then provably the same slot coming back.
|
|
worker_pool = PythonWorkerPool(python=sys.executable, size=1)
|
|
worker_pool.start()
|
|
yield worker_pool
|
|
worker_pool.stop()
|
|
|
|
|
|
def run(pool: PythonWorkerPool, code: str, node: str = "demo", **kwargs):
|
|
return pool.run(
|
|
"demo", node, code, kwargs, {"factor": 2}, f"demo.{node}", timeout=5
|
|
)
|
|
|
|
|
|
def test_a_node_returns_its_value_and_what_it_printed(pool, capsys):
|
|
result = run(
|
|
pool,
|
|
"def process(value, params):\n"
|
|
" print('seen', value)\n"
|
|
" return {'out': value * params['factor']}\n",
|
|
value=21,
|
|
)
|
|
assert result == {"out": 42}
|
|
# The proxy writes them to stdout, which is where the engine's tee is.
|
|
assert "seen 21" in capsys.readouterr().out
|
|
|
|
|
|
def test_a_failure_keeps_its_class_and_points_at_the_node(pool):
|
|
with pytest.raises(Exception) as caught:
|
|
run(pool, "def process(params):\n raise ValueError('bad input')\n")
|
|
|
|
# The engine renders a node error as "<class>: <message>", so both have to
|
|
# survive the trip.
|
|
assert type(caught.value).__name__ == "ValueError"
|
|
assert str(caught.value) == "bad input"
|
|
assert "<node demo." in caught.value.remote_traceback
|
|
assert "ValueError: bad input" in caught.value.remote_traceback
|
|
|
|
|
|
def test_a_node_that_kills_its_worker_is_an_ordinary_error(pool):
|
|
with pytest.raises(Exception, match="worker died"):
|
|
run(pool, "import os\n\n\ndef process(params):\n os._exit(1)\n")
|
|
|
|
assert run(pool, "def process(params):\n return {'out': 1}\n") == {"out": 1}
|
|
|
|
|
|
def test_a_node_that_runs_too_long_is_killed_and_the_pool_recovers(pool):
|
|
started = time.monotonic()
|
|
with pytest.raises(NodeTimeout):
|
|
pool.run(
|
|
"demo",
|
|
"slow",
|
|
"import time\n\n\ndef process(params):\n time.sleep(30)\n",
|
|
{},
|
|
{},
|
|
"demo.slow",
|
|
timeout=1,
|
|
)
|
|
assert time.monotonic() - started < 10
|
|
|
|
# The killed worker's slot is refilled on the next call.
|
|
assert run(pool, "def process(params):\n return {'out': 2}\n") == {"out": 2}
|
|
|
|
|
|
def test_a_running_node_can_be_cancelled(pool):
|
|
def stop_it() -> None:
|
|
for _ in range(100):
|
|
if pool.cancel("demo.slow"):
|
|
return
|
|
time.sleep(0.05)
|
|
|
|
stopper = threading.Thread(target=stop_it)
|
|
stopper.start()
|
|
try:
|
|
with pytest.raises(NodeCancelled):
|
|
pool.run(
|
|
"demo",
|
|
"slow",
|
|
"import time\n\n\ndef process(params):\n time.sleep(30)\n",
|
|
{},
|
|
{},
|
|
"demo.slow",
|
|
timeout=30,
|
|
)
|
|
finally:
|
|
stopper.join()
|
|
|
|
|
|
def test_a_result_that_is_not_json_is_refused(pool):
|
|
with pytest.raises(Exception, match="cannot be sent back as JSON"):
|
|
run(pool, "def process(params):\n return {'out': {1, 2}}\n")
|
|
|
|
|
|
def test_compiling_reports_where_the_source_is_wrong(pool):
|
|
assert pool.compile("demo", "broken", "def process(params)\n return {}\n")
|
|
assert pool.compile("demo", "fine", "def process(params):\n return {}\n") is None
|
|
|
|
|
|
def test_a_node_imports_the_standard_library_not_the_engines_own_modules(pool):
|
|
# The worker script lives in app/flow, which holds queue.py, secrets.py and
|
|
# more; the interpreter would put that directory first on sys.path.
|
|
result = run(
|
|
pool,
|
|
"import queue\nimport secrets\n\n\n"
|
|
"def process(params):\n"
|
|
" return {'out': [queue.Queue().qsize(), len(secrets.token_hex(4))]}\n",
|
|
)
|
|
assert result == {"out": [0, 8]}
|
|
|
|
|
|
def test_the_engines_secrets_are_not_in_a_workers_environment(pool, monkeypatch):
|
|
monkeypatch.setenv("SECRET_KEY", "not-for-nodes")
|
|
monkeypatch.setenv("POSTGRES_PASSWORD", "not-for-nodes")
|
|
monkeypatch.setenv("FLUKSIO_HARMLESS", "fine")
|
|
# A fresh process, so it is built from the environment set just now.
|
|
pool.respawn_all()
|
|
|
|
result = run(
|
|
pool,
|
|
"import os\n\n\n"
|
|
"def process(params):\n"
|
|
" return {'out': [k for k in ('SECRET_KEY', 'POSTGRES_PASSWORD',\n"
|
|
" 'FLUKSIO_HARMLESS') if k in os.environ]}\n",
|
|
)
|
|
assert result == {"out": ["FLUKSIO_HARMLESS"]}
|
|
|
|
|
|
def test_a_pool_can_stop_while_a_node_is_running(pool):
|
|
# One slot, taken by a node that will not finish on its own, and a second
|
|
# call queued behind it. The engine's node threads are not daemons, so a
|
|
# wait here is a shutdown that never completes.
|
|
outcomes: list[str] = []
|
|
|
|
def call(node: str) -> None:
|
|
try:
|
|
pool.run(
|
|
"demo",
|
|
node,
|
|
"import time\n\n\ndef process(params):\n time.sleep(60)\n",
|
|
{},
|
|
{},
|
|
f"demo.{node}",
|
|
timeout=60,
|
|
)
|
|
outcomes.append("returned")
|
|
except Exception as exc:
|
|
outcomes.append(type(exc).__name__)
|
|
|
|
busy = threading.Thread(target=call, args=("busy",))
|
|
busy.start()
|
|
# Let the first one take the slot, so the second is blocked acquiring it.
|
|
time.sleep(1)
|
|
waiting = threading.Thread(target=call, args=("waiting",))
|
|
waiting.start()
|
|
time.sleep(0.2)
|
|
|
|
pool.stop()
|
|
for thread in (busy, waiting):
|
|
thread.join(timeout=10)
|
|
assert not thread.is_alive()
|
|
assert len(outcomes) == 2
|
|
|
|
with pytest.raises(Exception, match="shutting down"):
|
|
run(pool, "def process(params):\n return {'out': 1}\n")
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Reporting from inside a node that has not returned yet
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
def test_a_node_reports_metrics_while_it_is_still_running(pool):
|
|
seen = []
|
|
result = pool.run(
|
|
"demo",
|
|
"train",
|
|
"import fluksio\n"
|
|
"def process(params):\n"
|
|
" for step in range(3):\n"
|
|
" fluksio.log_metric('loss', 1.0 / (step + 1), step)\n"
|
|
" fluksio.progress(0.5, 'halfway')\n"
|
|
" return {'out': 1}\n",
|
|
{},
|
|
{},
|
|
"demo.train",
|
|
timeout=5,
|
|
run_id="r1",
|
|
on_event=seen.append,
|
|
)
|
|
|
|
assert result == {"out": 1}
|
|
metrics = [event for event in seen if event["event"] == "metric"]
|
|
assert [(m["name"], m["step"]) for m in metrics] == [
|
|
("loss", 0),
|
|
("loss", 1),
|
|
("loss", 2),
|
|
]
|
|
assert metrics[0]["value"] == 1.0
|
|
# Every event says which call it belongs to, so a sweep can tell them apart.
|
|
assert {m["call_id"] for m in metrics} == {"r1:demo.train"}
|
|
assert [event["event"] for event in seen if event["event"] == "progress"] == [
|
|
"progress"
|
|
]
|
|
|
|
|
|
def test_events_hold_off_the_timeout_but_silence_does_not(pool):
|
|
# The deadline measures silence: a node reporting every 0.05s stays alive
|
|
# well past a 0.3s timeout, which is what a two-hour training needs.
|
|
result = pool.run(
|
|
"demo",
|
|
"slow",
|
|
"import time, fluksio\n"
|
|
"def process(params):\n"
|
|
" for step in range(12):\n"
|
|
" time.sleep(0.05)\n"
|
|
" fluksio.log_metric('beat', step, step)\n"
|
|
" return {'done': True}\n",
|
|
{},
|
|
{},
|
|
"demo.slow",
|
|
timeout=0.3,
|
|
run_id="r2",
|
|
on_event=lambda _event: None,
|
|
)
|
|
assert result == {"done": True}
|
|
|
|
with pytest.raises(NodeTimeout):
|
|
pool.run(
|
|
"demo",
|
|
"quiet",
|
|
"import time\ndef process(params):\n time.sleep(2)\n return {}\n",
|
|
{},
|
|
{},
|
|
"demo.quiet",
|
|
timeout=0.3,
|
|
)
|
|
|
|
|
|
def test_cancelling_one_run_leaves_the_same_node_in_another_alone(pool):
|
|
# Keyed by (run, node): cancelling a config of a sweep must not kill the
|
|
# rest of it. With one slot the second run is not executing, so the check
|
|
# is that the pool refuses to find it rather than killing the wrong worker.
|
|
started = threading.Event()
|
|
|
|
def hold():
|
|
try:
|
|
pool.run(
|
|
"demo",
|
|
"hold",
|
|
"import time\ndef process(params):\n time.sleep(5)\n return {}\n",
|
|
{},
|
|
{},
|
|
"demo.hold",
|
|
timeout=10,
|
|
run_id="run-a",
|
|
)
|
|
except Exception:
|
|
pass
|
|
finally:
|
|
started.set()
|
|
|
|
thread = threading.Thread(target=hold, daemon=True)
|
|
thread.start()
|
|
time.sleep(0.5)
|
|
|
|
assert pool.cancel("demo.hold", run_id="run-b") is False
|
|
assert pool.cancel("demo.hold", run_id="run-a") is True
|
|
started.wait(timeout=5)
|
|
thread.join(timeout=5)
|