How to deploy apache airflow (formally known as airbnb's airflow) scheduler in high availability?
I am not asking about the backend DB or RabbitMQ that should obviously be deployed in high availability configuration.
My main focus is the scheduler - is there something special needs to be done?
After a bit digging I found that it is not safe to run multiple schedulers simoultanously, this means that out of the box - airflow schedulers are not safe to use in high availablity environments.
The airflow team are planning to solve this issue by adding a lock mechanism on the DAG data structure, but this is not implemented yet (I checked by running 2 schedulers and saw that they schedule the same dag instances which is not good).
This is described here:
https://groups.google.com/forum/#!topic/airbnb_airflow/-1wKa3OcwME
I did found a way to workaround this high availalbilty issue by wrapping the schedulers with my own code and use cluster tools for leader election (I personanlly use consul for this purpose). This way only the elected master is running the scheduler and when the master is down the slave replaces him.
Please consider this when u use airflow in high availabilty environments since out of the box, airflow scheduler is currently not suitable for this (unless you solve this issue yourself).
Edit - an alternative approach to the master slave solution is to use a cluster manager/scheduler to make sure that only one airflow scheduler instance is always available. This approach relies on the self healing abilities of the cluster manager u have. For example both mesos and nomad supports this kind of configuration (I presonally chose nomad for its simplicity).
My personal experience was to follow the instructions I found for some best practices; that is to restart the scheduler every 10 runs ( -N 10 ) and use this software when possible:
https://github.com/teamclairvoyant/airflow-scheduler-failover-controller
I also use a DAG which pings a monitoring system to be sure that the scheduler has not gone away.
In my scenario, I have 2 schedulers (on 2 separate docker swarms), with the standby cluster scheduler turned off (using docker swarm service scale=0). I needed to make sure the primary scheduler had stopped fully before I started up the standby scheduler. What I found was that having 2 running schedulers (even for a brief time period) resulted in an occasional DAG scheduled to run on both clusters leading to duplicate reports generated from two different cluster zone.
Related
I have two servers in a HA mode. I'd like to know if is it possible to deploy an application on the slave server? If yes, how to configure it in jgroups? I need to run a specific program that access the master database, but I would not like to run on master serve to avoid overhead on it.
JGroups itself does not know much about WildFly and the deployments, it only creates a communication channel between nodes. I don't know where you get the notion of master/slave, but JGroups always has single* node marked as coordinator. You can check the membership through Channel.getView().
However, you still need to deploy the app on both nodes and just make it inactive if this is not its target node.
*) If there's no split-brain partition, or similar rare/temporal issues
I am using Mesos for container orchestration and get task history from Mesos using /task endpoint.
Mesos is running in a 7 nodes cluster and zookeeper is running in a 3 node cluster. I hope, Mesos uses Zookeeper to store the task History. We lost history sometimes when we restart Mesos. Does it store in memory? I am trying to understand what is happening here.
My questions are,
Where does it store task histories?
How can we configure the task history cleanup policy?
Why do we lose complete task history on restarting Mesos?
To answer your questions:
Task history/state for Mesos is stored in memory, and in the replicated_log (details here). The default is set to use the replicated_log, to store state completely in memory without the replicated_log you would have to specify this in your Mesos flags seen here in the configuration page as --registry=in_memory
Most users typically configure task history cleanup by using these three flags (there are more, but these are most common) --max_completed_frameworks=VALUE, --max_completed_tasks_per_framework=VALUE, and --max_unreachable_tasks_per_framework=VALUE as described in the previous document.
Yes, task history for the /tasks endpoint is lost every time a Mesos Master is restarted. However, the /state endpoint will still contain all task status changes over time.
**Edited to reflect information about the /tasks endpoint, not the /state endpoint.
Using dc/os we like to schedule tasks close to the data that the task requires that in our case is stored in hadoop/hdfs (on an HDP cluster). Issue is that the hadoop cluster is not run from within dc/os and so we are looking for a way to offer only a subset of the system resources.
For example: say we like to reserve 8GB of memory to data node services, then we like to provide the remainder to dc/os to schedule tasks.
From what i have read so far, the task can specify the resources it requires, but i have not found any means to specify what you want to offer from the node perspective.
I'm aware that a CDH cluster can be run on dc/os, that would be one way to go, but for now that is not provided for HDP.
Thanks for any idea's/tips,
Paul
Currently I am running my spark cluster as standalone mode. I am reading data from flat files or Cassandra(depending upon the job) and writing back the processed data to the Cassandra itself.
I was wondering if I switch to Hadoop and start using a Resource manager like YARN or mesos, does it give me an additional performance advantage like execution time and better resource management?
Currently sometime when I am processing huge chunk of data during shuffling with a possibility of stage failure. If I migrate to a YARN, can Resource manager address this issue?
Spark standalone cluster manager can also give you cluster mode capabilities.
Spark standalone cluster will provide almost all the same features as the other cluster managers if you are only running Spark.
When you submit your application in cluster mode all you job related files would be copied on to one of the machines on the cluster which would then submit the job on your behalf, if you submit the application in client mode the machine from which the job is being submitted would be taking care of driver related activities. This means that the machine from which the job has been submitted cannot go offline, whereas in cluster mode the machine from which the job has been submitted can go offline.
Having a Cassandra cluster would also not change any of these behaviors except it can save you network traffic if you can get the nearest contact point for the spark executor(Just like Data locality).
The failed stages gets rescheduled if you use either of the cluster managers.
I was wondering if I switch to Hadoop and start using a Resource manager like YARN or mesos, does it give me an additional performance advantage like execution time and better resource management?
In Standalone cluster model, each application uses all the available nodes in the cluster.
From spark-standalone documentation page:
The standalone cluster mode currently only supports a simple FIFO scheduler across applications. However, to allow multiple concurrent users, you can control the maximum number of resources each application will use. By default, it will acquire all cores in the cluster, which only makes sense if you just run one application at a time.
In other cases (when you are running multiple applications in the cluster) , you can prefer YARN.
Currently sometime when I am processing huge chunk of data during shuffling with a possibility of stage failure. If I migrate to a YARN, can Resource manager address this issue?
Not sure since your application logic is not known. But you can give a try with YARN.
Have a look at related SE question for benefits of YARN over Standalone and Mesos:
Which cluster type should I choose for Spark?
I have cluster of 3 Mesos slaves, where I have two applications: “redis” and “memcached”. Where redis depends on memcached and the requirement is both of the applications/services should start on same node instead of different slave nodes.
So I have created the application group and added the dependency properly in the JSON file. After launching the JSON file via “v2/groups” REST API, I observe that sometime both application group will start on same node but sometimes it will start on different slaves which breaks our requirement.
So intent/requirement is; if any application fails to start on a slave both the application should failover to other slave node. Also can I configure the JSON file to tell Marathon to start the application group on slave-1 (specific slave first) if it is available else start it on other slave in a cluster. Due to some reason if this application group will start on other slave can Marathon relaunch the application group to slave-1 if it is available to serve the request.
Thanks in advance for help.
Edit/Update (2):
Mesos, Marathon, and DC/OS support for PODs is available now:
DC/OS: https://dcos.io/docs/1.9/usage/pods/using-pods/
Mesos: https://github.com/apache/mesos/blob/master/docs/nested-container-and-task-group.md
Marathon: https://github.com/mesosphere/marathon/blob/master/docs/docs/pods.md
I assume you are talking about marathon apps.
Marathon application groups don't have any semantics concerning co-location on the same node and the same is the case for dependencies.
You seem to be looking for a Kubernetes like Pod abstraction in marathon, which is on the roadmap but not yet available (see update above :-)).
Hope this helps!
I think this should be possible (as a workaround) if you specify the correct app contraints within the group's JSON.
Have a look at the example request at
https://mesosphere.github.io/marathon/docs/generated/api.html#v2_groups_post
and the constraints syntax at
https://mesosphere.github.io/marathon/docs/constraints.html
e.g.
"constraints": [["hostname", "CLUSTER", "slave-1"]]
should do. Downside is that there will be no automatic failover to another slave that way. Still, I'd be curious why both apps need to specifically run on the same slave node...