def test_noop_pause_of_experiment_without_trials() -> None:
"""
Walk through starting, pausing, and resuming a single no-op experiment
which will never schedule a trial.
"""
config_obj = conf.load_config(
conf.fixtures_path("no_op/single-one-short-step.yaml"))
impossibly_large = 100
config_obj["max_restarts"] = 0
config_obj["resources"] = {"slots_per_trial": impossibly_large}
with tempfile.NamedTemporaryFile() as tf:
with open(tf.name, "w") as f:
yaml.dump(config_obj, f)
experiment_id = exp.create_experiment(tf.name,
conf.fixtures_path("no_op"),
None)
exp.pause_experiment(experiment_id)
exp.wait_for_experiment_state(experiment_id, "PAUSED")
exp.activate_experiment(experiment_id)
exp.wait_for_experiment_state(experiment_id, "ACTIVE")
for _ in range(5):
assert exp.experiment_state(experiment_id) == "ACTIVE"
time.sleep(1)
exp.cancel_single(experiment_id)
def test_cancel_one_experiment() -> None:
experiment_id = exp.create_experiment(
conf.fixtures_path("no_op/single-many-long-steps.yaml"),
conf.fixtures_path("no_op"),
)
exp.cancel_single(experiment_id)
def test_cancel_ten_experiments() -> None:
experiment_ids = [
exp.create_experiment(
conf.fixtures_path("no_op/single-many-long-steps.yaml"),
conf.fixtures_path("no_op"),
) for _ in range(10)
]
for experiment_id in experiment_ids:
exp.cancel_single(experiment_id)
def test_noop_experiment_config_override() -> None:
config_obj = conf.load_config(conf.fixtures_path("no_op/single-one-short-step.yaml"))
with tempfile.NamedTemporaryFile() as tf:
with open(tf.name, "w") as f:
yaml.dump(config_obj, f)
experiment_id = exp.create_experiment(
tf.name,
conf.fixtures_path("no_op"),
["--config", "reproducibility.experiment_seed=8200"],
)
exp_config = exp.experiment_config_json(experiment_id)
assert exp_config["reproducibility"]["experiment_seed"] == 8200
exp.cancel_single(experiment_id)
def test_cancel_one_active_experiment_ready() -> None:
experiment_id = exp.create_experiment(
conf.tutorials_path("mnist_pytorch/const.yaml"),
conf.tutorials_path("mnist_pytorch"),
)
while 1:
if exp.experiment_has_completed_workload(experiment_id):
break
time.sleep(1)
exp.cancel_single(experiment_id, should_have_trial=True)
exp.assert_performed_final_checkpoint(experiment_id)
def test_cancel_one_active_experiment() -> None:
experiment_id = exp.create_experiment(
conf.fixtures_path("no_op/single-many-long-steps.yaml"), conf.fixtures_path("no_op"),
)
for _ in range(15):
if exp.experiment_has_active_workload(experiment_id):
break
time.sleep(1)
else:
raise AssertionError("no workload active after 15 seconds")
exp.cancel_single(experiment_id, should_have_trial=True)
def test_drain_agent() -> None:
"""
Start an experiment, `disable --drain` the agent once the trial is running,
make sure the experiment still finishes, but the new ones won't schedule.
"""
slots = _fetch_slots()
assert len(slots) == 1
agent_id = slots[0]["agent_id"]
experiment_id = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_state(experiment_id, determinedexperimentv1State.STATE_ACTIVE)
exp.wait_for_experiment_active_workload(experiment_id)
exp.wait_for_experiment_workload_progress(experiment_id)
# Disable and quickly enable it back.
with _disable_agent(agent_id, drain=True):
pass
# Try to launch another experiment. It shouldn't get scheduled because the
# slot is still busy with the first experiment.
experiment_id_no_start = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
time.sleep(5)
exp.wait_for_experiment_state(experiment_id_no_start, determinedexperimentv1State.STATE_ACTIVE)
with _disable_agent(agent_id, drain=True):
# Check for 15 seconds it doesn't get scheduled into the same slot.
for _ in range(15):
trials = exp.experiment_trials(experiment_id_no_start)
assert len(trials) == 0
# Ensure the first one has finished with the correct number of workloads.
exp.wait_for_experiment_state(experiment_id, determinedexperimentv1State.STATE_COMPLETED)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
assert len(trials[0].workloads) == 7
# Ensure the slot is empty.
slots = _fetch_slots()
assert len(slots) == 1
assert slots[0]["enabled"] is False
assert slots[0]["draining"] is True
assert slots[0]["allocation_id"] == "FREE"
# Check agent state.
command = ["det", "-m", conf.make_master_url(), "agent", "list", "--json"]
output = subprocess.check_output(command).decode()
agent_data = cast(List[Dict[str, Any]], json.loads(output))[0]
assert agent_data["id"] == agent_id
assert agent_data["enabled"] is False
assert agent_data["draining"] is True
exp.cancel_single(experiment_id_no_start)
