def run_simple_tfjob(self, component):
api_client = k8s_client.ApiClient()
# Setup the ksonnet app
ks_util.setup_ks_app(self.app_dir, self.env, self.namespace, component,
self.params)
# Create the TF job
ks_cmd = ks_util.get_ksonnet_cmd(self.app_dir)
util.run([ks_cmd, "apply", self.env, "-c", component], cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name, self.namespace)
# Wait for the job to either be in Running state or a terminal state
logging.info("Wait for conditions Running, Succeeded, or Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Running", "Succeeded", "Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Current TFJob:\n %s", json.dumps(results, indent=2))
# Wait for the job to complete.
logging.info("Waiting for job to finish.")
results = tf_job_client.wait_for_job(
api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Final TFJob:\n %s", json.dumps(results, indent=2))
if not tf_job_client.job_succeeded(results):
self.failure = "Job {0} in namespace {1} in status {2}".format(
self.name, self.namespace, results.get("status", {}))
logging.error(self.failure)
return
# Check for creation failures.
creation_failures = tf_job_client.get_creation_failures_from_tfjob(
api_client, self.namespace, results)
if creation_failures:
# TODO(jlewi): Starting with
# https://github.com/kubeflow/tf-operator/pull/646 the number of events
# no longer seems to match the expected; it looks like maybe events
# are being combined? For now we just log a warning rather than an
# error.
logging.warning(creation_failures)
# Delete the TFJob.
tf_job_client.delete_tf_job(
api_client, self.namespace, self.name, version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(
api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
def run_tfjob_with_shutdown_policy(self, component, shutdown_policy):
tf_operator_util.load_kube_config()
api_client = k8s_client.ApiClient()
# Setup the ksonnet app
tf_operator_util.setup_ks_app(self.app_dir, self.env, self.namespace,
component, self.params)
# Create the TF job
ks_cmd = ks_util.get_ksonnet_cmd(self.app_dir)
util.run([ks_cmd, "apply", self.env, "-c", component],
cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name,
self.namespace)
# Wait for the job to either be in Running state or a terminal state
logging.info("Wait for conditions Running, Succeeded, or Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Running", "Succeeded", "Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Current TFJob:\n %s", json.dumps(results, indent=2))
if shutdown_policy == "worker":
tf_job_client.terminate_replicas(api_client, self.namespace,
self.name, "worker", 1)
else:
tf_job_client.terminate_replicas(api_client, self.namespace,
self.name, "chief", 1)
# Wait for the job to complete.
logging.info("Waiting for job to finish.")
results = tf_job_client.wait_for_job(
api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Final TFJob:\n %s", json.dumps(results, indent=2))
if not tf_job_client.job_succeeded(results):
self.failure = "Job {0} in namespace {1} in status {2}".format(
self.name, self.namespace, results.get("status", {}))
logging.error(self.failure)
return
# Delete the TFJob.
tf_job_client.delete_tf_job(api_client,
self.namespace,
self.name,
version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
def run_distributed_training_job(self, component):
api_client = k8s_client.ApiClient()
# Setup the ksonnet app
ks_util.setup_ks_app(self.app_dir, self.env, self.namespace, component,
self.params)
# Create the TF job
ks_cmd = ks_util.get_ksonnet_cmd(self.app_dir)
util.run([ks_cmd, "apply", self.env, "-c", component],
cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name,
self.namespace)
# Wait for the job to either be in Running state or a terminal state
logging.info("Wait for conditions Running, Succeeded, or Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Running", "Succeeded", "Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Current TFJob:\n %s", json.dumps(results, indent=2))
# Wait for the job to complete.
logging.info("Waiting for job to finish.")
results = tf_job_client.wait_for_job(
api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Final TFJob:\n %s", json.dumps(results, indent=2))
if not tf_job_client.job_succeeded(results):
self.failure = "Job {0} in namespace {1} in status {2}".format(
self.name, self.namespace, results.get("status", {}))
logging.error(self.failure)
return
# Check for creation failures.
creation_failures = tf_job_client.get_creation_failures_from_tfjob(
api_client, self.namespace, results)
if creation_failures:
logging.warning(creation_failures)
# Delete the TFJob.
tf_job_client.delete_tf_job(api_client,
self.namespace,
self.name,
version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
def test_tf_job_simple(test_case): # pylint: disable=redefined-outer-name
args = parse_args()
namespace = "default"
name = "tf-job-simple"
util.load_kube_config()
api_client = k8s_client.ApiClient()
create_app_and_job(args, namespace, name)
try:
tf_job_client.wait_for_condition(
api_client,
namespace,
name, ["Running"],
status_callback=tf_job_client.log_status)
logging.info("TFJob launched successfully")
except Exception as e:
logging.error("Test failed waiting for job; %s", e)
test_case.add_failure_info(e.message)
def test_invalid_tfjob_spec(self):
api_client = k8s_client.ApiClient()
component = INVALID_TFJOB_COMPONENT_NAME + "_" + self.tfjob_version
# Setup the ksonnet app
ks_util.setup_ks_app(self.app_dir, self.env, self.namespace, component,
self.params)
# Create the TF job
ks_cmd = ks_util.get_ksonnet_cmd(self.app_dir)
util.run([ks_cmd, "apply", self.env, "-c", component],
cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name,
self.namespace)
logging.info("Wait for conditions Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Final TFJob:\n %s", json.dumps(results, indent=2))
last_condition = results.get("status", {}).get("conditions", [{}])[-1]
if last_condition.get("type", "").lower() != "failed":
self.failure = "Job {0} in namespace {1} did not fail; status {2}".format(
self.name, self.namespace, results.get("status", {}))
logging.error(self.failure)
return
pattern = ".*the spec is invalid.*"
condition_message = last_condition.get("message", "")
if not re.match(pattern, condition_message):
self.failure = "Condition message {0} did not match pattern {1}".format(
condition_message, pattern)
logging.error(self.failure)
# Delete the TFJob.
tf_job_client.delete_tf_job(api_client,
self.namespace,
self.name,
version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
def test_tfjob_and_verify_runconfig(self):
api_client = k8s_client.ApiClient()
masterHost = api_client.configuration.host
component = COMPONENT_NAME + "_" + self.tfjob_version
# Setup the ksonnet app
ks_util.setup_ks_app(self.app_dir, self.env, self.namespace, component,
self.params)
# Create the TF job
ks_cmd = ks_util.get_ksonnet_cmd(self.app_dir)
util.run([ks_cmd, "apply", self.env, "-c", component],
cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name,
self.namespace)
# Wait for the job to either be in Running state or a terminal state
logging.info("Wait for conditions Running, Succeeded, or Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Running", "Succeeded", "Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Current TFJob:\n %s", json.dumps(results, indent=2))
num_ps = results.get("spec", {}).get("tfReplicaSpecs",
{}).get("PS",
{}).get("replicas", 0)
num_workers = results.get("spec",
{}).get("tfReplicaSpecs",
{}).get("Worker",
{}).get("replicas", 0)
verify_runconfig(masterHost, self.namespace, self.name, "chief",
num_ps, num_workers)
verify_runconfig(masterHost, self.namespace, self.name, "worker",
num_ps, num_workers)
verify_runconfig(masterHost, self.namespace, self.name, "ps", num_ps,
num_workers)
verify_runconfig(masterHost, self.namespace, self.name, "evaluator",
num_ps, num_workers)
tf_job_client.terminate_replicas(api_client, self.namespace, self.name,
"chief", 1)
# Wait for the job to complete.
logging.info("Waiting for job to finish.")
results = tf_job_client.wait_for_job(
api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Final TFJob:\n %s", json.dumps(results, indent=2))
if not tf_job_client.job_succeeded(results):
self.failure = "Job {0} in namespace {1} in status {2}".format(
self.name, self.namespace, results.get("status", {}))
logging.error(self.failure)
# Delete the TFJob.
tf_job_client.delete_tf_job(api_client,
self.namespace,
self.name,
version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
def run_tfjob_with_cleanpod_policy(self, component, clean_pod_policy):
api_client = k8s_client.ApiClient()
# Setup the ksonnet app
ks_util.setup_ks_app(self.app_dir, self.env, self.namespace, component,
self.params)
# Create the TF job
util.run(["ks", "apply", self.env, "-c", component], cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name,
self.namespace)
# Wait for the job to either be in Running state or a terminal state
logging.info("Wait for conditions Running, Succeeded, or Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Running", "Succeeded", "Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Current TFJob:\n %s", json.dumps(results, indent=2))
# Wait for the job to complete.
logging.info("Waiting for job to finish.")
results = tf_job_client.wait_for_job(
api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Final TFJob:\n %s", json.dumps(results, indent=2))
if not tf_job_client.job_succeeded(results):
self.failure = "Job {0} in namespace {1} in status {2}".format(
self.name, self.namespace, results.get("status", {}))
logging.error(self.failure)
return
# All pods are deleted.
if clean_pod_policy == "All":
pod_labels = tf_job_client.get_labels(self.name)
pod_selector = tf_job_client.to_selector(pod_labels)
k8s_util.wait_for_pods_to_be_deleted(api_client, self.namespace,
pod_selector)
# Only running pods (PS) are deleted, completed pods are not.
elif clean_pod_policy == "Running":
tf_job_client.wait_for_replica_type_in_phases(
api_client, self.namespace, self.name, "Chief", ["Succeeded"])
tf_job_client.wait_for_replica_type_in_phases(
api_client, self.namespace, self.name, "Worker", ["Succeeded"])
pod_labels = tf_job_client.get_labels(self.name, "PS")
pod_selector = tf_job_client.to_selector(pod_labels)
k8s_util.wait_for_pods_to_be_deleted(api_client, self.namespace,
pod_selector)
# No pods are deleted.
elif clean_pod_policy == "None":
tf_job_client.wait_for_replica_type_in_phases(
api_client, self.namespace, self.name, "Chief", ["Succeeded"])
tf_job_client.wait_for_replica_type_in_phases(
api_client, self.namespace, self.name, "Worker", ["Succeeded"])
tf_job_client.wait_for_replica_type_in_phases(
api_client, self.namespace, self.name, "PS", ["Running"])
# Delete the TFJob.
tf_job_client.delete_tf_job(api_client,
self.namespace,
self.name,
version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
def run_tfjob_with_replica_restart_policy(self, component,
replica_restart_policy,
exit_code):
tf_operator_util.load_kube_config()
api_client = k8s_client.ApiClient()
# Setup the ksonnet app
tf_operator_util.setup_ks_app(self.app_dir, self.env, self.namespace,
component, self.params)
# Create the TF job
ks_cmd = ks_util.get_ksonnet_cmd(self.app_dir)
util.run([ks_cmd, "apply", self.env, "-c", component],
cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name,
self.namespace)
# Wait for the job to either be in Running state or a terminal state
logging.info("Wait for conditions Running, Succeeded, or Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Running", "Succeeded", "Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Current TFJob:\n %s", json.dumps(results, indent=2))
if replica_restart_policy == "Always" and exit_code == 0:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
True)
elif replica_restart_policy == "Always" and exit_code == 1:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
True)
elif replica_restart_policy == "OnFailure" and exit_code == 1:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
True)
elif replica_restart_policy == "OnFailure" and exit_code == 0:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
False)
elif replica_restart_policy == "Never" and exit_code == 1:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
False)
elif replica_restart_policy == "Never" and exit_code == 0:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
False)
elif replica_restart_policy == "ExitCode" and exit_code == 1:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
False)
else:
res = tf_job_client.terminate_and_verify_start_time(
api_client, self.namespace, self.name, "ps", 0, exit_code,
True)
if res is False:
self.failure = "Job {0} in namespace {1} with restart policy {2} failed test \
with exit_code {3}".format(self.name, self.namespace,
replica_restart_policy, exit_code)
logging.error(self.failure)
return
# Delete the TFJob.
tf_job_client.delete_tf_job(api_client,
self.namespace,
self.name,
version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
def test_pod_names(self):
api_client = k8s_client.ApiClient()
component = COMPONENT_NAME + "_" + self.tfjob_version
ks_util.setup_ks_app(self.app_dir, self.env, self.namespace, component,
self.params)
util.run(["ks", "apply", self.env, "-c", component], cwd=self.app_dir)
logging.info("Created job %s in namespaces %s", self.name,
self.namespace)
logging.info("Wait for conditions Running, Succeeded, or Failed")
results = tf_job_client.wait_for_condition(
api_client,
self.namespace,
self.name, ["Running", "Succeeded", "Failed"],
version=self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Current TFJob:\n %s", json.dumps(results, indent=2))
job_specs = extract_job_specs(
results.get("spec", {}).get("tfReplicaSpecs", {}))
expected_pod_names = []
for replica_type, replica_num in job_specs.items():
logging.info("job_type = %s, replica = %s", replica_type,
replica_num)
for i in range(replica_num):
expected_pod_names.append(
POD_NAME_FORMAT.format(name=self.name,
replica=replica_type,
index=i))
expected_pod_names = set(expected_pod_names)
actual_pod_names = tf_job_client.get_pod_names(api_client,
self.namespace,
self.name)
# We are not able to guarantee pods selected with default namespace and job
# name are only for this test run only. Therefore we only do partial check,
# e.g. make sure expected set of pod names are in the selected pod names.
if not (expected_pod_names & actual_pod_names) == expected_pod_names:
msg = "Actual pod names doesn't match. Expected: {0} Actual: {1}".format(
str(expected_pod_names), str(actual_pod_names))
logging.error(msg)
raise RuntimeError(msg)
tf_job_client.terminate_replicas(api_client, self.namespace, self.name,
"chief", 1)
# Wait for the job to complete.
logging.info("Waiting for job to finish.")
results = tf_job_client.wait_for_job(
api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
logging.info("Final TFJob:\n %s", json.dumps(results, indent=2))
if not tf_job_client.job_succeeded(results):
self.failure = "Job {0} in namespace {1} in status {2}".format(
self.name, self.namespace, results.get("status", {}))
logging.error(self.failure)
# Delete the TFJob.
tf_job_client.delete_tf_job(api_client,
self.namespace,
self.name,
version=self.tfjob_version)
logging.info("Waiting for job %s in namespaces %s to be deleted.",
self.name, self.namespace)
tf_job_client.wait_for_delete(api_client,
self.namespace,
self.name,
self.tfjob_version,
status_callback=tf_job_client.log_status)
