def patch_jobs(active_jobs=None, desired_jobs=None):
"""
patch jobs check current state of the job and applies desired goal state
for the job. It can yield to create a job or updating a job.
"""
jobs = {}
for job_name, job_spec in desired_jobs.items():
if job_name in active_jobs.keys():
# job exists -> update to desired state
patch_job(active_jobs[job_name], job_spec)
jobs[job_name] = active_jobs[job_name].get_job_id()
else:
# job does not exist -> create
job = StatelessJob(job_config=job_spec,
config=IntegrationTestConfig(
pool_file=RESPOOL_FILE_NAME,
max_retry_attempts=MAX_RETRY_ATTEMPTS))
job.create()
time.sleep(10)
job.wait_for_all_pods_running()
jobs[job_name] = job.get_job_id()
# TODO: Kill any undesired active job running in the canary cluster
return jobs
def test__placement_exclusive_job(exclusive_host):
excl_constraint = pod_pb2.Constraint(
type=1, # Label constraint
label_constraint=pod_pb2.LabelConstraint(
kind=2, # Host
condition=2, # Equal
requirement=1,
label=peloton_pb2_v1alpha.Label(key="peloton/exclusive",
value="exclusive-test-label"),
),
)
# We have 1 exclusive host and 2 non-exclusive hosts. Set number of
# instances to be a few more than what can run simulatenously on
# a single exclusive host
job = StatelessJob(job_file="test_stateless_job_cpus_large_spec.yaml")
job.job_spec.default_spec.constraint.CopyFrom(excl_constraint)
job.job_spec.instance_count = 6
job.create()
job.wait_for_state(goal_state="RUNNING")
job.wait_for_all_pods_running(num_pods=4)
job.stop()
job.wait_for_terminated()
# check that all of them ran on exclusive host
pod_summaries = job.list_pods()
for s in pod_summaries:
if s.status.host:
assert "exclusive" in s.status.host
def test__in_place_kill_job_release_host():
job1 = StatelessJob(
job_file="test_stateless_job_spec.yaml",
)
job1.create()
job1.wait_for_state(goal_state="RUNNING")
job2 = StatelessJob(
job_file="test_stateless_job_spec.yaml",
)
job2.create()
job2.wait_for_state(goal_state="RUNNING")
update1 = StatelessUpdate(job1,
updated_job_file=UPDATE_STATELESS_JOB_SPEC,
batch_size=0)
update1.create(in_place=True)
# stop the update
job1.stop()
update2 = StatelessUpdate(job2,
updated_job_file=UPDATE_STATELESS_JOB_SPEC,
batch_size=0)
update2.create()
# both updates should complete
update1.wait_for_state(goal_state="SUCCEEDED")
update2.wait_for_state(goal_state="SUCCEEDED")
def test__simple_revocable_batch_and_stateless_colocate():
revocable_stateless_job = StatelessJob(
job_file="test_stateless_job_revocable_spec.yaml")
revocable_stateless_job.create()
revocable_stateless_job.wait_for_state(goal_state="RUNNING")
revocable_stateless_job.wait_for_all_pods_running()
non_revocable_stateless_job = StatelessJob(
job_file="test_stateless_job_spec.yaml")
non_revocable_stateless_job.create()
non_revocable_stateless_job.wait_for_state(goal_state="RUNNING")
non_revocable_stateless_job.wait_for_all_pods_running()
revocable_batch_job = Job(
job_file="test_job_revocable.yaml",
config=IntegrationTestConfig(pool_file='test_stateless_respool.yaml'),
)
revocable_batch_job.create()
revocable_batch_job.wait_for_state(goal_state="RUNNING")
non_revocable_batch_job = Job(
job_file="test_job.yaml",
config=IntegrationTestConfig(pool_file='test_stateless_respool.yaml'),
)
non_revocable_batch_job.create()
non_revocable_batch_job.wait_for_state(goal_state="RUNNING")
revocable_batch_job.wait_for_state()
non_revocable_batch_job.wait_for_state()
def test__delete_running_job_without_force_flag():
job = StatelessJob()
job.create()
job.wait_for_state(goal_state="RUNNING")
try:
job.delete()
except grpc.RpcError as e:
assert e.code() == grpc.StatusCode.ABORTED
return
raise Exception("job in non-terminal state error not received")
def test__create_update_to_unset_health_check():
job = StatelessJob(job_file=UPDATE_STATELESS_JOB_WITH_HEALTH_CHECK_SPEC,
config=IntegrationTestConfig(max_retry_attempts=100))
job.create()
job.wait_for_state(goal_state='RUNNING')
update = StatelessUpdate(job,
updated_job_file=UPDATE_STATELESS_JOB_SPEC,
max_failure_instances=1,
max_instance_attempts=1)
update.create()
update.wait_for_state(goal_state='SUCCEEDED')
def test__create_update_to_disable_health_check():
job = StatelessJob(job_file=UPDATE_STATELESS_JOB_WITH_HEALTH_CHECK_SPEC,
config=IntegrationTestConfig(max_retry_attempts=100))
job.create()
job.wait_for_state(goal_state='RUNNING')
job.job_spec.default_spec.containers[0].liveness_check.enabled = False
update = StatelessUpdate(job,
updated_job_spec=job.job_spec,
max_failure_instances=1,
max_instance_attempts=1)
update.create()
update.wait_for_state(goal_state='SUCCEEDED')
def test__kill_sla_violated_job():
"""
1. Create a stateless job(instance_count=5) with host-limit-1 constraint and
MaximumUnavailableInstances=1. Since there are only 3 UP hosts, 2 of
the instances will not get placed (hence unavailable).
2. Kill job and wait for the job to reach KILLED state
"""
job = StatelessJob(job_file="test_stateless_job_spec_sla.yaml", )
job.job_spec.instance_count = 5
job.create()
job.wait_for_all_pods_running(num_pods=3)
job.stop()
job.wait_for_state(goal_state='KILLED')
def test__health_check_detects_healthy_tasks():
job = StatelessJob(
job_file='test_stateless_job_successful_health_check_spec.yaml',
config=IntegrationTestConfig(max_retry_attempts=100))
job.job_spec.instance_count = 1
job.create()
job.wait_for_state(goal_state='RUNNING')
def task_has_healthy_events():
for pod_event in job.get_pod(0).get_pod_events():
if pod_event.healthy == 'HEALTHY':
return True
job.wait_for_condition(task_has_healthy_events)
def test__delete_running_job_with_force_flag():
job = StatelessJob()
job.create()
job.wait_for_state(goal_state="RUNNING")
job.delete(force_delete=True)
time.sleep(10)
try:
job.get_job()
except grpc.RpcError as e:
assert e.code() == grpc.StatusCode.NOT_FOUND
return
raise Exception("job not found error not received")
def test__delete_initialized_job_with_force_flag():
job = StatelessJob()
job.create()
# the job might have transitioned to INITIALIZED/PENDING
# since there is no way to fine control the job transitions
job.delete(force_delete=True)
time.sleep(10)
try:
job.get_job()
except grpc.RpcError as e:
assert e.code() == grpc.StatusCode.NOT_FOUND
return
raise Exception("job not found error not received")
def test__failed_task_throttled_by_exponential_backoff():
job = StatelessJob(job_file='test_stateless_job_exit_with_err_spec.yaml',
config=IntegrationTestConfig(max_retry_attempts=100))
job.create()
job.wait_for_state(goal_state='RUNNING')
time.sleep(40)
pod_events = job.get_pod(0).get_pod_events()
# if throttle is effective, the task should not create many
# pod events. Otherwise it can generate many pod events, during
# the time window
pod_id = pod_events[0].pod_id.value
run_id = int(pod_id[pod_id.rindex('-') + 1:])
assert 1 < run_id < 20
def test__restart_killed_job():
job = StatelessJob(job_file="test_stateless_job_spec_k8s.yaml")
job.create()
job.wait_for_state(goal_state="RUNNING")
old_pod_infos = job.query_pods()
job.stop()
job.wait_for_state(goal_state="KILLED")
job.restart(in_place=False)
job.wait_for_all_pods_running()
new_pod_infos = job.query_pods()
assert_pod_id_changed(old_pod_infos, new_pod_infos)
def test__delete_job_bad_version():
job = StatelessJob()
job.create()
job.wait_for_state(goal_state="RUNNING")
job.stop()
job.wait_for_state(goal_state="KILLED")
try:
job.delete(entity_version="1-2-3")
except grpc.RpcError as e:
assert e.code() == grpc.StatusCode.ABORTED
assert INVALID_ENTITY_VERSION_ERR_MESSAGE in e.details()
return
raise Exception("entity version mismatch error not received")
def test__health_check_detects_unhealthy_tasks():
job = StatelessJob(
job_file="test_stateless_job_failed_health_check_spec.yaml",
config=IntegrationTestConfig(max_retry_attempts=100),
)
job.job_spec.instance_count = 1
job.create()
job.wait_for_state(goal_state="RUNNING")
def task_has_unhealthy_events():
for pod_event in job.get_pod(0).get_pod_events():
if pod_event.healthy == "HEALTH_STATE_UNHEALTHY":
return True
job.wait_for_condition(task_has_unhealthy_events)
def test_update_killed_job(in_place):
job = StatelessJob(job_file=UPDATE_STATELESS_JOB_ADD_INSTANCES_SPEC)
job.create()
job.wait_for_state(goal_state="RUNNING")
job.stop()
job.wait_for_state(goal_state="KILLED")
update = StatelessUpdate(
job,
updated_job_file=UPDATE_STATELESS_JOB_UPDATE_REDUCE_INSTANCES_SPEC)
update.create(in_place=in_place)
update.wait_for_state(goal_state="SUCCEEDED")
assert job.get_spec().instance_count == 3
assert job.get_status().state == stateless_pb2.JOB_STATE_KILLED
def patch_jobs(active_jobs=None, desired_jobs=None):
"""
patch jobs check current state of the job and applies desired goal state
for the job. It can yield to create a job or updating a job.
"""
jobs = {}
for job_name, job_spec in desired_jobs.items():
if job_name in active_jobs.keys():
j = active_jobs[job_name]
# failfast is not None then do not run canary test
# until dirty jobs are restored manually.
if os.getenv("FAILFAST") == "NO":
# job exists -> update to desired state
patch_job(j, job_spec)
jobs[job_name] = j.get_job_id()
else:
# if job update diff has non-nil result means that previous
# canary test run failed and we want more runs to block
# until issue is manually debugged and state is restored.
job_spec.respool_id.value = j.get_spec().respool_id.value
resp = j.get_replace_job_diff(job_spec=job_spec)
print resp
if len(resp.instances_removed) > 0 or \
len(resp.instances_updated) > 0 or \
len(resp.instances_added) > 0:
pytest.exit(
"canary test run was aborted, since jobs are dirty!!")
jobs[job_name] = j.get_job_id()
else:
# job does not exist -> create
job = StatelessJob(
job_config=job_spec,
config=IntegrationTestConfig(
pool_file=RESPOOL_FILE_NAME,
max_retry_attempts=MAX_RETRY_ATTEMPTS,
),
)
job.create()
time.sleep(10)
job.wait_for_all_pods_running()
jobs[job_name] = job.get_job_id()
# TODO: Kill any undesired active job running in the canary cluster
return jobs
def test__delete_killed_job():
job = StatelessJob()
job.create()
job.wait_for_state(goal_state="RUNNING")
job.stop()
job.wait_for_state(goal_state="KILLED")
job.delete()
time.sleep(10)
try:
job.get_job()
except grpc.RpcError as e:
assert e.code() == grpc.StatusCode.NOT_FOUND
return
raise Exception("job not found error not received")
def test__failed_task_automatically_restart():
job = StatelessJob(job_file='test_stateless_job_exit_with_err_spec.yaml',
config=IntegrationTestConfig(max_retry_attempts=100))
job.create()
job.wait_for_state(goal_state='RUNNING')
old_pod_id = job.get_pod(0).get_pod_status().pod_id.value
def job_not_running():
return job.get_status().state != 'JOB_STATE_RUNNING'
job.wait_for_condition(job_not_running)
def pod_id_changed():
new_pod_id = job.get_pod(0).get_pod_status().pod_id.value
return old_pod_id != new_pod_id
job.wait_for_condition(pod_id_changed)
def test__placement_non_exclusive_job(exclusive_host):
# We have 1 exclusive host and 2 non-exclusive hosts. Set number of
# instances to be a few more than what can run simulatenously
# on 2 non-exclusive hosts
job = StatelessJob(job_file="test_stateless_job_cpus_large_spec.yaml")
job.job_spec.instance_count = 10
job.create()
job.wait_for_state(goal_state="RUNNING")
job.wait_for_all_pods_running(num_pods=5)
job.stop()
job.wait_for_terminated()
# check that none of them ran on exclusive host
pod_summaries = job.list_pods()
for s in pod_summaries:
if s.status.host:
assert "exclusive" not in s.status.host
def test_stop_running_job_with_active_update_remove_instances(in_place):
stateless_job = StatelessJob(
job_file=UPDATE_STATELESS_JOB_ADD_INSTANCES_SPEC)
stateless_job.create()
stateless_job.wait_for_state(goal_state="RUNNING")
assert len(stateless_job.query_pods()) == 5
update = StatelessUpdate(
stateless_job,
updated_job_file=UPDATE_STATELESS_JOB_UPDATE_REDUCE_INSTANCES_SPEC,
batch_size=1,
)
update.create(in_place=in_place)
update.wait_for_state(goal_state="ROLLING_FORWARD")
stateless_job.stop()
update.wait_for_state(goal_state="SUCCEEDED")
assert stateless_job.get_spec().instance_count == 3
def test__create_update_to_set_health_check(in_place):
job = StatelessJob(
job_file=UPDATE_STATELESS_JOB_SPEC,
config=IntegrationTestConfig(
max_retry_attempts=100,
pool_file='test_stateless_respool.yaml',
),
)
job.create()
job.wait_for_state(goal_state="RUNNING")
update = StatelessUpdate(
job,
updated_job_file=UPDATE_STATELESS_JOB_WITH_HEALTH_CHECK_SPEC,
max_failure_instances=1,
max_instance_attempts=1,
)
update.create(in_place=in_place)
update.wait_for_state(goal_state="SUCCEEDED")
def test__create_update_to_change_health_check_config(in_place):
job = StatelessJob(
job_file=UPDATE_STATELESS_JOB_WITH_HEALTH_CHECK_SPEC,
config=IntegrationTestConfig(max_retry_attempts=100),
)
job.job_spec.default_spec.containers[0].liveness_check.enabled = False
job.create()
job.wait_for_state(goal_state="RUNNING")
job.job_spec.default_spec.containers[
0].liveness_check.initial_interval_secs = 2
update = StatelessUpdate(
job,
updated_job_spec=job.job_spec,
max_failure_instances=1,
max_instance_attempts=1,
)
update.create(in_place=in_place)
update.wait_for_state(goal_state="SUCCEEDED")
def test__revocable_job_slack_limit():
revocable_job = StatelessJob(
job_file="test_stateless_job_revocable_slack_limit_spec.yaml")
revocable_job.create()
revocable_job.wait_for_state(goal_state="RUNNING")
# 2 tasks are running out of 3
def partial_tasks_running():
count = 0
for pod_id in range(0, revocable_job.job_spec.instance_count):
pod_state = revocable_job.get_pod(pod_id).get_pod_status().state
if pod_state == pod.POD_STATE_RUNNING:
count += 1
return count == 2
revocable_job.wait_for_condition(partial_tasks_running)
# cleanup job from jobmgr
revocable_job.stop()
revocable_job.wait_for_terminated()
def test__in_place_update_success_rate():
stateless_job = StatelessJob(job_file="test_stateless_job_spec_k8s.yaml")
stateless_job.job_spec.instance_count = 30
stateless_job.create()
stateless_job.wait_for_all_pods_running()
old_pod_infos = stateless_job.query_pods()
job_spec_dump = load_test_config("test_stateless_job_spec_k8s.yaml")
updated_job_spec = JobSpec()
json_format.ParseDict(job_spec_dump, updated_job_spec)
updated_job_spec.instance_count = 30
updated_job_spec.default_spec.containers[0].resource.mem_limit_mb = 0.1
update = StatelessUpdate(stateless_job,
updated_job_spec=updated_job_spec,
batch_size=0)
update.create(in_place=True)
update.wait_for_state(goal_state='SUCCEEDED')
new_pod_infos = stateless_job.query_pods()
old_pod_dict = {}
new_pod_dict = {}
for old_pod_info in old_pod_infos:
split_index = old_pod_info.status.pod_id.value.rfind('-')
pod_name = old_pod_info.status.pod_id.value[:split_index]
old_pod_dict[pod_name] = old_pod_info.status.host
for new_pod_info in new_pod_infos:
split_index = new_pod_info.status.pod_id.value.rfind('-')
pod_name = new_pod_info.status.pod_id.value[:split_index]
new_pod_dict[pod_name] = new_pod_info.status.host
count = 0
for pod_name, pod_id in old_pod_dict.items():
if new_pod_dict[pod_name] != old_pod_dict[pod_name]:
log.info("%s, prev:%s, cur:%s", pod_name, old_pod_dict[pod_name],
new_pod_dict[pod_name])
count = count + 1
assert count == 0
def test__delete_sla_violated_job():
"""
1. Create a stateless job(instance_count=5) with host-limit-1 constraint and
MaximumUnavailableInstances=1. Since there are only 3 UP hosts, 2 of
the instances will not get placed (hence unavailable).
2. Force delete the job and verify that the job is deleted
"""
job = StatelessJob(job_file="test_stateless_job_spec_sla.yaml", )
job.job_spec.instance_count = 5
job.create()
job.wait_for_all_pods_running(num_pods=3)
job.delete(force_delete=True)
time.sleep(10)
try:
job.get_job()
except grpc.RpcError as e:
assert e.code() == grpc.StatusCode.NOT_FOUND
return
raise Exception("job not found error not received")
def test__revocable_tasks_move_to_revocable_queue():
revocable_job1 = StatelessJob(job_file='test_stateless_job_revocable_spec.yaml')
revocable_job1.create()
revocable_job1.wait_for_state(goal_state='RUNNING')
revocable_job1.wait_for_all_pods_running()
# 1 task is running out of 3
def partial_tasks_running():
count = 0
for pod_id in range(0, revocable_job2.job_spec.instance_count):
pod_state = revocable_job2.get_pod(pod_id).get_pod_status().state
if pod_state == pod.POD_STATE_RUNNING:
count += 1
return count == 1
revocable_job2 = StatelessJob(
job_file='test_stateless_job_revocable_slack_limit_spec.yaml')
revocable_job2.create()
# sleep for 5 seconds to make sure job has enough time
time.sleep(5)
revocable_job2.wait_for_condition(partial_tasks_running)
non_revocable_job = StatelessJob(job_file='test_stateless_job_spec.yaml')
non_revocable_job.create()
non_revocable_job.wait_for_state('RUNNING')
non_revocable_job.wait_for_all_pods_running()
# cleanup jobs from jobmgr
revocable_job1.stop()
revocable_job2.stop()
non_revocable_job.stop()
revocable_job1.wait_for_terminated()
revocable_job2.wait_for_terminated()
non_revocable_job.wait_for_terminated()
def test__create_revocable_job():
revocable_job1 = StatelessJob(
job_file="test_stateless_job_revocable_spec.yaml")
revocable_job1.create()
revocable_job1.wait_for_state(goal_state="RUNNING")
revocable_job1.wait_for_all_pods_running()
revocable_job2 = StatelessJob(
job_file="test_stateless_job_revocable_spec.yaml")
revocable_job2.create()
revocable_job2.wait_for_state(goal_state="RUNNING")
revocable_job2.wait_for_all_pods_running()
non_revocable_job = StatelessJob(
job_file="test_stateless_job_cpus_large_spec.yaml")
non_revocable_job.create()
non_revocable_job.wait_for_state(goal_state="RUNNING")
non_revocable_job.wait_for_all_pods_running()
# cleanup jobs from jobmgr
revocable_job1.stop()
revocable_job2.stop()
non_revocable_job.stop()
revocable_job1.wait_for_terminated()
revocable_job2.wait_for_terminated()
non_revocable_job.wait_for_terminated()
def test__in_place_update_multi_component_restart(jobmgr, resmgr, hostmgr,
placement_engines,
batch_size):
# need extra retry attempts, since in-place update would need more time
# to process given hostmgr would be restarted
job1 = StatelessJob(
job_file="test_stateless_job_spec.yaml",
config=IntegrationTestConfig(max_retry_attempts=300),
)
job1.create()
job1.wait_for_all_pods_running()
job2 = StatelessJob(
job_file="test_stateless_job_spec.yaml",
config=IntegrationTestConfig(max_retry_attempts=300),
)
job2.create()
job2.wait_for_all_pods_running()
update1 = StatelessUpdate(job1,
updated_job_file=UPDATE_STATELESS_JOB_SPEC,
batch_size=batch_size)
update1.create(in_place=True)
update2 = StatelessUpdate(job2,
updated_job_file=UPDATE_STATELESS_JOB_SPEC,
batch_size=batch_size)
update2.create()
jobmgr.restart()
time.sleep(random.randint(1, 10))
resmgr.restart()
time.sleep(random.randint(1, 10))
hostmgr.restart()
time.sleep(random.randint(1, 10))
placement_engines.restart()
update1.wait_for_state(goal_state="SUCCEEDED")
update2.wait_for_state(goal_state="SUCCEEDED")
def test__stop_nonrevocable_job_to_free_resources_for_revocable_job():
non_revocable_job1 = StatelessJob(
job_file="test_stateless_job_memory_large_spec.yaml"
)
non_revocable_job1.create()
non_revocable_job1.wait_for_state("RUNNING")
non_revocable_job2 = StatelessJob(
job_file="test_stateless_preemptible_job_memory_large_spec.yaml"
)
non_revocable_job2.create()
non_revocable_job2.wait_for_state("RUNNING")
non_revocable_job1.wait_for_all_pods_running()
non_revocable_job2.wait_for_all_pods_running()
revocable_job = StatelessJob(
job_file="test_stateless_job_revocable_spec.yaml"
)
revocable_job.create()
# no tasks should be running
def no_task_running():
count = 0
for pod_id in range(0, revocable_job.job_spec.instance_count):
pod_state = revocable_job.get_pod(pod_id).get_pod_status().state
if pod_state == pod.POD_STATE_RUNNING:
count += 1
return count == 0
# give job 5 seconds to run, even after that no tasks should be running
time.sleep(5)
revocable_job.wait_for_condition(no_task_running)
# stop non_revocable job to free up resources for revocable job
non_revocable_job2.stop()
non_revocable_job2.wait_for_terminated()
# After non_revocable job is killed, all revocable tasks should be running
revocable_job.wait_for_all_pods_running()
# cleanup jobs from jobmgr
non_revocable_job1.stop()
revocable_job.stop()
non_revocable_job1.wait_for_terminated()
revocable_job.wait_for_terminated()