We're using AWS EMR for our spark jobs. All our jobs are submitted in yarn cluster mode, so the driver will run in one of the cluster nodes. We use on-demand node for master, and spot-instances for the core nodes. Now, although we almost always choose instances with < 5% interruption rate, sometimes it so happens that a significant fraction of our cluster nodes get terminated prematurely (probably because of higher demands).
So, I was wondering, in the above situation, what happens if a node containing the driver process goes down? Is there any chance of recovery for the spark job in that case? Or is the job gone forever?
The Spark driver is a single point of failure because it holds all cluster state for the running App.
In practice non-ephemeral storage can be used for check-pointing batch Apps after expensive expensive transformations. That said, trying to re-start after such a situation can be done, but when I looked into it, it is quite difficult to say the least. I asked such a question under my name some time ago, you can find it. I am quite technical but felt: gosh what a lot of hard work.
So, the recovery means rolling your own stuff, or accepting a re-run. Since I last evaluated EMR I see that the driver can run on the Master and that can be failed-over, but that is not the same thing as far as I can see, nor what you wish.
EMR has node leveling for CORE nodes in Yarn. Your spark driver/ Application master only gets created in CORE nodes. And HDFS also resides in CORE nodes only.
So to handle your situation in a best way, you may consider to use both CORE and TASK group.
What you can do to tackle this -
MASTER: On-demand
CORE: On-demand. Minimum no of Instances can be 1.
TASK: Spot with autoscaling with minimal EBS volume. Minimum no of Instances can be 0 this case.
This will reduce your cost also ensure that node containing the driver process never goes down.
https://docs.aws.amazon.com/emr/latest/ManagementGuide/emr-master-core-task-nodes.html
Related
My technical task is to synchronize data from GCS (Google Cloud Storage) to our Elasticsearch cluster.
We use Apache Spark 2.4 with the Elastic Hadoop connector on a Google Dataproc cluster (autoscaling enabled).
During the execution, if the Dataproc cluster downscaled, all tasks on the decommissioned node are lost and the processed data on this node are never pushed to elastic.
This problem does not exist when I save to GCS or HDFS for example.
How to make resilient this task even when nodes are decommissioned?
An extract of the stacktrace:
Lost task 50.0 in stage 2.3 (TID 427, xxxxxxx-sw-vrb7.c.xxxxxxx, executor 43): FetchFailed(BlockManagerId(30, xxxxxxx-w-23.c.xxxxxxx, 7337, None), shuffleId=0, mapId=26, reduceId=170, message=org.apache.spark.shuffle.FetchFailedException: Failed to connect to xxxxxxx-w-23.c.xxxxxxx:7337
Caused by: java.net.UnknownHostException: xxxxxxx-w-23.c.xxxxxxx
Task 50.0 in stage 2.3 (TID 427) failed, but the task will not be re-executed (either because the task failed with a shuffle data fetch failure, so the previous stage needs to be re-run, or because a different copy of the task has already succeeded).
Thanks.
Fred
I'll go through a bit of background on the "downscaling problem" and how to mitigate it. Note that this information applies to both manual downscaling as well as preemptible VMs getting preempted.
Background
Autoscaling removes nodes based on the amount of "available" YARN memory in the cluster. It does not take into account shuffle data on the cluster. Here's an illustration from a recent presentation we gave.
In a MapReduce-style job (Spark jobs are a set of MapReduce-style shuffles between stages), data from all mappers must get to all reducers. Mappers write their shuffle data to local disk, and then reducers fetch data from each mapper. There is a server on every node dedicated to serving shuffle data, and it runs outside of YARN. Therefore, a node can appear idle in YARN even though it needs to stay around to serve its shuffle data.
When a single node gets removed, pretty much all reducers will fail, since they all need to fetch data from every node. Reducers will specifically fail with FetchFailedException (as you saw), indicating they were unable to get shuffle data from a particular node. The driver will eventually re-run necessary mappers, and then re-run the reduce stage. Spark is a bit inefficient (https://issues.apache.org/jira/browse/SPARK-20178), but it works.
Note that you can lose nodes in one of three scenarios:
Intentionally removing nodes (autoscaling or manual downscaling)
Preemptible VMs
getting preempted. Preemptible VMs get preempted at least every 24 hours.
(Relatively rare) a standard GCE VM is
ungracefully terminated by GCE, and restarted. Usually, standard VMs
are transparently live migrated.
When you create an autoscaling cluster, Dataproc adds several properties to improve
job resiliency in the face of losing nodes:
yarn:yarn.resourcemanager.am.max-attempts=10
mapred:mapreduce.map.maxattempts=10
mapred:mapreduce.reduce.maxattempts=10
spark:spark.task.maxFailures=10
spark:spark.stage.maxConsecutiveAttempts=10
spark:spark.yarn.am.attemptFailuresValidityInterval=1h
spark:spark.yarn.executor.failuresValidityInterval=1h
Note that if you enable autoscaling on an existing cluster, it will not have these properties set. (But you can set them manually when creating a cluster).
Mitigations
1) Use graceful decommissioning
Dataproc integrates with YARN's Graceful Decommissioning, and can be set on Autoscaling Policies or manual downscale operations.
When gracefully decommissioning a node, YARN keeps it around until applications that ran containers on the node finish, but does not let it run new containers. That gives nodes an opportunity to serve their shuffle data before being removed.
You will need to ensure that your graceful decommission timeout is long enough to encompass your longest jobs. The autoscaling docs suggest 1h as a starting point.
Note that graceful decommissioning only really makes sense only long-running clusters that process lots of short jobs.
On ephemeral clusters, you would be better off "right-sizing" the cluster from the start, or disabling downscaling unless the cluster is completely idle (set scaleDownMinWorkerFraction=1.0).
2) Avoid preemptible VMs
Even when using graceful decommissioning, preemptible VMs will be periodically terminated through "preemptions". GCE guarantees preemptible VMs will get preempted within 24 hours, and preemptions on large clusters are very spread out.
If you are using graceful decommissioning, and the FetchFailedException error messages include -sw-, you are likely seeing fetch failures due to nodes being preempted.
You have two options to avoid using preemptible VMs:
1. In your autoscaling policy, you can set secondaryWorkerConfig to have 0 min and max instances, and instead put all workers in the primary group.
2. Alternatively, you can keep using "secondary" workers, but set --properties dataproc:secondary-workers.is-preemptible.override=false. That will make your secondary workers be standard VMs.
3) Long term: Enhanced Flexibility Mode
Dataproc's Enhanced Flexibility Mode is the long term answer to the shuffle problem.
The downscaling problem is caused by shuffle data getting stored on local disk. EFM will include new shuffle implementations that allow placing shuffle data on a fixed set of nodes (e.g. just primary workers), or on storage outside of the cluster.
That will make secondary workers stateless, which means they can be removed at any time. This makes autoscaling far more compelling.
At the moment, EFM is still in Alpha, and does not scale to real-world workloads, but look out for a production-ready Beta by the summer.
I'm trying to fetch some data from Cloudera's Quick Start Hadoop distribution (a Linux VM for us) on our SAP HANA database using SAP Spark Controller. Every time I trigger the job in HANA, it gets stuck and I see the following warning being logged continuously every 10-15 seconds in SPARK Controller's log file, unless I kill the job.
WARN org.apache.spark.scheduler.cluster.YarnScheduler: Initial job has not accepted any resources; check your cluster UI to ensure that workers are registered and have sufficient resources
Although it's logged like a warning it looks like it's a problem that prevents the job from executing on Cloudera. From what I read, it's either an issue with the resource management on Cloudera, or an issue with blocked ports. In our case we don't have any blocked ports so it must be the former.
Our Cloudera is running a single node and has 16GB RAM with 4 CPU cores.
Looking at the overall configuration I have a bunch of warnings, but I can't determine if they are relevant to the issue or not.
Here's also how the RAM is distributed on Cloudera
It would be great if you can help me pinpoint the cause for this issue because I've been trying various combinations of things over the past few days without any success.
Thanks,
Dimitar
You're trying to use the Cloudera Quickstart VM for a purpose beyond it's capacity. It's really meant for someone to play around with Hadoop and CDH and should not be used for any production level work.
Your Node Manager only has 5GB of memory to use for compute resources. In order to do any work, you need to create an Application Master(AM) and a Spark Executor and then have reserve memory for your executors which you won't have on a Quickstart VM.
I have a daily pipeline running on Spark Standalone 2.1. Its deployed in and runs on AWS EC2 and uses S3 for its persistence layer. For the most part, the pipeline runs without a hitch, but occasionally the job hangs on a single worker node during a reduceByKey operation. When I work into the worker, I notice that the CPU (as seen via top) is pegged at 100%. My remedy so far is to reboot the worker node so that Spark re-assigns the task and the job proceeds fine from there.
I would like to be able to mitigate this issue. I gather that I can prevent CPU pegging by switching to use YARN as my cluster manager, but I wonder whether I could configure Spark Standalone to prevent CPU pegging by maybe limiting the number of cores that get assigned to the Spark job ? Any suggestions would be greatly appreciated.
I'm running a Spark Streaming process on a 16 CPU's 64 GB RAM host with Mesos.
When I'm running it using Mesos as a cluster manager (by setting --master mesos://leader.mesos:5050) it's running much slower than when it is run in local mode (--master local[4]).
I can't find the reason for that and I have no clue. One of the things I've noticed is that there is one specific task that is taking significantly more time on Mesos than in Local.
The weird thing (maybe that should be the questions' title) is that the task itself takes 6s and its stage (it has only one stage) takes less than a second. See attached pictures (Mesos (1) and (2)). How come? Isn't a job equal to the sum of its parts?
Local:
Mesos:
(1)
(2)
Another note: I did manage to run this exact same Spark Streaming process on another Mesos cluster, and it runs in a sensible amount of time, pretty much like in the local mode described above. The only difference that I can think of is that this cluster has more than one host, and that Spark is running with 2 executors rather than 1. (I couldn't find a way to run more than 1 executor on the same host on Mesos). Is this may be the reason?
Any clues would be much appreciated.
Spark can run over Mesos in two modes: coarse-grained (default) and fine-grained (see documentation).
In coarse-grained mode Spark launches exactly one executor on each machine it was assigned to by Mesos. Inside this task Spark launches other mini-tasks. It has the benefit of lower startup overhead (in your case you don't want to change this mode).
Could you be more specific about your streaming job? Is it CPU, disk, or network bounded? You can easily compare performance if you run some of Spark examples.
If your task is CPU intensive you might consider setting spark.mesos.extra.cores. By default Spark tries to acquire all cores that are being offered by Mesos. So, if there's no other task running on that cluster it shouldn't be a problem.
I have 2 spark streaming jobs that I want to run, as well as keeping some available resources for batch jobs and other operations.
I evaluated Spark Standalone cluster manager, but I realized that I would have to fix the resources for two jobs, which would leave almost no computing power to batch jobs.
I started evaluating Mesos, because it has "fine grained" execution model, where resources are shifted between Spark applications.
1) Does it mean that a single core can be shifted between 2 streaming applications?
2) Although I have spark & cassandra, in order to exploit data locality, do I need to have dedicated core on each of the slave machines to avoid shuffling?
3) Would you recommend running Streaming jobs in "fine grained" or "course grained" mode. I know that logical answer is course grained (in order to minimize the latency of streaming apps) but what when resource in total cluster are limited (cluster of 3 nodes, 4 cores each - there are 2 streaming applications to run and multiple time to time batch jobs)
4) In Mesos, when I run spark streaming job in cluster mode, will it occupy 1 core permanently (like Standalone cluster manager is doing), or will that core execute driver process and sometimes act as executor?
Thank you
Fine grained mode is actually deprecated now. Even with it, each core is allocated to task until completion, but in Spark Streaming, each processing interval is a new job, so tasks only last as long the time it takes to process each interval's data. Hopefully that time is less than the interval time or your processing will back up, eventually running out of memory to store all those RDDs waiting for processing.
Note also that you'll need to have one core dedicated to each stream's Reader. Each will be pinned for the life of the stream! You'll need extra cores in case the stream ingestion needs to be restarted; Spark will try to use a different core. Plus you'll have a core tied up by your driver, if it's also running on the cluster (as opposed to on your laptop or something).
Still, Mesos is a good choice, because it will allocate the tasks to nodes that have capacity to run them. Your cluster sounds pretty small for what you're trying to do, unless the data streams are small themselves.
If you use the Datastax connector for Spark, it will try to keep input partitions local to the Spark tasks. However, I believe that connector assumes it will manage Spark itself, using Standalone mode. So, before you adopt Mesos, check to see if that's really all you need.