YARN

When an application wants to run, the client launches the ApplicationMaster, which then negotiates with the ResourceManager to get resources in the cluster in the form of containers. A container represents CPUs (cores) and memory allocated on a single node to be used to run tasks and processes. Containers are supervised by the NodeManager and scheduled by the ResourceManager.

Examples of containers:

• One core and 4 GB RAM
• Two cores and 6 GB RAM
• Four cores and 20 GB RAM

Some containers are assigned to be mappers and others to be reducers; all this is coordinated by the ApplicationMaster in conjunction with the ResourceManager. This framework is called YARN:

Using YARN, several different applications can request for and execute tasks on containers, sharing the cluster resources pretty well. However, as the size of the clusters grows and the variety of applications and requirements change, the efficiency of the resource utilization is not as good over time.

Opportunistic containers

Opportunistic containers can be transmitted to a NodeManager even if their execution at that particular time cannot begin immediately, unlike YARN containers, which are scheduled in a node if and only if there are unallocated resources.

In these types of scenarios, opportunistic containers will be queued at the NodeManager till the required resources are available for use. The ultimate goal of these containers is to enhance the cluster resource utilization and in turn improve task throughput.

Types of container execution 

There are two types of container, as follows:

• Guaranteed containers: These containers correspond to the existing YARN containers. They are assigned by the capacity scheduler. They are transmitted to a node if and only if there are resources available to begin their execution immediately. 
• Opportunistic containers: Unlike guaranteed containers, in this case we cannot guarantee that there will be resources available to begin their execution once they are dispatched to a node. On the contrary, they will be queued at the NodeManager itself until resources become available.

YARN timeline service v.2

The YARN timeline service v.2 addresses the following two major challenges:

• Enhancing the scalability and reliability of the timeline service
• Improving usability by introducing flows and aggregation

Enhancing scalability and reliability

Version 2 adopts a more scalable distributed writer architecture and backend storage, as opposed to v.1 which does not scale well beyond small clusters as it used a single instance of writer/reader architecture and backend storage.

Since Apache HBase scales well even to larger clusters and continues to maintain a good read and write response time, v.2 prefers to select it as the primary backend storage.

Usability improvements

Many a time, users are more interested in the information obtained at the level of flows or in logical groups of YARN applications. For this reason, it is more convenient to launch a series of YARN applications to complete a logical workflow.

In order to achieve this, v.2 supports the notion of flows and aggregates metrics at the flow level.

Architecture

YARN Timeline Service v.2 uses a set of collectors (writers) to write data to the back-end
storage. The collectors are distributed and co-located with the application masters to which
they are dedicated. All data that belong to that application are sent to the application level
timeline collectors with the exception of the resource manager timeline collector.

For a given application, the application master can write data for the application to the co-located timeline collectors (which is an NM auxiliary service in this release). In addition,
node managers of other nodes that are running the containers for the application also write
data to the timeline collector on the node that is running the application master. 

The resource manager also maintains its own timeline collector. It emits only YARN-generic
life-cycle events to keep its volume of writes reasonable.

The timeline readers are separate daemons separate from the timeline collectors, and they
are dedicated to serving queries via REST API:

The following diagram illustrates the design at a high level: