def test_job_queue_adjust_weight() -> None:
config = conf.tutorials_path("mnist_pytorch/const.yaml")
model = conf.tutorials_path("mnist_pytorch")
for _ in range(2):
exp.create_experiment(config, model)
jobs = JobInfo()
ok = jobs.refresh_until_populated()
assert ok
ordered_ids = jobs.get_ids()
subprocess.run(["det", "job", "update", ordered_ids[0], "--weight", "10"])
sleep(2)
jobs.refresh()
new_weight = jobs.get_job_weight(ordered_ids[0])
assert new_weight == "10"
subprocess.run(
["det", "job", "update-batch", f"{ordered_ids[1]}.weight=10"])
sleep(2)
jobs.refresh()
new_weight = jobs.get_job_weight(ordered_ids[1])
assert new_weight == "10"
def test_drain_agent_sched() -> None:
"""
Start an experiment, drain it. Start a second one and make sure it schedules
on the second agent *before* the first one has finished.
"""
slots = _wait_for_slots(2)
assert len(slots) == 2
exp_id1 = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_workload_progress(exp_id1)
slots = _fetch_slots()
used_slots = [s for s in slots if s["allocation_id"] != "FREE"]
assert len(used_slots) == 1
agent_id1 = used_slots[0]["agent_id"]
with _disable_agent(agent_id1, drain=True):
exp_id2 = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_state(exp_id2,
determinedexperimentv1State.STATE_ACTIVE)
# Wait for a state when *BOTH* experiments are scheduled.
for _ in range(20):
slots = _fetch_slots()
assert len(slots) == 2
used_slots = [s for s in slots if s["allocation_id"] != "FREE"]
if len(used_slots) == 2:
# All good.
break
else:
pytest.fail(
"Second experiment didn't schedule on the second agent "
"while the first agent was draining")
exp.wait_for_experiment_state(
exp_id1, determinedexperimentv1State.STATE_COMPLETED)
exp.wait_for_experiment_state(
exp_id2, determinedexperimentv1State.STATE_COMPLETED)
trials1 = exp.experiment_trials(exp_id1)
trials2 = exp.experiment_trials(exp_id2)
assert len(trials1) == len(trials2) == 1
assert len(trials1[0].workloads) == len(trials2[0].workloads) == 7
def test_master_restart_reattach_recover_experiment(
managed_cluster_restarts: ManagedCluster, downtime: int) -> None:
_sanity_check(managed_cluster_restarts)
try:
exp_id = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
# TODO(ilia): don't wait for progress.
exp.wait_for_experiment_workload_progress(exp_id)
if downtime >= 0:
managed_cluster_restarts.kill_master()
time.sleep(downtime)
managed_cluster_restarts.restart_master()
exp.wait_for_experiment_state(exp_id,
EXP_STATE.STATE_COMPLETED,
max_wait_secs=downtime + 60)
trials = exp.experiment_trials(exp_id)
assert len(trials) == 1
train_wls = exp.workloads_with_training(trials[0].workloads)
assert len(train_wls) == 5
except Exception:
managed_cluster_restarts.restart_master()
managed_cluster_restarts.restart_agent()
raise
def test_master_restart_kill_works(
managed_cluster_restarts: ManagedCluster) -> None:
_sanity_check(managed_cluster_restarts)
try:
exp_id = exp.create_experiment(
conf.fixtures_path("no_op/single-many-long-steps.yaml"),
conf.fixtures_path("no_op"),
[
"--config", "searcher.max_length.batches=10000", "--config",
"max_restarts=0"
],
)
exp.wait_for_experiment_workload_progress(exp_id)
managed_cluster_restarts.kill_master()
time.sleep(0)
managed_cluster_restarts.restart_master()
command = [
"det", "-m",
conf.make_master_url(), "e", "kill",
str(exp_id)
]
subprocess.check_call(command)
exp.wait_for_experiment_state(exp_id,
EXP_STATE.STATE_CANCELED,
max_wait_secs=10)
managed_cluster_restarts.ensure_agent_ok()
except Exception:
managed_cluster_restarts.restart_master()
managed_cluster_restarts.restart_agent()
def test_agent_reconnect_keep_experiment(
managed_cluster_restarts: ManagedCluster) -> None:
managed_cluster_restarts.ensure_agent_ok()
try:
exp_id = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_workload_progress(exp_id)
managed_cluster_restarts.kill_proxy()
time.sleep(1)
managed_cluster_restarts.restart_proxy()
exp.wait_for_experiment_state(exp_id, EXP_STATE.STATE_COMPLETED)
trials = exp.experiment_trials(exp_id)
assert len(trials) == 1
train_wls = exp.workloads_with_training(trials[0].workloads)
assert len(train_wls) == 5
except Exception:
managed_cluster_restarts.restart_proxy(wait_for_reconnect=False)
managed_cluster_restarts.restart_agent()
raise
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_noop_pause() -> None:
"""
Walk through starting, pausing, and resuming a single no-op experiment.
"""
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, "ACTIVE")
# Wait for the only trial to get scheduled.
workload_active = False
for _ in range(conf.MAX_TASK_SCHEDULED_SECS):
workload_active = exp.experiment_has_active_workload(experiment_id)
if workload_active:
break
else:
time.sleep(1)
check.true(
workload_active,
f"The only trial cannot be scheduled within {conf.MAX_TASK_SCHEDULED_SECS} seconds.",
)
# Wait for the only trial to show progress, indicating the image is built and running.
num_steps = 0
for _ in range(conf.MAX_TRIAL_BUILD_SECS):
trials = exp.experiment_trials(experiment_id)
if len(trials) > 0:
only_trial = trials[0]
num_steps = len(only_trial["steps"])
if num_steps > 1:
break
time.sleep(1)
check.true(
num_steps > 1,
f"The only trial cannot start training within {conf.MAX_TRIAL_BUILD_SECS} seconds.",
)
# Pause the experiment. Note that Determined does not currently differentiate
# between a "stopping paused" and a "paused" state, so we follow this check
# up by ensuring the experiment cleared all scheduled workloads.
exp.pause_experiment(experiment_id)
exp.wait_for_experiment_state(experiment_id, "PAUSED")
# Wait at most 20 seconds for the experiment to clear all workloads (each
# train step should take 5 seconds).
for _ in range(20):
workload_active = exp.experiment_has_active_workload(experiment_id)
if not workload_active:
break
else:
time.sleep(1)
check.true(
not workload_active, "The experiment cannot be paused within 20 seconds.",
)
# Resume the experiment and wait for completion.
exp.activate_experiment(experiment_id)
exp.wait_for_experiment_state(experiment_id, "COMPLETED")
def test_labels() -> None:
experiment_id = exp.create_experiment(
conf.fixtures_path("no_op/single-one-short-step.yaml"),
conf.fixtures_path("no_op"), None)
label = "__det_test_dummy_label__"
# Add a label and check that it shows up.
subprocess.check_call([
"det", "-m",
conf.make_master_url(), "e", "label", "add",
str(experiment_id), label
])
output = subprocess.check_output([
"det", "-m",
conf.make_master_url(), "e", "describe",
str(experiment_id)
]).decode()
assert label in output
# Remove the label and check that it doesn't show up.
subprocess.check_call([
"det", "-m",
conf.make_master_url(), "e", "label", "remove",
str(experiment_id), label
])
output = subprocess.check_output([
"det", "-m",
conf.make_master_url(), "e", "describe",
str(experiment_id)
]).decode()
assert label not in output
def test_disable_agent_experiment_resume() -> None:
"""
Start an experiment with max_restarts=0 and ensure that being killed due to an explicit agent
disable/enable (without draining) does not count toward the number of restarts.
"""
slots = _fetch_slots()
assert len(slots) == 1
agent_id = slots[0]["agent_id"]
exp_id = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
["--config", "max_restarts=0"],
)
exp.wait_for_experiment_workload_progress(exp_id)
with _disable_agent(agent_id):
# Wait for the allocation to go away.
for _ in range(20):
slots = _fetch_slots()
print(slots)
if not any(s["allocation_id"] != "FREE" for s in slots):
break
time.sleep(1)
else:
pytest.fail("Experiment stayed scheduled after agent was disabled")
exp.wait_for_experiment_state(exp_id, determinedexperimentv1State.STATE_COMPLETED)
def test_streaming_observability_metrics_apis(
framework_base_experiment: str, framework_timings_enabled: bool
) -> None:
# TODO: refactor tests to not use cli singleton auth.
certs.cli_cert = certs.default_load(conf.make_master_url())
authentication.cli_auth = authentication.Authentication(conf.make_master_url(), try_reauth=True)
config_path = conf.tutorials_path(f"../{framework_base_experiment}/const.yaml")
model_def_path = conf.tutorials_path(f"../{framework_base_experiment}")
config_obj = conf.load_config(config_path)
config_obj = conf.set_profiling_enabled(config_obj)
with tempfile.NamedTemporaryFile() as tf:
with open(tf.name, "w") as f:
yaml.dump(config_obj, f)
experiment_id = exp.create_experiment(
tf.name,
model_def_path,
)
exp.wait_for_experiment_state(experiment_id, "COMPLETED")
trials = exp.experiment_trials(experiment_id)
trial_id = trials[0]["id"]
gpu_enabled = conf.GPU_ENABLED
request_profiling_metric_labels(trial_id, framework_timings_enabled, gpu_enabled)
if gpu_enabled:
request_profiling_system_metrics(trial_id, "gpu_util")
if framework_timings_enabled:
request_profiling_pytorch_timing_metrics(trial_id, "train_batch")
def test_agent_restart_recover_experiment(
managed_cluster_restarts: ManagedCluster, downtime: int) -> None:
if not managed_cluster_restarts.reattach:
pytest.skip()
managed_cluster_restarts.ensure_agent_ok()
try:
exp_id = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_workload_progress(exp_id)
if downtime >= 0:
managed_cluster_restarts.kill_agent()
time.sleep(downtime)
managed_cluster_restarts.restart_agent(wait_for_amnesia=False)
exp.wait_for_experiment_state(exp_id, EXP_STATE.STATE_COMPLETED)
trials = exp.experiment_trials(exp_id)
assert len(trials) == 1
train_wls = exp.workloads_with_training(trials[0].workloads)
assert len(train_wls) == 5
except Exception:
managed_cluster_restarts.restart_agent()
raise
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_streaming_metrics_api() -> None:
auth.initialize_session(conf.make_master_url(), try_reauth=True)
pool = mp.pool.ThreadPool(processes=7)
experiment_id = exp.create_experiment(
conf.fixtures_path("mnist_pytorch/adaptive_short.yaml"),
conf.tutorials_path("mnist_pytorch"),
)
# To fully test the streaming APIs, the requests need to start running immediately after the
# experiment, and then stay open until the experiment is complete. To accomplish this with all
# of the API calls on a single experiment, we spawn them all in threads.
# The HP importance portion of this test is commented out until the feature is enabled by
# default
metric_names_thread = pool.apply_async(request_metric_names,
(experiment_id, ))
train_metric_batches_thread = pool.apply_async(
request_train_metric_batches, (experiment_id, ))
valid_metric_batches_thread = pool.apply_async(
request_valid_metric_batches, (experiment_id, ))
train_trials_snapshot_thread = pool.apply_async(
request_train_trials_snapshot, (experiment_id, ))
valid_trials_snapshot_thread = pool.apply_async(
request_valid_trials_snapshot, (experiment_id, ))
train_trials_sample_thread = pool.apply_async(request_train_trials_sample,
(experiment_id, ))
valid_trials_sample_thread = pool.apply_async(request_valid_trials_sample,
(experiment_id, ))
metric_names_results = metric_names_thread.get()
train_metric_batches_results = train_metric_batches_thread.get()
valid_metric_batches_results = valid_metric_batches_thread.get()
train_trials_snapshot_results = train_trials_snapshot_thread.get()
valid_trials_snapshot_results = valid_trials_snapshot_thread.get()
train_trials_sample_results = train_trials_sample_thread.get()
valid_trials_sample_results = valid_trials_sample_thread.get()
if metric_names_results is not None:
pytest.fail("metric-names: %s. Results: %s" % metric_names_results)
if train_metric_batches_results is not None:
pytest.fail("metric-batches (training): %s. Results: %s" %
train_metric_batches_results)
if valid_metric_batches_results is not None:
pytest.fail("metric-batches (validation): %s. Results: %s" %
valid_metric_batches_results)
if train_trials_snapshot_results is not None:
pytest.fail("trials-snapshot (training): %s. Results: %s" %
train_trials_snapshot_results)
if valid_trials_snapshot_results is not None:
pytest.fail("trials-snapshot (validation): %s. Results: %s" %
valid_trials_snapshot_results)
if train_trials_sample_results is not None:
pytest.fail("trials-sample (training): %s. Results: %s" %
train_trials_sample_results)
if valid_trials_sample_results is not None:
pytest.fail("trials-sample (validation): %s. Results: %s" %
valid_trials_sample_results)
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_nan_validations() -> None:
"""
Ensure that NaN validation metric values don't prevent an experiment from completing.
"""
experiment_id = exp.create_experiment(
conf.fixtures_path("no_op/single-nan-validations.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_state(
experiment_id, bindings.determinedexperimentv1State.STATE_COMPLETED
)
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_experiment_archive_unarchive() -> None:
experiment_id = exp.create_experiment(
conf.fixtures_path("no_op/single.yaml"), conf.fixtures_path("no_op"),
["--paused"])
describe_args = [
"det",
"-m",
conf.make_master_url(),
"experiment",
"describe",
"--json",
str(experiment_id),
]
# Check that the experiment is initially unarchived.
infos = json.loads(subprocess.check_output(describe_args))
assert len(infos) == 1
assert not infos[0]["archived"]
# Check that archiving a non-terminal experiment fails, then terminate it.
with pytest.raises(subprocess.CalledProcessError):
subprocess.check_call([
"det", "-m",
conf.make_master_url(), "experiment", "archive",
str(experiment_id)
])
subprocess.check_call([
"det", "-m",
conf.make_master_url(), "experiment", "cancel",
str(experiment_id)
])
# Check that we can archive and unarchive the experiment and see the expected effects.
subprocess.check_call([
"det", "-m",
conf.make_master_url(), "experiment", "archive",
str(experiment_id)
])
infos = json.loads(subprocess.check_output(describe_args))
assert len(infos) == 1
assert infos[0]["archived"]
subprocess.check_call([
"det", "-m",
conf.make_master_url(), "experiment", "unarchive",
str(experiment_id)
])
infos = json.loads(subprocess.check_output(describe_args))
assert len(infos) == 1
assert not infos[0]["archived"]
def test_hp_importance_api() -> None:
auth.initialize_session(conf.make_master_url(), try_reauth=True)
pool = mp.pool.ThreadPool(processes=1)
experiment_id = exp.create_experiment(
conf.fixtures_path("mnist_pytorch/random.yaml"),
conf.tutorials_path("mnist_pytorch"),
)
hp_importance_thread = pool.apply_async(request_hp_importance, (experiment_id,))
hp_importance_results = hp_importance_thread.get()
if hp_importance_results is not None:
pytest.fail("hyperparameter-importance: %s. Results: %s" % hp_importance_results)
def test_agent_restart_exp_container_failure(
managed_cluster_restarts: ManagedCluster) -> None:
managed_cluster_restarts.ensure_agent_ok()
try:
exp_id = exp.create_experiment(
conf.fixtures_path("no_op/single-medium-train-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_workload_progress(exp_id)
container_ids = list(_local_container_ids_for_experiment(exp_id))
if len(container_ids) != 1:
pytest.fail(
f"unexpected number of local containers for the experiment: {len(container_ids)}"
)
# Get task id / allocation id
tasks_data = _task_list_json(managed_cluster_restarts.master_url)
assert len(tasks_data) == 1
exp_task_before = list(tasks_data.values())[0]
managed_cluster_restarts.kill_agent()
subprocess.run(["docker", "kill", container_ids[0]],
check=True,
stdout=subprocess.PIPE)
except Exception:
managed_cluster_restarts.restart_agent()
raise
else:
managed_cluster_restarts.restart_agent()
# As soon as the agent is back, the original allocation should be considered dead,
# but the new one should be allocated.
state = exp.experiment_state(exp_id)
assert state == EXP_STATE.STATE_ACTIVE
tasks_data = _task_list_json(managed_cluster_restarts.master_url)
assert len(tasks_data) == 1
exp_task_after = list(tasks_data.values())[0]
assert exp_task_before["task_id"] == exp_task_after["task_id"]
assert exp_task_before["allocation_id"] != exp_task_after[
"allocation_id"]
exp.wait_for_experiment_state(exp_id, EXP_STATE.STATE_COMPLETED)
def test_noop_pause() -> None:
"""
Walk through starting, pausing, and resuming a single no-op experiment.
"""
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, bindings.determinedexperimentv1State.STATE_ACTIVE)
# Wait for the only trial to get scheduled.
exp.wait_for_experiment_active_workload(experiment_id)
# Wait for the only trial to show progress, indicating the image is built and running.
exp.wait_for_experiment_workload_progress(experiment_id)
# Pause the experiment. Note that Determined does not currently differentiate
# between a "stopping paused" and a "paused" state, so we follow this check
# up by ensuring the experiment cleared all scheduled workloads.
exp.pause_experiment(experiment_id)
exp.wait_for_experiment_state(experiment_id, bindings.determinedexperimentv1State.STATE_PAUSED)
# Wait at most 20 seconds for the experiment to clear all workloads (each
# train step should take 5 seconds).
for _ in range(20):
workload_active = exp.experiment_has_active_workload(experiment_id)
if not workload_active:
break
else:
time.sleep(1)
check.true(
not workload_active,
"The experiment cannot be paused within 20 seconds.",
)
# Resume the experiment and wait for completion.
exp.activate_experiment(experiment_id)
exp.wait_for_experiment_state(
experiment_id, bindings.determinedexperimentv1State.STATE_COMPLETED
)
def run_gc_checkpoints_test(checkpoint_storage: Dict[str, str]) -> None:
fixtures = [
(
conf.fixtures_path("no_op/gc_checkpoints_decreasing.yaml"),
{
"COMPLETED": {8, 9, 10},
"DELETED": {1, 2, 3, 4, 5, 6, 7}
},
),
(
conf.fixtures_path("no_op/gc_checkpoints_increasing.yaml"),
{
"COMPLETED": {1, 2, 3, 9, 10},
"DELETED": {4, 5, 6, 7, 8}
},
),
]
all_checkpoints = []
for base_conf_path, result in fixtures:
config = conf.load_config(str(base_conf_path))
config["checkpoint_storage"].update(checkpoint_storage)
with tempfile.NamedTemporaryFile() as tf:
with open(tf.name, "w") as f:
yaml.dump(config, f)
experiment_id = exp.create_experiment(tf.name,
conf.fixtures_path("no_op"))
exp.wait_for_experiment_state(experiment_id, "COMPLETED")
# Checkpoints are not marked as deleted until gc_checkpoint task starts.
retries = 5
for retry in range(retries):
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
checkpoints = sorted(
(step.checkpoint for step in trials[0].steps),
key=operator.itemgetter("step_id"),
)
assert len(checkpoints) == 10
by_state = {} # type: Dict[str, Set[int]]
for checkpoint in checkpoints:
by_state.setdefault(checkpoint.state,
set()).add(checkpoint.step_id)
if by_state == result:
all_checkpoints.append((config, checkpoints))
break
if retry + 1 == retries:
assert by_state == result
time.sleep(1)
# Check that the actual checkpoint storage (for shared_fs) reflects the
# deletions. We want to wait for the GC containers to exit, so check
# repeatedly with a timeout.
max_checks = 30
for check in range(max_checks):
time.sleep(1)
try:
for config, checkpoints in all_checkpoints:
checkpoint_config = config["checkpoint_storage"]
if checkpoint_config["type"] == "shared_fs" and (
"storage_path" not in checkpoint_config):
if "tensorboard_path" in checkpoint_config:
checkpoint_config[
"storage_path"] = checkpoint_config.get(
"tensorboard_path", None)
else:
checkpoint_config[
"storage_path"] = checkpoint_config.get(
"checkpoint_path", None)
root = os.path.join(checkpoint_config["host_path"],
checkpoint_config["storage_path"])
for checkpoint in checkpoints:
dirname = os.path.join(root, checkpoint.uuid)
if checkpoint.state == "COMPLETED":
assert os.path.isdir(dirname)
elif checkpoint.state == "DELETED":
assert not os.path.exists(dirname)
except AssertionError:
if check == max_checks - 1:
raise
else:
break
def run_gc_checkpoints_test(checkpoint_storage: Dict[str, str]) -> None:
fixtures = [
(
conf.fixtures_path("no_op/gc_checkpoints_decreasing.yaml"),
{
"COMPLETED": {8, 9, 10},
"DELETED": {1, 2, 3, 4, 5, 6, 7}
},
),
(
conf.fixtures_path("no_op/gc_checkpoints_increasing.yaml"),
{
"COMPLETED": {1, 2, 3, 9, 10},
"DELETED": {4, 5, 6, 7, 8}
},
),
]
all_checkpoints = []
for base_conf_path, result in fixtures:
config = conf.load_config(str(base_conf_path))
config["checkpoint_storage"].update(checkpoint_storage)
with tempfile.NamedTemporaryFile() as tf:
with open(tf.name, "w") as f:
yaml.dump(config, f)
experiment_id = exp.create_experiment(tf.name,
conf.fixtures_path("no_op"))
exp.wait_for_experiment_state(experiment_id, "COMPLETED")
# Checkpoints are not marked as deleted until gc_checkpoint task starts.
retries = 5
for retry in range(retries):
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
checkpoints = sorted(
(step["checkpoint"] for step in trials[0]["steps"]),
key=operator.itemgetter("step_id"),
)
assert len(checkpoints) == 10
by_state = {} # type: Dict[str, Set[int]]
for checkpoint in checkpoints:
by_state.setdefault(checkpoint["state"],
set()).add(checkpoint["step_id"])
if by_state == result:
all_checkpoints.append((config, checkpoints))
break
if retry + 1 == retries:
assert by_state == result
time.sleep(1)
# Check that the actual checkpoint storage (for shared_fs) reflects the
# deletions. We want to wait for the GC containers to exit, so check
# repeatedly with a timeout.
max_checks = 30
for i in range(max_checks):
time.sleep(1)
try:
for config, checkpoints in all_checkpoints:
checkpoint_config = config["checkpoint_storage"]
if checkpoint_config["type"] == "shared_fs":
deleted_exception = check.CheckFailedError
elif checkpoint_config["type"] == "s3":
deleted_exception = botocore.exceptions.ClientError
else:
raise NotImplementedError(
f'unsupported storage type {checkpoint_config["type"]}'
)
storage_manager = storage.build(checkpoint_config,
container_path=None)
for checkpoint in checkpoints:
metadata = storage.StorageMetadata.from_json(checkpoint)
if checkpoint["state"] == "COMPLETED":
with storage_manager.restore_path(metadata):
pass
elif checkpoint["state"] == "DELETED":
try:
with storage_manager.restore_path(metadata):
raise AssertionError("checkpoint not deleted")
except deleted_exception:
pass
except AssertionError:
if i == max_checks - 1:
raise
else:
break
def run_gc_checkpoints_test(checkpoint_storage: Dict[str, str]) -> None:
fixtures = [
(
conf.fixtures_path("no_op/gc_checkpoints_decreasing.yaml"),
{
(bindings.determinedexperimentv1State.STATE_COMPLETED.value):
{800, 900, 1000},
(bindings.determinedexperimentv1State.STATE_DELETED.value): {
100,
200,
300,
400,
500,
600,
700,
},
},
),
(
conf.fixtures_path("no_op/gc_checkpoints_increasing.yaml"),
{
(bindings.determinedexperimentv1State.STATE_COMPLETED.value): {
100,
200,
300,
900,
1000,
},
(bindings.determinedexperimentv1State.STATE_DELETED.value): {
400,
500,
600,
700,
800,
},
},
),
]
all_checkpoints: List[Tuple[Any, List[bindings.v1CheckpointWorkload]]] = []
for base_conf_path, result in fixtures:
config = conf.load_config(str(base_conf_path))
config["checkpoint_storage"].update(checkpoint_storage)
with tempfile.NamedTemporaryFile() as tf:
with open(tf.name, "w") as f:
yaml.dump(config, f)
experiment_id = exp.create_experiment(tf.name,
conf.fixtures_path("no_op"))
exp.wait_for_experiment_state(
experiment_id,
bindings.determinedexperimentv1State.STATE_COMPLETED)
# In some configurations, checkpoint GC will run on an auxillary machine, which may have to
# be spun up still. So we'll wait for it to run.
wait_for_gc_to_finish(experiment_id)
# Checkpoints are not marked as deleted until gc_checkpoint task starts.
retries = 5
for retry in range(retries):
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
cpoints = exp.workloads_with_checkpoint(trials[0].workloads)
sorted_checkpoints = sorted(
cpoints,
key=lambda ckp: int(ckp.totalBatches),
)
assert len(sorted_checkpoints) == 10
by_state = {} # type: Dict[str, Set[int]]
for ckpt in sorted_checkpoints:
by_state.setdefault(ckpt.state.value,
set()).add(ckpt.totalBatches)
if by_state == result:
all_checkpoints.append((config, sorted_checkpoints))
break
if retry + 1 == retries:
assert by_state == result
time.sleep(1)
# Check that the actual checkpoint storage (for shared_fs) reflects the
# deletions. We want to wait for the GC containers to exit, so check
# repeatedly with a timeout.
max_checks = 30
for i in range(max_checks):
time.sleep(1)
try:
storage_states = []
for config, checkpoints in all_checkpoints:
checkpoint_config = config["checkpoint_storage"]
storage_manager = storage.build(checkpoint_config,
container_path=None)
storage_state = {} # type: Dict[str, Any]
for checkpoint in checkpoints:
assert checkpoint.uuid is not None
storage_id = checkpoint.uuid
storage_state[storage_id] = {}
if checkpoint.state == bindings.determinedcheckpointv1State.STATE_COMPLETED:
storage_state[storage_id]["found"] = False
try:
with storage_manager.restore_path(storage_id):
storage_state[storage_id]["found"] = True
except errors.CheckpointNotFound:
pass
elif checkpoint.state == bindings.determinedcheckpointv1State.STATE_DELETED:
storage_state[storage_id] = {
"deleted": False,
"checkpoint": checkpoint
}
try:
with storage_manager.restore_path(storage_id):
pass
except errors.CheckpointNotFound:
storage_state[storage_id]["deleted"] = True
storage_states.append(storage_state)
for storage_state in storage_states:
for state in storage_state.values():
if state.get("deleted", None) is False:
json_states = json.dumps(storage_states)
raise AssertionError(
f"Some checkpoints were not deleted: JSON:{json_states}"
)
if state.get("found", None) is False:
json_states = json.dumps(storage_states)
raise AssertionError(
f"Some checkpoints were not found: JSON:{json_states}"
)
except AssertionError:
if i == max_checks - 1:
raise
else:
break
def test_allocation_resources_incremental_release() -> None:
"""
Start an two container experiment and ensure one container exits before the other. Ensure
resources are released and schedule-able without the other needing to be released.
"""
cleanup_exp_ids = []
try:
slots = _wait_for_slots(2)
assert len(slots) == 2
with tempfile.TemporaryDirectory() as context_dir, open(
os.path.join(context_dir, "const.yaml"), "w") as config_file:
# Launch an experiment that has one resource (docker container) that exits immediately.
config_obj = conf.load_config(
conf.fixtures_path("no_op/single.yaml"))
config_obj["resources"] = {
**config_obj.get("resources", {}),
**{
"slots": 2
}
}
config_obj["hyperparameters"] = {
**config_obj.get("hyperparameters", {}),
**{
"non_chief_exit_immediately": True
},
}
yaml.dump(config_obj, config_file)
shutil.copy(conf.fixtures_path("no_op/model_def.py"),
os.path.join(context_dir, "model_def.py"))
exp_id = exp.create_experiment(config_file.name, context_dir, None)
cleanup_exp_ids.append(exp_id)
# Wait for the experiment to start and run some.
exp.wait_for_experiment_state(
exp_id,
determinedexperimentv1State.STATE_ACTIVE,
)
exp.wait_for_experiment_active_workload(exp_id)
# And wait for exactly one of the resources to free, while one is still in use.
confirmations = 0
for _ in range(RANK_ONE_WAIT_TIME):
free_agents = list_free_agents()
if len(free_agents) == 1:
confirmations += 1
if confirmations == 2:
# Just for the race where one container has exited and the other hasn't quite yet,
# but is going to, make sure we see it at least twice.
break
# Still waiting on partial exit
time.sleep(1)
else:
pytest.fail(
"exactly one agent did not free after {} seconds".format(
RANK_ONE_WAIT_TIME))
# Ensure we can schedule on the free slot, not only that the API says its available.
exp_id_2 = exp.create_experiment(
conf.fixtures_path("no_op/single.yaml"),
conf.fixtures_path("no_op"),
None,
)
cleanup_exp_ids.append(exp_id_2)
exp.wait_for_experiment_workload_progress(exp_id_2)
exp.wait_for_experiment_state(
exp_id_2, determinedexperimentv1State.STATE_COMPLETED)
cleanup_exp_ids = cleanup_exp_ids[:-1]
# And check the hung experiment still is holding on to its hung slot.
free_agents = list_free_agents()
if len(free_agents) != 1:
pytest.fail(
f"should still have exactly one agent scheduled: {free_agents}"
)
finally:
for exp_id in cleanup_exp_ids:
bindings.post_KillExperiment(determined_test_session(), id=exp_id)
exp.wait_for_experiment_state(
exp_id, determinedexperimentv1State.STATE_CANCELED)
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)
def test_stress_agents_reconnect(steps: int, num_agents: int, should_disconnect: bool) -> None:
random.seed(42)
master_host = "localhost"
master_port = "8080"
conf.MASTER_IP = master_host
conf.MASTER_PORT = master_port
master_up([])
# Start all agents.
agents_are_up = [True] * num_agents
for i in range(num_agents):
agent_up(["--agent-name", f"agent-{i}"], fluent_offset=i)
time.sleep(3)
for _ in range(steps):
for agent_id, agent_is_up in enumerate(agents_are_up):
if random.choice([True, False]): # Flip agents status randomly.
continue
if should_disconnect:
# Can't just randomly deploy up/down due to just getting a Docker name conflict.
if agent_is_up:
agent_down(["--agent-name", f"agent-{agent_id}"])
else:
agent_up(["--agent-name", f"agent-{agent_id}"], fluent_offset=agent_id)
agents_are_up[agent_id] = not agents_are_up[agent_id]
else:
if random.choice([True, False]):
agent_disable([f"agent-{agent_id}"])
agents_are_up[agent_id] = False
else:
agent_enable([f"agent-{agent_id}"])
agents_are_up[agent_id] = True
time.sleep(10)
# Validate that our master kept track of the agent reconnect spam.
agent_list = json.loads(
subprocess.check_output(
[
"det",
"agent",
"list",
"--json",
]
).decode()
)
assert sum(agents_are_up) <= len(agent_list)
for agent in agent_list:
agent_id = int(agent["id"].replace("agent-", ""))
assert agents_are_up[agent_id] == agent["enabled"]
# Can we still schedule something?
if any(agents_are_up):
experiment_id = exp.create_experiment(
conf.fixtures_path("no_op/single-one-short-step.yaml"),
conf.fixtures_path("no_op"),
None,
)
exp.wait_for_experiment_state(
experiment_id, bindings.determinedexperimentv1State.STATE_COMPLETED
)
for agent_id in range(num_agents):
agent_down(["--agent-name", f"agent-{agent_id}"])
master_down([])
