Skip to content
Fetching contributors…
Cannot retrieve contributors at this time
138 lines (87 sloc) 5.18 KB

Hoya's needs of an application

Hoya installs and runs applications in a YARN cluster -applications that do not need to be written for YARN.

What they do need to be is deployable by Hoya, which means installable by YARN, configurable by Hoya, and, finally, executable by YARN. YARN will kill the executed process when destroying a container, so the deployed application must expect this to happen and be able to start up from a kill-initiated shutdown without any manual recovery process.

They need to locate each other dynamically, both at startup and during execution, because the location of processes will be unknown at startup, and may change due to server and process failures.


  • Install and run from a tarball -and be run from a user that is not root.

  • Be self contained or have all dependencies pre-installed.

  • Support dynamic discovery of nodes -such as via ZK.

  • Nodes to rebind themselves dynamically -so if nodes are moved, the application can continue

  • Handle kill as a normal shutdown mechanism.

  • Support multiple instances of the application running in the same cluster, with processes from different application instances sharing the same servers.

  • Operate correctly when more than one role instance in the application is deployed on the same physical host. (If YARN adds anti-affinity options in container requests this will no longer be a requirement)

  • Dynamically allocate any RPC or web ports -such as supporting 0 as the number of the port to listen on in configuration options.

  • Be trusted. YARN does not run code in a sandbox.

  • If it talks to HDFS or other parts of Hadoop, be built against/ship with libaries compatible with the version of Hadoop running on the cluster.

  • Store persistent data in HDFS (directly or indirectly) with the exact storage location configurable. Specifically: not to the local filesystem, and not in a hard coded location such as hdfs://app/data. Hoya creates per-Hoya application directories for persistent data.

  • Be configurable as to where any configuration directory is (or simply relative to the tarball). The application must not require it to be in a hard-coded location such as /etc.

  • Not have a fixed location for log output -such as /var/log/something

  • Run until explicitly terminated. Hoya treats an application termination (which triggers a container release) as a failure -and reacts to it by restarting the container.


  • Require human intervention at startup or termination.


These are the features that we'd like from a service:

  • Publish the actual RPC and HTTP ports in a way that can be picked up, such as via ZK or an admin API.

  • Be configurable via the standard Hadoop mechanisms: text files and XML configuration files. If not, custom parsers/configuration generators will be required.

  • Support an explicit parameter to define the configuration directory.

  • Take late bindings params via -D args or similar

  • Be possible to exec without running a complex script, so that process inheritance works everywhere, including (for testing) OS/X

  • Provide a way for Hoya to get list of nodes in cluster and status. This will let Hoya detect failed worker nodes and react to it.

  • FUTURE: If a graceful decommissioning is preferred, have an RPC method that a Hoya provider can call to invoke this.

  • Be location aware from startup. Example: worker nodes to be allocated tables to serve based on which tables are stored locally/in-rack, rather than just randomly. This will accelerate startup time.

  • Support simple liveness probes (such as an HTTP GET operations).

  • Return a well documented set of exit codes, so that failures can be propagated and understood.

  • Support cluster size flexing: the dynamic addition and removal of nodes.

  • Support a management platform such as Apache Ambari -so that the operational state of a Hoya application can be monitored.


  • Include a single process that will run at a fixed location and whose termination can trigger application termination. Such a process will be executed in the same container as the Hoya AM, and so known before all other containers are requested. If a live cluster is unable to handle restart/migration of such a process, then the Hoya application will be unable to handle Hoya AM restarts.

  • Ideally: report on load/cost of decommissioning. E.g amount of data; app load.


  • Be written for YARN.

  • Be (pure) Java. If the tarball contains native binaries for the cluster's hardware & OS, they should be executable.

  • Be dynamically reconfigurable, except for the special requirement of handling movement of manager/peer containers in an application-specific manner.

Something went wrong with that request. Please try again.