Best Hadoop Monitoring Tools

In an era where enterprises are looking to leverage the power of big data, Apache Hadoop has become a key tool for storing and processing large datasets efficiently. However, like all pieces of infrastructure, it needs to be monitored.

Here is our list of the best Hadoop monitoring tools:

  1. Datadog – Cloud monitoring software with a customizable Hadoop dashboard, integrations, alerts, and more.
  2. LogicMonitor – Infrastructure monitoring software with a HadoopPackage, REST API, alerts, reports, dashboards, and more.
  3. Dynatrace – Application performance management software with Hadoop monitoring with NameNode/DataNode metrics, dashboards, analytics,, custom alerts, and more.

What is Apache Hadoop?

Apache Hadoop is an open-source software framework that can process and distribute large data sets across multiple clusters of computers. Hadoop was designed to break down data management workloads over a cluster of computers. It divides data processing between multiple nodes, which manages the datasets more efficiently than a single device could.

How to Monitor Hadoop: Metrics You Need to Keep Track of to  Monitor Hadoop Clusters

Like any computing resource, Hadoop clusters need to be monitored to ensure that they keep performing at their best. Hadoop’s architecture may be resilient to system failures, but it still needs maintenance to prevent jobs from being disrupted. When monitoring the status of clusters, there are four main categories of metrics you need to be aware of:

  • HDFS metrics (NameNode metrics and DataNode metrics)
  • MapReduce counters
  • YARN metrics
  • ZooKeeper metrics

Below, we’re going to break each of these metric types down, explaining what they are and providing a brief guide for how you can monitor them.

HDFS Metrics

Apache Hadoop Distributed File System (HDFS) is a distributed file system with a NameNode and DataNode architecture. Whenever the HDFS receives data it breaks it down into blocks and sends it to multiple nodes. The HDFS is scalable and can support thousands of nodes.

Monitoring key HDFS metrics is important because it helps you to: monitor the capacity of the DFS, monitor the space available, track the status of blocks, and optimize the storage of your data.

There are two main categories of HDFS metrics:

  • NameNode metrics
  • DataNode metrics

NameNodes and DataNodes

HDFS follows a master-slave architecture where every cluster in the HDFS is composed of a single NameNode (master) and multiple DataNodes (slave). The NameNode controls access to files, records metadata of files stored in the cluster, and monitors the state of DataNodes.

A DataNode is a process that runs on each slave machine, which performs low-level read/write requests from the system’s clients, and sends periodic heartbeats to the NameNode, to report on the health of the HDFS. The NameNode then uses the health information to monitor the status of DataNodes and verify that they’re live.

When monitoring, it’s important to prioritize analyzing metrics taken from the NameNode because if a NameNode fails, all the data within a cluster will become inaccessible to the user.

Prioritizing monitoring NameNode also makes sense as it enables you to ascertain the health of all the data nodes within a cluster. NameNode metrics can be broken down into two groups:

  • NameNode-emitted metrics
  • NameNode Java Virtual Machine (JVM) metrics

Below we’re going to list each group of metrics you can monitor and then show you a way to monitor these metrics for HDFS.

NameNode-emitted metrics

  • CapacityRemaining – Records the available capacity
  • CorruptBlocks/MissingBlocks – Records number of corrupt/missing blocks
  • VolumeFailuresTotal – Records number of failed volumes
  • NumLiveDataNodes/NumDeadDataNodes – Records count of alive or dead DataNodes
  • FilesTotal – Total count of files tracked by the NameNode
  • Total Load – Measure of file access across all DataNodes
  • BlockCapacity/BlocksTotal – Maximum number of blocks allocable/count of blocks tracked by NameNode
  • UnderReplicated Blocks – Number of under-replicated blocks
  • NumStaleDataNodes – Number of stale DataNodes

NameNode JVM Metrics

  • ConcurrentMarkSweep count – Number of old-generation collections
  • ConcurrentMarkSweep time – The elapsed time of old-generation collections, in milliseconds

How to Monitor HDFS Metrics

One way that you can monitor HDFS metrics is through Java Management Extensions (JMX) and the HDFS daemon web interface. To view a summary of NameNode and performance metrics enter the following URL into your web browser to access the web interface (which is available by default at port 50070):

http://<namenodehost>:50070

Here you’ll be able to see information on Configured Capacity, DFS Used, Non-DFS Used, DFS Remaining, Block Pool Used, DataNodes usage%, and more.

If you require more in-depth information, you can enter the following URL to view more metrics with a JSON output:

http://<namenodehost>:50070jmx 

MapReduce Counters

MapReduce is a software framework used by Hadoop to process large datasets in-parallel across thousands of nodes. The framework breaks down a dataset into chunks and stores them in a file system. MapReduce jobs are responsible for splitting the datasets and map tasks then process the data.

For performance monitoring purposes, you need to monitor MapReduce counters, so that you view information/statistics about job execution. Monitoring MapReduce counters enables you to monitor the number of rows read, and the number of rows written as output.

You can use MapReduce counters to find performance bottlenecks. There are two main types of MapReduce counters:

  • Built-in Counters – Counters that are included with MapReduce by default
  • Custom counters – User-defined counters that the user can create with custom code

Below we’re going to look at some of the built-in counters you can use to monitor Hadoop.

Built-In MapReduce Counters

Built-in Counters are counters that come with MapReduce by default. There are five main types of built-in counters:

  • Job counters
  • Task counters
  • File system counters
  • FileInputFormat Counters
  • FileOutput Format Counters

Job Counters 

MapReduce job counters measure statistics at the job level, such as the number of failed maps or reduces.

  • MILLIS_MAPS/MILLIS_REDUCES – Processing time for maps/reduces
  • NUM_FAILED_MAPS/NUM_FAILED_REDUCES – Number of failed maps/reduces
  • RACK_LOCAL_MAPS/DATA_LOCAL_MAPS/OTHER_LOCAL_MAPS – Counters tracking where map tasks were executed

Task Counters 

Task counters collect information about tasks during execution, such as the number of input records for reduce tasks.

  • REDUCE_INPUT_RECORDS – Number of input records for reduce tasks
  • SPILLED_RECORDS – Number of records spilled to disk
  • GC_TIME_MILLIS – Processing time spent in garbage collection

FileSystem Counters 

FileSystem Counters record information about the file system, such as the number of bytes read by the FileSystem.

  • FileSystem bytes read – The number of bytes read by the FileSystem
  • FileSystem bytes written – The number of bytes written to the FileSystem

FIleInputFormat Counters 

FileInputFormat Counters record information about the number of bytes read by map tasks

  • Bytes read – Displays the bytes read by map tasks with the specific input format

File OutputFormat Counters 

FileOutputFormat counters gather information on the number of bytes written by map tasks or reduce tasks in the output format.

  • Bytes written – Displays the bytes written by map and reduce tasks with the specified format

How to Monitor MapReduce Counters

You can monitor MapReduce counters for jobs through the ResourceManager web UI. To load up the ResourceManager web UI, go to your browser and enter the following URL:

http://<resourcemanagerhost>:8088

Here you will be shown a list of All Applications in a table format. Now, go to the application you want to monitor and click the History hyperlink in the Tracking UI column.

On the application page, click on the Counters option on the left-hand side. You will now be able to view counters associated with the job monitored.

YARN Metrics

Yet Another Resource Negotiator (YARN) is the component of Hadoop that’s responsible for allocating system resources to the applications or tasks running within a Hadoop cluster.

There are three main categories of YARN metrics:

  • Cluster metrics – Enable you to monitor high-level YARN application execution
  • Application metrics -Monitor execution of individual YARN applications
  • NodeManager metrics – Monitor information at the individual node level

Cluster Metrics 

Cluster metrics can be used to view a YARN application execution.

  • unhealthyNodes – Number of unhealthy nodes
  • activeNodes – Number of currently active nodes
  • lostNodes – Number of lost nodes
  • appsFailed – Number of failed applications
  • totalMB/allocatedMB – Total amount of memory/amount of memory allocated

Application metrics 

Application metrics provide in-depth information on the execution of YARN applications.

  • progress – Application execution progress meter

NodeManager metrics 

NodeManager metrics display information on resources within individual nodes.

  • containersFailed – Number of containers that failed to launch

How to Monitor YARN Metrics

To collect metrics for YARN, you can use the HTTP API. With your resource manager, host query the yarn metrics located on port 8088 by entering the following (use the qry parameter to specify the MBeans you want to monitor).

Resourcemanagerhost:8088/jmx?qry=java.lang:type=memory

ZooKeeper Metrics

ZooKeeper is a centralized service that maintains configuration information and delivers distributed synchronization across a Hadoop cluster. ZooKeeper is responsible for maintaining the availability of the HDFS NameNode and YARNs ResourceManager.

Some key ZooKeeper metrics you should monitor include:

  • zk_followers – Number of active followers
  • zk_avg_latency – Amount of time it takes to respond to a client request (in ms)
  • zk_num_alive_connections – Number of clients connected to ZooKeeper

How to Collect Zookeeper Metrics

There are a number of ways you can collect metrics for Zookeeper, but the easiest is by using the 4 letter word commands through Telnet or Netcat at the client port. To keep things simple, we’re going to look at the mntr, arguably the most important of the four 4 letter word commands.

$ echo mntr | nc localhost 2555

Entering the mntr command will return you information on average latency, maximum latency, packets received, packets sent, outstanding requests, number of followers, and more. You can view a list of four-letter word commands on the Apache ZooKeeper site.

Hadoop Monitoring Software

Monitoring Hadoop metrics through JMX or an HTTP API enables you to see the key metrics, but it isn’t the most efficient method of monitoring performance. The most efficient way to collect and analyze HDFS, MapReduce, Yarn, and ZooKeeper metrics, is to use an infrastructure monitoring tool or Hadoop monitoring software.

Many network monitoring providers have designed platforms with the capacity to monitor frameworks like Hadoop, with state-of-the-art dashboards and analytics to help the user monitor the performance of clusters at a glance. Many also come with custom alerts systems that provide you with email and SMS notifications when a metric hits a problematic threshold.

In this section, we’re going to look at some of the top Hadoop monitoring tools on the market. We’ve prioritized tools with high-quality visibility, configurable alerts systems, and complete data visualizations.

1. Datadog

Datadog App Analytics - Filters view

Datadog is a cloud monitoring tool that can monitor services and applications. With Datadog you can monitor the health and performance of Apache Hadoop. There is a Hadoop dashboard that displays information on DataNodes and NameNodes.

For example, you can view a graph of Disk remaining by DataNode, and TotalLoad by NameNode. Dashboards can be customized to add information from other systems as well. Integrations for HDFS, MapReduce, YARN, and ZooKeeper enable you to monitor the most significant performance indicators.

The alerts system makes it easy for you to track performance changes when they occur by providing you with automatic notifications. For example, Datadog can notify you if Hadoop jobs fail. The alerts system uses machine learning, which has the ability to identify anomalous behavior.

To give you greater control over your monitoring experience, Datadog provides full API access so that you can create new integrations. You can use the API access to complete tasks such as querying Datadog in the command-line or creating JSON-formatted dashboards.

Datadog is a great starting point for enterprises that want comprehensive Hadoop monitoring with wider cloud monitoring capabilities. The Infrastructure package of Datadog starts at $15 (£11.47) per host, per month. You can start the free trial version via this link here.

Key features:

  • Hadoop monitoring dashboard
  • Integrations for HDFS, MapReduce, YARN, and ZooKeeper
  • Alerts
  • Full API access

2. LogicMonitor

LogicMonitor Modules - Apache

LogicMonitor is an infrastructure monitoring platform that can be used for monitoring Apache Hadoop. LogicMonitor comes with a Hadoop package that can monitor HDFS NameNode, HDFS DataNode, Yarn, and MapReduce metrics. For monitoring Hadoop all you need to do is add Hadoop hosts to monitor, enable JMX on the Hadoop hosts, and assign properties to each resource.  The tool then collects Hadoop metrics through a REST API.

To monitor these metrics you can set alert trigger conditions to determine when alerts will be raised. Alerts can be assigned a numeric priority value to determine the severity of a breach. There is also an escalation chain you can use to escalate alerts that haven’t been responded to.

For more general monitoring activity, LogicMonitor includes a dashboard that you can use to monitor your environment with key metrics and visualizations including graphs and charts. The software also allows you to schedule reports to display performance data. For example, the Alert Trends report provides a summary of alerts that occurred for resources/groups over a period of time.

LogicMonitor is ideal for enterprises that want to monitor Apache Hadoop alongside other applications and services. The tool has a custom pricing model so you need to contact the company directly to request a quote. You can start the free trial version via this link here.

Key features:

  • Monitors HDFS NameNode, HDFS DataNode, Yarn, and MapReduce metrics
  • REST API
  • Custom alert thresholds
  • Dashboard
  • Reports

3. Dynatrace

Dynatrace Technology Overview dashboard - Apache Hadoop monitoring

Dynatrace is an application performance management tool you can use to monitor services and applications. Dynatrace also offers users performance monitoring for Hadoop. The platform can automatically detect Hadoop components and display performance metrics for HDFS and MapReduce. Whenever a new host running Hadoop is added to your environment the tool detects it automatically.

You can monitor a range of NameNode and DataNode metrics. NameNode metrics include Total, Used, Remaining, Total load, Total, Pending deletion, Files total, Under replicated, Live, Capacity, and more. Types of DataNode metrics include Capacity, Used, Cached, Failed to Cache, Blocks, Removed, Replicated, and more.

Dynatrace Apache Hadoop monitoring dashboard

Dashboards provide a range of information with rich data visualizations. For example, you can view a chart of MapReduce maps failed or a bar graph of Jobs preparing and running. Custom alerts powered by anomaly detection enable you to identify performance issues, helping you to make sure that your service stays available.

Dynatrace is not only a top application monitoring tool but a formidable choice for monitoring Hadoop as well. There are a range of packages available including the Infrastructure monitoring package at $2 (£1.53) per month and the Full-stack monitoring package at $69 (£52.78) per month. You can start the 15-day free trial via this link here.

Key features:

  • Automatically detects Hadoop components
  • DataNode and NameNode metrics
  • Analytics and Data visualizations
  • Dashboards
  • Custom alerts

Choosing Hadoop Monitoring Software for Cluster Performance 

Monitoring Hadoop metrics is vital for making sure that your clusters stay up and running. While you can attempt to monitor Hadoop metrics through JMX or an HTTP API, it doesn’t offer the complete monitoring experience that many infrastructure monitoring tools like Datadog, LogicMonitor, and Dynatrace do.

These tools offer features like custom dashboards and alerts that provide you with a more holistic perspective of what’s going on. By collecting all of your Hadoop performance data and putting it in one place, you’ll be able to monitor the performance of your systems much more effectively.