Introduction

Hadoop slave node will regularly emit some metrics information to reflect the service healthy, service team will look the metrics to understand if the service is in healthy state, and trace back to understand the history behavior. Some typical use cases are:

• Pre-caution for un-healthy HBase RegionServer (the heap usage), RPC handling metrics and region aliveness, etc.

• Troubleshooting through the metrics history dashboard

• NameNode RPC traffic from client is very high, identify the source of the client, grep the user from audit log as well

• User can flexibly set threshold for each monitored metric and get alert notification w/o re-writing or create policies from scratch

• Notification about the HDFS clients which generating abnormal RPC traffic

• Extract the DN/RS list with abnormal RPC processing time

High-Level Monitoring/Alert Flow

Metrics Collector

Operation team is always struggling with the metrics monitoring for HBase cluster, e.g. HBase RegionServer heap usage, the RPC handling metrics for RegionServer, and region aliveness in regionserver. So we need solutions to get all those metrics. One option is to deploy a standalone JMX client in each node; another is to add JMX sink in Hadoop’s metrics system.

• A JMX client is expected to be deployed at each RegionServer slave node, or deployed at a collection of nodes to retrieve the JMX information. Since we have thousands of slave nodes, it’s not suppose to deploy those clients in a single server, which may lead to heavy load on that machine.

• JMX sink should be developed according to Hadoop’s metrics interface and plug into the Hadoop runtime environment

We tend to write the data into Kafka as a “distributed caching layer” to decouple the JMX client and the back-end storage system, and to avoid the storing latency of JMX data as well.

Agent

If we tend to build the JMX client to collect JMX metrics, we’d better have an agent to monitor whether the JMX client is working well, otherwise there may be some JMX data loss if we have some JMX clients stop working.

If we use JMX sink to collect data, no agent is required for JMX client, the data collection lifecycle is the same as the daemon lifecycle.

Metrics Storage

A scalable backend for large scale metrics information storage, as well as the query engine for time-series data, along with the min/max/average aggregation semantics.

NameNode Metrics

DataNode Metrics

HBase Master Metrics

RegionServer Metrics