Debugging Apache Hadoop YARN Cluster in Production

Debugging Apache
Hadoop YARN Cluster in
Production
Jian He, Junping Du and Xuan Gong
Hortonworks YARN Team
06/30/2016

2 © Hortonworks Inc. 2011 – 2016. All Rights Reserved
Who are We
 Junping Du
– Apache Hadoop Committer and PMC
Member
– Dev Lead in Hortonworks YARN team
 Xuan Gong
Member
– Software Engineer
 Jian He
Member
– Staff Software Engineer

Today’s Agenda
 YARN in a Nutshell
 Trouble-shooting Process and Tools
 Case Study
 Enhanced YARN Log Tool Demo
 Summary and Future

Agenda
YARN in a Nutshell

YARN Architecture
 ResourceManager
 NodeManager
 ApplicationMaster
 Other daemons:
– Application
History/Timeline Server
– Job History Server (for
MR only)
– Proxy Server
– Etc.

RM and NM in a nutshell

Agenda
YARN in a Nutshell
Trouble-shooting Process and Tools

“Troubles” to start troubleshooting effort on a YARN cluster
 Applications Failed
 Applications Hang/Slow
 YARN configuration doesn’t work
 YARN APIs (CLI, WebService, etc.) doesn’t work
 YARN daemons crashed (OOM issue, etc.)
 YARN daemons’ log has error/warnings
 YARN cluster monitoring tools (like Ambari) alert
Problem Type Distribution
Configuration
Executing Jobs
Cluster Administration
Installation
Application Development
Performance

Process: Phenomenon -> Root Cause -> Solution
 Solution:
– Infrastructure/Hardware issue
• Replace disks
• Fix network
– Mis-configuration
• Fix configuration
• Enhance documentation
– Setup issue
• Fix setup
• Restart services
– Application issue
• Update application
• Workaround
– A YARN Bug
• Report/fix it in Apache
community!
 Phenomenon:
– Application Failed
 Root cause:
– Container Launch failures
• Classpath issue
• Resource localization
failures
– Too many attempt failures
• Network connection
issue
• NM disk issues
• AM failed caused by
node restarted
– Application logic issue
• Container failed with
OOM, etc.
– Security issue
• Token related issues

Iceberg of troubleshooting – Case Study
 "java.lang.RuntimeException: java.io.FileNotFoundException:
/etc/hadoop/2.3.4.0-3485/0/core-site.xml (Too many open files in
system)”
 That actually due to too many TCP connections issue

Iceberg of troubleshooting – Dig Deeply
 Most connections are from local NM to DNs
– LogAggregationService
– ResourceLocalizationService
 We found the root cause is threads leak on NM LogAggregationService:
– YARN-4697
NM aggregation thread pool is not bound by limits
– YARN-4325
Purge app state from NM state-store should cover more LOG_HANDLING cases
– YARN-4984
LogAggregationService shouldn't swallow exception in handling createAppDir() which cause thread
leak.

Lesson Learned for Trouble-shooting on a production cluster
 What’s mean by a “Production” Cluster?
– Cannot afford stop/restart cluster for trouble shooting
– Most operations on cluster are “Read Only”
– In fenced network, remote debugging with local cluster admin.
 Lesson learned:
1. Get related info (screenshots, log files, jstack, memory heap dump, etc.) as much as you can
2. Work closely with the end user to gain an understanding of the issue and symptoms
3. Setup knowledge base used to compare to previous cases
4. If possible, reproduce the issue on test/backup cluster – easy to trouble shooting and verify
5. Version your configuration!

Handy Tools for YARN Troubleshooting
 Log
 UI
 Historic Info
– JobHistoryServer (for MR only)
– Application Timeline Service (v1, v1.5, v2.0)
 Monitoring tools, like: AMBARI
 Runtime info
– Memory Dump
– Jstack
– System Metrics

Log
 Log CLI
– yarn logs -applicationId <application ID> [OPTIONS]
– Discuss more later
 Enable Debug log
– When daemons are NOT running
• Put log level settings like: export YARN_ROOT_LOGGER = “DEBUG, console” to yarn-env.sh
• Start the daemons
– When Daemons are running
• Dynamic change log level via daemon’s logLevel UI/CLI
• CLI:
– yarn daemonlog [-getlevel <host:httpPort> <classname>]
– yarn daemonlog [-setlevel <host:httpPort> <classname> <level>]
– for YARN Client side
• Similar setting as daemons not running

Runtime Log Level settings in YARN UI
 RM: http://<rm_addr>:8088/logLevel
 NM: http://<nm_addr>:8042/logLevel
 ATS: http://<ats_addr>:8188/logLevel

UI (Ambari and YARN)

Job History Server

Memory dump analysis

Hadoop metrics
 RPC metrics
– RpcQueueTimeAvgTime
– ReceivedBytes
…
 JVM metrics
– MemHeapUsedM
– ThreadsBlocked
…
 Documentation:
– http://s.apache.org/UwSu

YARN top
 top like command line view for application stats, queue stats

Agenda
YARN in a Nutshell
Case Study

Why is my job hung ?
 Job can be stuck at 3 states.
NEW_SAVING: Waiting for app to be persisted in state-store
- Connection error with state-store (zookeeper etc.)
Accepted: Waiting to allocate ApplicationMaster container.
- Low max- AM-resource-percentage config
Running: waiting for containers to be allocated?
- Are there resources available for the app
- Otherwise, application land issue, stuck on
socket read/write.
App
states

Case Study
 Friday evening, Customer experiences cluster outages.
 Large amount of jobs getting stuck.
 There are resources available in the cluster.
 Restarting Resource Manger can resolve issue temporarily
 But after several hours, cluster again goes back to the bad state

Case Study
 Are there any resources available in the queue ?

Case Study
 Are there any resources available for the app ?
– Sometimes, even if cluster has resources, user may still not be able to run their applications
because they hit the user-limit.
– User-limit controls how much resources a single user can use
– Check user-limit info on the scheduler UI
– Check application head room on application UI

Case study
 Not a problem of resource contention.
 Use yarn logs command to get hung application logs.
– Found app waiting for containers to be allocated.
 Problem: cluster has free resources, but app is not able to use it.

Case study
 May be a scheduling issue.
 Analyze the scheduler log. (Most difficult)
– User not much familiar with the scheduler log.
– RM log is too huge, hard to do text searching in the logs.
– Getting worse if enabling debug log.
 Dump the scheduling log into a separate file

Case study
 Scheduler log shows several apps are skipped for scheduling.
 Pick one of the applications, go to the application attempt UI,
 Check the resource requests table (see below), notice billions of containers are asked by
the application.
8912124

Case study
 Tried to kill those misbehaving jobs, cluster went fine.
 Find the user who submit those jobs and stop him/her from doing that.
 Big achievement so far, unblock the cluster.
 Offline debugging and find product bug.
 Surprisingly, we use int for measuring memory size in the scheduler.
 That misbehaving app asked too much resources, which caused integer overflow in the
scheduler.
 YARN-4844, replace int with long for resource memory API.

What we learn
 Rebooting service can solve many problems. 
– Thanks to working-preserving RM and NM recovery (YARN-556 & YARN-1336).
 Denial of Service - Poorly written, or accidental configuration for workloads can cause
component outages.
– Carefully code against DOS scenarios.
– Example: User RPC method (getQueueInfo) holds scheduler lock
 UI enhancement
– Small change, big impact.
– Example: Resource requests table on application very useful in this case.
 Alerts
– Ask too many containers, alerting to the users.

Case study 2
 10 % of the jobs are failing every day.
 After they re-run, jobs sometime finish successfully.
 No resource contention when jobs are running
 Logs contain a lot of mysterious connection errors (unable to read call parameters)

Case study 2
 Initial attempt,
– Dig deeper into the code to see under what conditions, this exception may throw.
– Not able to figure out.

Case study 2
 Requested more failed application logs
 Identify pattern for these applications
 Finally, we realize all apps failed on a certain set of nodes.
 Ask customer to exclude those nodes. Jobs running fine after that.
 Customer checked “/var/log/messages” and found disk issues for those nodes.
When dealing with mysterious connection
failures, hung problems, try to find
correlation between failed apps and nodes.

Agenda
YARN in a Nutshell
Case Study
Enhanced YARN Log Tool Demo

Enhanced YARN Log CLI (YARN-4904)
 Useful Log CLIs
– Get container logs for running apps
• yarn logs –applicationId ${appId}
– Get a specific container log
• yarn logs –applicationId ${appId} –containerId ${containerId}
– Get AM Container logs.
• yarn logs -applicationId ${appId} –am 1
– Get a specific log file
• yarn logs -applicationId ${appId} –logFiles syslog
• Support java regular expression
– Get the log file's first 'n' bytes or the last 'n' bytes
• yarn logs –applicationId ${appId} –size 100
– Dump the application/container logs
• yarn logs –applicationId ${appId} –out ${local_dir}– List application/container log information
• yarn logs –applicationId ${appId} -show_application_log_info
• yarn logs –applicationId ${appId} –containerId ${containerId} -show_container_log_info

Agenda
YARN in a Nutshell
Case Study
Enhanced YARN Log Tool Demo
Summary and Future

Summary and Future
 Summary
– Methodology and Tools for trouble-shooting on YARN
– Case Study
– Enhanced YARN Log CLI
• YARN-4904
 Future Enhancement
– ATS (Application Timeline Service) v2
• YARN-2928
• #hs16sj “How YARN Timeline Service v.2 Unlocks 360-Degree Platform Insights at
Scale”
– New ResourceManager UI
• YARN-3368

New RM UI (YARN-3368)

Thank You

Backup Slides

YARN Log Command example Screenshot
yarn logs –applicationId application_1467090861129_0001 –am 1

yarn logs –applicationId application_1467090861129_0001 –containerId container_1467090861129_0001_01_000002

yarn logs –applicationId application_1467090861129_0001 –am 1 –logFiles stderr

yarn logs –applicationId application_1467090861129_0001 –am 1 –logFiles stderr –size -1000

yarn logs –applicationId application_1467090861129_0001 –out ${localDir}

yarn logs –applicationId application_1467090861129_0001 –-show_application_log_info
yarn logs –applicationId application_1467090861129_0001 –-show_container_log_info

Debugging Apache Hadoop YARN Cluster in Production

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Debugging Apache Hadoop YARN Cluster in Production (20)

Recently uploaded (20)

Debugging Apache Hadoop YARN Cluster in Production