Deploying TF 1.0 on GC ML with GPUs queues forever - google-cloud-ml

I'm using the following parameters:
scale_tier = CUSTOM
master_type = standard_gpu
worker_count = 0
ps_count = 0
runtime_version = 1.0
The logs show the message Job <job_name> is queued. without ever printing the next log message I usually see (Waiting for job to be provisioned.). My job has been stuck like this for 30 minutes, and I've tried twice already.
I tried submitting the same job without GPUs (i.e., scale_tier = BASIC), and it moved onto the provisioning stage and eventually finished.
What might be going on?

It turns out this error was because we ran out of quota for GPUs. When I killed some jobs that were using GPU, the queued job moved forward into the provisioning phase.
Thanks!

Related

MWAA Airflow Scaling: what do I do when I have to run frequent & time consuming scripts? (Negsignal.SIGKILL)

I have an MWAA Airflow env in my AWS account. The DAG I am setting up is supposed to read massive data from S3 bucket A, filter what I want and dump the filtered results to S3 bucket B. It needs to read every minute since the data is coming in every minute. Every run processes about 200MB of json data.
My initial setting was using env class mw1.small with 10 worker machines, if I only run the task once in this setting, it takes about 8 minutes to finish each run, but when I start the schedule to run every minute, most of them could not finish, starts to take much longer to run (around 18 mins) and displays the error message:
[2021-09-25 20:33:16,472] {{local_task_job.py:102}} INFO - Task exited with return code Negsignal.SIGKILL
I tried to expand env class to mw1.large with 15 workers, more jobs were able to complete before the error shows up, but still could not catch up with the speed of ingesting every minute. The Negsignal.SIGKILL error would still show before even reaching worker machine max.
At this point, what should I do to scale this? I can imagine opening another Airflow env but that does not really make sense. There must be a way to do it within one env.
I've found the solution to this, for MWAA, edit the environment and under Airflow configuration options, setup these configs
celery.sync_parallelism = 1
celery.worker_autoscale = 1,1
This will make sure your worker machine runs 1 job at a time, preventing multiple jobs to share the worker, hence saving memory and reduces runtime.

Why is Google Cloud Task queue so slow?

I have a google cloud tasks queue that processes 1000's of HTTP requested (cloud functions). I've set up the task queue with the default settings except updated "Max attempts" = 2
Each task is dispatched using python using "from google.cloud import tasks_v2" package.
the issue I'm facing is that it just takes too long to finish processing all the tasks within the queue, I would have expected with a setting of "max concurrent" = 1000 I would see more tasks running in one go? When refreshing all observing the "running tasks" indicator I've only seen it at a maximum of 15.
Have I missed something or are there other settings I can play with to get these tasks process quicker?
It turns out that the issues had to do with my cloud function. I have try-catch statements that would return a status of 500 when an error occurred.
It seems that cloud tasks will back off when it sees an increase in error responses. I ended up changing my catch statements to return a 200 and my task queue is finishing substantially quicker now.
Hope this helps someone else in the future.

Google AI Platform training - wait for the job to finish

I've built an AI Platform pipeline with a lot of parallel processes. Each process launches a training job on the AI Platform, like this:
gcloud ai-platform jobs submit training ...
Then it has to wait for the job to finish to pass to the next step. For doing this, I've tried to add the parameter --stream-logs to the above command. In this way, it streams all the logs until the job is done.
The problem is, with so many parallel processes, I run out of requests for getting logs:
Quota exceeded for quota metric 'Read requests' and limit 'Read requests per minute'
of service 'logging.googleapis.com'
But I do not need to actually stream the logs, I just need a way to tell the process to "wait" until the training job is done. Is there a smarter and simpler way of doing this?
I've just found that I can use the Python API to launch and monitor the job:
training_inputs = {
'scaleTier': 'CUSTOM',
'masterType': 'n1-standard-8',
...
}
job_spec = {'jobId': 'your_job_name', 'trainingInput': training_inputs}
project_name = 'your-project'
project_id = 'projects/{}'.format(project_name)
cloudml = discovery.build('ml', 'v1')
request = cloudml.projects().jobs().create(
body=job_spec,
parent=project_id
)
response = request.execute()
Now I can set up a loop that checks the job state every 60 seconds
state = 'RUNNING'
while state == 'RUNNING':
time.sleep(60)
status_req = cloudml.projects().jobs().get(name=f'{project_id}/jobs/{job_name}')
state = status_req.execute()['state']
print(state)
Regarding the error message you are experiencing, indeed you are hitting a quota exceeded for Cloud Logging, what you can do is to request a quota increase.
On the other hand, about an smarter way to check the status of a job without streaming logs, what you can do is to check the status once in a while by running gcloud ai-platform jobs describe <job_name> or create a Python script to check the status, this is explained in the following documentation.

Long-running Dataflow job fails with no errors in user code

After running for 17 hours, my Dataflow job failed with the following message:
The job failed because a work item has failed 4 times. Look in previous log entries for the cause of each one of the 4 failures.
The 4 failures consist of 3 workers losing contact with the service, and one worker reported dead:
****-q15f Root cause: The worker lost contact with the service.
****-pq33 Root cause: The worker lost contact with the service.
****-fzdp Root cause: The worker ****-fzdp has been reported dead. Aborting lease 4624388267005979538.
****-nd4r Root cause: The worker lost contact with the service.
I don't see any errors in the worker logs for the job in Stackdriver. Is this just bad luck? I don't know how frequently work items need to be retried, so I don't know what the probability is that a single work item will fail 4 times over the course of a 24 hour job. But this same type of job failure happens frequently for this long-running job, so it seems like we need some way to either decrease the failure rate of work items, or increase the allowed number of retries. Is either possible? This doesn't seem related to my pipeline code, but in case it's relevant, I'm using the Python SDK with apache-beam==2.15.0. I'd appreciate any advice on how to debug this.
Update: The "STACK TRACES" section in the console is totally empty.
I was having the same problem and it was solved by scaling up my workers resources. Specifically, I set --machine_type=n1-highcpu-96 in my pipeline configs. See this for a more extensive list on machine type options.
Edit: Set it to highcpu or highmem depending on the requirements of your pipeline process

Celery/SQS task retry gone haywire - how to get rid of it?

We've got Celery/SQS set up for asynchronous task management. We're running Django for our framework. We have a celery task that has a self.retry() in it. Max_retries is set to 15. The retry is happening with an exponential backoff and takes 182 hours to complete all 15 retries.
Last week, this task went haywire, I think due to a bug in our code not properly handling a service outage. It resulted in exponential creation (retrying?) of the same celery task. It eventually used up all available memory and the worker crashed. Restarting the worker results in another crash a couple hours later, since all those tasks (and their retries) keep retrying and spawning new retries until we run out of memory again. Ultimately we ended up with nearly 600k tasks created!
We need our workers to ignore all the tasks with a specific celery GUID. Ideally we could just get rid of them for good. I was going to use revoke() but, per documentation (http://docs.celeryproject.org/en/3.1/userguide/workers.html#commands), this is only implemented for Redis and RabbitMQ, not SQS. Furthermore, when I go to the SQS service in the AWS console, it's showing zero messages in flight so it's not like I can just flush it.
Is there a way to delete or revoke a specific message from SQS using the Celery task ID? Or is there another way to fix this problem? Obviously we need to fix our code so we don't get into this situation again, but first we need to get our worker up and running because without it our website has reduced functionality. Thanks!