How Apache Hadoop YARN HA Works
- by Karthik Kambatla, Wing Yew Poon, and Vikram Srivastava
- May 08, 2014
- 1 comment
Thanks to recent work upstream, YARN is now a highly available service. This post explains its architecture and configuration details.
YARN, the next-generation compute and resource management framework in Apache Hadoop, until recently had a single point of failure: the ResourceManager, which coordinates work in a YARN cluster. With planned (upgrades) or unplanned (node crashes) events, this central service, and YARN itself, could become unavailable.
This post details Cloudera’s recent work in the Hadoop community (YARN-149) to make the ResourceManager (and thus YARN) highly available. We’ll also explain the high-level design for how the ResourceManager’s active state is preserved (state store), and how ResourceManagers fail-over to achieve high availability (HA). We also outline how to deploy an HA cluster, and review some important configuration options.
CDH 5 (and Apache Hadoop 2.x) ship YARN as the vehicle to manage cluster resources, and share the said resources among compute frameworks like MapReduce, Impala, and Apache Spark. In previous posts, you have learned the high-level architecture of YARN, and how to migrate from MR1 to MR2/YARN for users and cluster admins.
To briefly re-cap, YARN has a master/worker architecture (see below); the master (the ResourceManager) manages the resources on the workers and schedules work in the cluster. Furthermore, the ResourceManager handles all client interactions.
YARN’s master/worker architecture
As you can see, the ResourceManager is an obvious single point of failure!
Design for HA
ResourceManager HA is realized through an Active/Standby architecture: one ResourceManager is Active, and one or more ResourceManagers are in Standby mode waiting to take over should anything happen to the Active:
YARN HA design
Below, you’ll learn how the store enables restarting the ResourceManager without losing any state, restarting an ResourceManager in a single-ResourceManager cluster, and failing over between ResourceManagers when there are multiple ResourceManagers in the cluster.
ResourceManager State Store
ResourceManagerStateStore enables storing the internal state of the ResourceManager (applications and their attempts, delegation tokens, and version information). The state of the cluster (resource consumption on individual nodes) is not stored because it can be easily reconstructed when the NodeManagers heartbeat to the “new” ResourceManager.
There are three possible implementations of the state store:
- Memory-based implementation (primarily for testing)
- Filesystem-based implementation (HDFS or local filesystems can be used)
- Apache ZooKeeper-based implementation
We recommend using the ZooKeeper-based implementation when there are multiple ResourceManagers to ensure one assumes the Active role.
With the state store enabled, an ResourceManager, on restart, loads the internal application state from the store. The scheduler reconstructs its state when each node heartbeats. A new attempt is spawned for each managed application previously submitted to the ResourceManager. Applications can checkpoint periodically to avoid losing any work; the MR ApplicationMaster checkpoints completed tasks, so the completed tasks need not be re-run on ResourceManager restart.
ResourceManager failover is an extension of ResourceManager restart. When the cluster is configured to have multiple ResourceManagers (two or more – one Active, others Standby), one of the Standby ResourceManagers takes over if/when the Active ResourceManager goes down. On startup, each ResourceManager starts in the Standby mode — the process is started, but the state is not loaded. When transitioning to Active, the ResourceManager loads the internal state from the designated state store and starts all the internal services to assume the Active ResourceManager duties. The stimulus to transition-to-active comes from either the admin (through CLI) or through the integrated failover controller when automatic failover is enabled.
In a multi-ResourceManager cluster, there is a risk of more than one ResourceManager assuming the Active role. This can lead to a split-brain situation, where each ResourceManager controls a subset of the resources in the cluster and handles client requests. To avoid this split-brain condition, a “fencing” mechanism is required by which an Active ResourceManager disallows other ResourceManagers to assume the Active role. The ZooKeeper-based state store (ZKResourceManagerStateStore) implicitly allows write access to a single ResourceManager at any point in time, and hence is the recommended state store to use in an HA cluster. When using the ZKResourceManagerStateStore, there is no need for a separate fencing mechanism to address a potential split-brain situation where multiple ResourceManagers assume the Active role.
The ZooKeeper state store achieves this implicit fencing through ACLs. All the ResourceManagers have shared read-write-admin access to the store, but only the Active has create-delete access. A ResourceManager claims this create-delete access while transitioning to Active. At this point, any other ResourceManager that previously had create-delete access loses access, fails to make changes to the store, and transitions itself to Standby. By having each ResourceManager create a dummy znode every so often (10 seconds, by default), a ResourceManager is always informed of its access to the store.
Automatic failover though embedded leader election (Recommended)
ResourceManagers use ZooKeeper-based
ActiveStandbyElector to decide which ResourceManager should be the Active. When the Active goes down or becomes unresponsive, another ResourceManager is automatically elected to be the Active and takes over. (Note that this elector is embedded in the ResourceManager, and hence there is no need to run a separate
ZooKeeperFailoverController (ZKFC) daemon as is the case for HDFS NameNode HA.)
Manual failover (Not recommended)
When automatic failover is not enabled, admins have to manually transition one of the ResourceManagers to Active. To failover from one ResourceManager to the other, they are expected to first transition the Active ResourceManager to Standby and transition a Standby ResourceManager to Active. All this can be done using the
yarn rmadmin CLI.
When there are multiple ResourceManagers, the YARN configuration (
yarn-site.xml) used by clients and nodes is expected to list all the ResourceManagers. Clients, ApplicationMasters, and NodeManagers try connecting to the ResourceManagers in a round-robin fashion until they hit the Active ResourceManager. If the Active goes down, they resume the round-robin polling until they hit the “new” Active.
Deploying an HA Cluster
One can set up an HA cluster with or without using Cloudera Manager. We recommend using it, because all you have to do is click a few buttons!
Deploying an HA Cluster using Cloudera Manager (Recommended)
Cloudera Manager 5 provides a simple wizard to enable high availability for YARN ResourceManager similar to other services like HDFS, MapReduce, and Oozie. The wizard takes care of adding the second ResourceManager and setting all configuration properties needed for HA setup. To get to the wizard, go to your YARN service in Cloudera Manager and under “Actions”, click on “Enable High Availability”.
Kicking-off the Cloudera Manager wizard for YARN HA
The next page in the wizard asks you to specify the host where the new ResourceManager should start.
Specifying the host
Finally, Cloudera Manager starts the workflow that enables HA, restarts all the dependent services of YARN so that they can pick up the new ResourceManager, and deploys the client configuration.
YARN HA workflow in progress
Deploying an HA Cluster Manually
- To enable HA, set
yarn.resourcemanager.ha.enabledto true in
- YARN uses identifiers (rm-ids) as logical names for each resourcemanager. Set
yarn.resourcemanager.ha.rm-idsto the list of identifiers you wish to use; for example,
rm1,rm2(the value is a comma-separated list).
- For each of the identifiers (corresponding to each ResourceManager), define
yarn.resourcemanager.hostname.(should be the hostname of the ResourceManager). Alternatively, you can define the addresses for that ResourceManager:
yarn.resourcemanager.scheduler.address, and so on. (See the CDH5 documentation for more details.) NodeManagers and clients use the configured addresses to find the ResourceManager service they need to talk to. They go through the list of rm-ids and successively try the address corresponding to each rm-id.
For a YARN application to complete successfully, both the ApplicationMaster and the ResourceManager must be live. Containers with completed tasks need to report to the ApplicationMaster, which in turn needs to report to the ResourceManager to commit the job completion. Recovery is controlled via configs:
- To enable recovering applications on ResourceManager restart/failover, set
- The maximum number of times any YARN application can be recovered is set by
yarn.resourcemanager.am.max-attempts; we recommend setting this global limit to a high value.
When a ResourceManager dies and is restarted, or fails over to another ResourceManager in the case of an HA cluster, the newly active ResourceManager instructs running ApplicationMasters to abort (YARN-556). This uses up an application attempt. Also, if the ResourceManager is down for some time and the ApplicationMaster is unable to connect, it will timeout and abort. That uses up an application attempt too. When a new ResourceManager becomes active, it can recover applications with failed attempts that have not exceeded their max-attempts.
For MapReduce running on YARN (aka MR2), the MR ApplicationMaster plays the role of a per-job jobtracker. MRAM failure recovery is controlled by the property,
mapreduce.am.max-attempts. This property may be set per job. If its value is greater than 1, then when the ApplicationMaster dies, a new one is spun up for a new application attempt, up to the max-attempts. When a new application attempt is started, in-flight tasks are aborted and rerun but completed tasks are not rerun.
State Store Configuration
As mentioned previously, we recommend using the ZooKeeper-based ResourceManager state store implementation:
yarn.resourcemanager.zk-addressto the address of the ZooKeeper ensemble (a comma-separated list of hosts, optionally with port if not using the default port of 2181).
- To configure automatic ResourceManager failover, set
- In addition, we recommend setting
yarn.resourcemanager.ha.automatic-failover.embeddedto true; this uses the
EmbeddedElectorServiceto pick an active ResourceManager.
- Also, you need to configure
yarn.resourcemanager.cluster-idfor the cluster; this name is used for electing the active ResourceManager.
The web UI of the standby ResourceManager will automatically redirect to that of the active ResourceManager, which is convenient. If you want to ascertain the HA status of a particular ResourceManager in an HA cluster, you can go to the
/cluster/cluster page of the ResourceManager web address (this page does not redirect), or you can use the REST API, by going to
/ws/v1/cluster/info of the web address.
Currently, all in-flight work is interrupted on ResourceManager restart/failover. While this approach doesn’t hurt applications with short-lived tasks like MapReduce, restarting long-running applications can lead to losing significant amount of internal task/application state and would be nice to avoid. Working-preserving ResourceManager restart aims to do exactly that: all ApplicationMaster and task containers continue running and “resync” with the “new” ResourceManager.
The community is also working on work-preserving restart for the NodeManager (worker; YARN-1336) and the ApplicationMaster (YARN-1489) to complement ResourceManager high availability. With all these pieces in place, one would be able to do a rolling restart on a YARN cluster without losing any work at all!
Several people in the Hadoop community have contributed to improving the availability of YARN. In particular, Alejandro Abdelnur, Arpit Gupta, Bikas Saha, Hitesh Shah, Jian He, Sandy Ryza, Tom White, Tsuyoshi Ozawa, Vinod Kumar Vavilapalli, Xuan Gong, and Zhijie Shen all played a major role in design, implementation, testing, and code reviews. Insights from Aaron T Myers, Patrick Hunt, Phil Zeyliger, Santhosh Srinivasan, and Todd Lipcon also helped us arrive at the final design.
Karthik Kambatla, Wing Yew Poon, and Vikram Srivastava are Software Engineers at Cloudera. Karthik is also a Hadoop committer.