Service Discovery

This document describes the mechanism of service discovery on Hadoop and the steps for enabling it.


A DNS server is implemented to enable discovering services on Hadoop via the standard mechanism: DNS lookup.

The framework ApplicationMaster posts the container information such as hostname and IP address into the Hadoop service registry. The DNS server exposes the information in Hadoop service registry by translating them into DNS records such as A record and SRV record. Clients can then discover the IPs of containers via standard DNS lookup.

For non-docker containers (containers with null Artifact or with Artifact type set to TARBALL), since all containers on the same host share the same ip address, the DNS supports forward DNS lookup, but not support reverse DNS lookup. With docker, it supports both forward and reverse lookup, since each container can be configured to have its own unique IP. In addition, the DNS also supports configuring static zone files for both foward and reverse lookup.

Docker Container IP Management in Cluster

To support the use-case of per container per IP, containers must be launched with bridge network. However, with bridge network, containers running on one node are not routable from other nodes by default. This is not an issue if you are only doing single node testing, however, for a multi-node environment, containers must be made routable from other nodes.

There are several approaches to solve this depending on the platforms like GCE or AWS. Please refer to specific platform documentations for how to enable this. For on-prem cluster, one way to solve this issue is, on each node, configure the docker daemon to use a custom bridge say br0 which is routable from all nodes. Also, assign an exclusive, contiguous range of IP addresses expressed in CIDR form e.g (64 IPs) to each docker daemon using the fixed-cidr option like below in the docker daemon.json:

"bridge": "br0"
"fixed-cidr": ""

Check how to customize docker bridge network for details.

Naming Convention with Registry DNS

With the DNS support, user can simply access their services in a well-defined naming format as below:


For example, in a cluster whose domain name is yarncluster (as defined by the hadoop.registry.dns.domain-name in core-site.xml), a service named hbase deployed by user devuser with two components hbasemaster and regionserver can be accessed as below:

This URL points to the usual hbase master UI


Note that YARN service framework assigns COMPONENT_INSTANCE_NAME for each container in a sequence of monotonically increasing integers. For example, hbasemaster-0 gets assigned 0 since it is the first and only instance for the hbasemaster component. In case of regionserver component, it can have multiple containers and so be named as such: regionserver-0, regionserver-1, regionserver-2 … etc

Each YARN service component also has Multi-A Records for container fault tolerance or load balancing via RegistryDNS. The naming format is defined as:


For example, a component named www for application app launched by Chuck with 3 containers will have DNS records that look like: IN A IN A IN A

Disclaimer: The DNS implementation is still experimental. It should not be used as a fully-functional DNS.