def run(containers, node, persist):
container_ids = RedisJobContainers.get_containers()
gpu_resources = get_gpu_resources()
if gpu_resources:
gpu_resources = {gpu_resource['index']: gpu_resource for gpu_resource in gpu_resources}
update_cluster_node(gpu_resources)
for container_id in container_ids:
container = get_container(containers, container_id)
if not container:
continue
try:
payload = get_container_resources(node, containers[container_id], gpu_resources)
except KeyError:
payload = None
if payload:
payload = payload.to_dict()
# todo: Re-enable publishing
# logger.debug("Publishing resources event")
# celery_app.send_task(
# K8SEventsCeleryTasks.K8S_EVENTS_HANDLE_RESOURCES,
# kwargs={'payload': payload, 'persist': persist})
job_uuid = payload['job_uuid']
# Check if we should stream the payload
# Check if we have this container already in place
experiment_uuid = RedisJobContainers.get_experiment_for_job(job_uuid)
set_last_resources_cond = (
RedisToStream.is_monitored_job_resources(job_uuid) or
RedisToStream.is_monitored_experiment_resources(experiment_uuid))
if set_last_resources_cond:
RedisToStream.set_latest_job_resources(job_uuid, payload)
def update_job_containers(event: Mapping, status: str,
job_container_name: str) -> None:
job_containers = RedisJobContainers()
if JobLifeCycle.is_done(status):
# Remove the job monitoring
job_uuid = event['metadata']['labels']['job_uuid']
logger.info('Stop monitoring job_uuid: %s', job_uuid)
job_containers.remove_job(job_uuid)
if event['status']['container_statuses'] is None:
return
def get_container_id(container_id):
if not container_id:
return None
if container_id.startswith('docker://'):
return container_id[len('docker://'):]
return container_id
for container_status in event['status']['container_statuses']:
if container_status['name'] != job_container_name:
continue
container_id = get_container_id(container_status['container_id'])
if container_id:
job_uuid = event['metadata']['labels']['job_uuid']
if container_status['state']['running'] is not None:
logger.info('Monitoring (container_id, job_uuid): (%s, %s)',
container_id, job_uuid)
job_containers.monitor(container_id=container_id,
job_uuid=job_uuid)
else:
job_containers.remove_container(container_id=container_id)
def run(cls):
results = {}
result = cls.redis_health(RedisEphemeralTokens.connection())
if not result.is_healthy:
results['REDIS_EPH_TOKENS'] = result
result = cls.redis_health(RedisSessions.connection())
if not result.is_healthy:
results['REDIS_SESSIONS'] = result
result = cls.redis_health(RedisTTL.connection())
if not result.is_healthy:
results['REDIS_TTL'] = result
result = cls.redis_health(RedisToStream.connection())
if not result.is_healthy:
results['REDIS_TO_STREAM'] = result
result = cls.redis_health(RedisJobContainers.connection())
if not result.is_healthy:
results['REDIS_CONTAINERS'] = result
if not results:
results = {'REDIS': Result()}
return results
def experiment_job_status_post_save(sender, **kwargs):
instance = kwargs['instance']
job = instance.job
job.status = instance
set_job_started_at(instance=job, status=instance.status)
set_job_finished_at(instance=job, status=instance.status)
job.save(update_fields=['status', 'started_at', 'finished_at'])
# check if the new status is done to remove the containers from the monitors
if job.is_done:
from db.redis.containers import RedisJobContainers
RedisJobContainers.remove_job(job.uuid.hex)
# Check if we need to change the experiment status
auditor.record(event_type=EXPERIMENT_JOB_NEW_STATUS, instance=job)
def test_update_job_containers(self):
update_job_containers(event=status_experiment_job_event_with_conditions['object'],
status=JobLifeCycle.BUILDING,
job_container_name=settings.CONTAINER_NAME_EXPERIMENT_JOB)
# Assert it's still 0 because no job was created with that job_uuid
assert len(RedisJobContainers.get_containers()) == 0 # pylint:disable=len-as-condition
# Create a job with a specific uuid
labels = status_experiment_job_event_with_conditions['object']['metadata']['labels']
ExperimentJobFactory(uuid=labels['job_uuid'])
job = ExperimentJob.objects.get(uuid=labels['job_uuid'])
update_job_containers(event=status_experiment_job_event_with_conditions['object'],
status=JobLifeCycle.BUILDING,
job_container_name=settings.CONTAINER_NAME_EXPERIMENT_JOB)
# Assert now it has started monitoring the container
assert len(RedisJobContainers.get_containers()) == 1
container_id = '539e6a6f4209997094802b0657f90576fe129b7f81697120172836073d9bbd75'
assert RedisJobContainers.get_containers() == [container_id]
job_uuid, experiment_uuid = RedisJobContainers.get_job(container_id)
assert job.uuid.hex == job_uuid
assert job.experiment.uuid.hex == experiment_uuid
def experiment_job_status_post_save(sender, **kwargs):
instance = kwargs['instance']
job = instance.job
job.status = instance
set_job_started_at(instance=job, status=instance.status)
set_job_finished_at(instance=job, status=instance.status)
job.save(update_fields=['status', 'started_at', 'finished_at'])
# check if the new status is done to remove the containers from the monitors
if job.is_done:
from db.redis.containers import RedisJobContainers
RedisJobContainers.remove_job(job.uuid.hex)
# Check if we need to change the experiment status
experiment = instance.job.experiment
if experiment.is_done:
return
celery_app.send_task(SchedulerCeleryTasks.EXPERIMENTS_CHECK_STATUS,
kwargs={'experiment_id': experiment.id},
countdown=1)
def get_container_resources(node, container, gpu_resources):
# Check if the container is running
if container.status != ContainerStatuses.RUNNING:
logger.debug("`%s` container is not running", container.name)
RedisJobContainers.remove_container(container.id)
return
job_uuid, experiment_uuid = RedisJobContainers.get_job(container.id)
if not job_uuid:
logger.debug("`%s` container is not recognised", container.name)
return
logger.debug(
"Streaming resources for container %s in (job, experiment) (`%s`, `%s`) ",
container.id, job_uuid, experiment_uuid)
try:
stats = container.stats(decode=True, stream=False)
except json.decoder.JSONDecodeError:
logger.info("Error streaming states for `%s`", container.name)
except NotFound:
logger.debug("`%s` was not found", container.name)
RedisJobContainers.remove_container(container.id)
return
except requests.ReadTimeout:
return
precpu_stats = stats['precpu_stats']
cpu_stats = stats['cpu_stats']
pre_total_usage = float(precpu_stats['cpu_usage']['total_usage'])
total_usage = float(cpu_stats['cpu_usage']['total_usage'])
delta_total_usage = total_usage - pre_total_usage
pre_system_cpu_usage = float(precpu_stats['system_cpu_usage'])
system_cpu_usage = float(cpu_stats['system_cpu_usage'])
delta_system_cpu_usage = system_cpu_usage - pre_system_cpu_usage
percpu_usage = cpu_stats['cpu_usage']['percpu_usage']
num_cpu_cores = len(percpu_usage)
if num_cpu_cores >= node.cpu * 1.5:
logger.warning('Docker reporting num cpus `%s` and kubernetes reporting `%s`',
num_cpu_cores, node.cpu)
num_cpu_cores = node.cpu
cpu_percentage = 0.
percpu_percentage = [0.] * num_cpu_cores
if delta_total_usage > 0 and delta_system_cpu_usage > 0:
cpu_percentage = (delta_total_usage / delta_system_cpu_usage) * num_cpu_cores * 100.0
percpu_percentage = [cpu_usage / total_usage * cpu_percentage for cpu_usage in percpu_usage]
memory_used = int(stats['memory_stats']['usage'])
memory_limit = int(stats['memory_stats']['limit'])
container_gpu_resources = None
if gpu_resources:
gpu_indices = get_container_gpu_indices(container)
container_gpu_resources = [gpu_resources[gpu_indice] for gpu_indice in gpu_indices]
return ContainerResourcesConfig.from_dict({
'job_uuid': job_uuid,
'job_name': job_uuid, # it will be updated during the streaming
'experiment_uuid': experiment_uuid,
'container_id': container.id,
'cpu_percentage': cpu_percentage,
'n_cpus': num_cpu_cores,
'percpu_percentage': percpu_percentage,
'memory_used': memory_used,
'memory_limit': memory_limit,
'gpu_resources': container_gpu_resources
})
def test_update_job_containers_with_no_container_statuses(self):
update_job_containers(
event=status_experiment_job_event['object'],
status=JobLifeCycle.BUILDING,
job_container_name=conf.get(CONTAINER_NAME_EXPERIMENT_JOBS))
assert len(RedisJobContainers().get_containers()) == 0 # pylint:disable=len-as-condition
