Get most out of Spark on YARN

© Hortonworks Inc. 2014
Get most out of Spark on YARN
Oleg Zhurakousky, Hortonworks @z_oleg

Spark
“Apache Spark™ is a general engine for data processing.”
val file = spark.textFile("hdfs://...")
val counts = file.flatMap(line => line.split(" "))
.map(word => (word, 1))
.reduceByKey(_ + _)
counts.saveAsTextFile("hdfs://...")
Word Count in Spark’s Scala API

Spark 101
• Spark provides:
– API
– Task Execution engine
– Libraries for SQL, Machine Learning, Graph
• Resilient Distributed Dataset (RDD)
– Immutable, distributed data collection
– 2 Types of RDD Functions
– Transformations (e.g. map, filter….) create new RDD
– Actions (e.g. collect, count) lead of Spark DAG execution
– Spark Transformations are pipelined until actions are called
– Spark provides factory methods to create RDDs
– From in-memory Collection
– from various data sources – E.g. HDFS, Local files
– You can create RDD manually as well

Spark 101 - 2
• Spark Driver
– Client side application that creates Spark Context
• Spark Context
– Talks to Spark Driver, Cluster Manager to Launch Spark Executors
• Cluster Manager – E.g YARN, Spark Standalone, MESOS
• Executors – Spark worker bees

Demo

Extensibility – Why?
• Integrate with native features of target system (e.g., YARN Shuffle)
• Optimize by benefiting form side effects
• KV Grouping
• Sorting
• Unified security and monitoring
• Hybrid execution environment
• Streaming and Micro-batching
• Cross-context data sharing
• Specialization over generalization

Extensibility
Spark is a Framework
with many extensibility points
• RDD
• Add, additional operations, optimize existing operations
• SparkContext
• Execution delegation, hybrid execution environments
• ShuffleReaders/ShuffleWriters
• Delegate reads/writes to other source/target storage systems
• Additional operations
• RDD and Context
• Data Broadcasting and Caching
• Cross context sharing of cached/broadcasted data
• More. . .

© Hortonworks Inc. 2011 – 2014. All Rights Reserved
Spark on YARN – Why?
Page 10
Workloads Run Natively IN Hadoop
HDFS (Redundant, Reliable Storage)
YARN (Cluster Resource Management)
BATCH
(MapReduce)
SQL
(Tez)
STREAMING
(Storm, S4,…)
GRAPH
(Giraph)
SPARK
HPC MPI
(OpenMPI)
ONLINE
(HBase)
OTHER
(Search)
(Weave…)
Run the Spark workload IN Hadoop
Instead of shipping the data to the code
with Predictable Performance and Quality of Service

Spark On YARN Benefits
Ship the code to the data instead of the other way around
Efficient locality-aware access to data stored in HDFS
Leverage existing Hadoop clusters
Single provisioning, management, monitoring story
Simplified deployment and operations
Scale-up for production use with minimal IT involvement
Secure, multi-tenant workload management

Demo

Spark on YARN – Key Features
Flexible deployment options
Support for secure mode
Monitoring and metrics
Distributed cache and local resource management

Integration with Hadoop Security
• Kerberos support - kinit & submit:
• kinit -kt /etc/security/keytabs/myuser.headless.keytab myuser
• HiveContext - accessing the Hive Metastore in secure mode
• Minimum configuration: hive.metastore.uris
• Be careful with extra configuration, as spark cannot support some of hive configurations.
• Need the patch Spark-5111 for basic Kerberos setup working against Hadoop-2.6 (connecting
to hive metastore).
• Spark ThriftServer:
• Minimum configuration: Kerberos keytab/principal, sasl, authorization,
• Spark thrift server has to be co-located with hive thrift server.
• Spark user has to be able to access the keytab of hive.

Application Timeline Service (ATS) Integration

OrcFile support
• saveAsOrcFile
• save the table into ORC Format File
• orcFile
• Read orc format file as table source.
• Other features:
• Column pruning, self-contained schema support, predicate push down, different compression
method.
• External data source API:
• Refer to latest PR in the JIRA(SPARK-2883) with the support.
• Import org.apache.spark.sql.hive.orc._
• Operate under HiveContext
Tech Preview in HDP 2.2, targeted for
rewrite to Spark Data Source API

Spark on YARN - Beyond the basics -
Multi-tenancy & workload management

Key Themes
• Security: Kerberos Token Renewal
• Log Aggregation: View logs for running applications
• Fault Tolerance: AM Retry and Container keep alive
• Service Registry: Directory for services running in YARN
Long
Running
Services
Support
THEME
• CPU Scheduling
• CPU Isolation through CGroups
• Node Labels for scheduling constraints
Workload
Mgmt
THEME
• YARN Rolling Upgrades support
• Work Preserving Restart
• Timeline Server support in Secure clusters
Reliable and
Secure
Operations
THEME

Enhancements to Support Long Running Services
• YARN-941: YARN updates Kerberos token for a Long Running Service
after the token expires
Security
Capability
• YARN-2468: Aggregate and capture logs during the lifetime of a
Long Running Service
Log
Aggregation
Capability
• YARN-1489: When ApplicationMaster(AM) restarts, do not kill all
associated containers – reconnect to the AM
• YARN-611/YARN-614: Tolerate AM failures for Long Running Services
Fault
Tolerance
Capability
• YARN-913: Service Registry that publishes host and ports that each
service comes up on
Service
Registry
Capability

Node Labels: Apply Node Constraints
A
App
L1 L1 L1
Deploy/Allocate
B
App
L1 L1 L1
Isolate
A A A
nodes
labels

CPU Scheduling
What
• Admin tells YARN how much CPU capacity is available in cluster
• Applications specify CPU capacity needed for each container
• YARN Capacity Scheduler schedules application taking CPU capacity availability into account
Why
• Applications (for example Storm, HBase, Machine Learning) need predictable access to CPU
as a resource
• CPU has become bottleneck instead of memory in certain clusters (128 GB RAM, 6 CPUs)

CGroup Isolation
What
• Admin enables CGroups for CPU Isolation for YARN application workloads
Why
• Applications need guaranteed access to CPU resources
• To ensure SLAs, need to enforce CPU allocations given to an Application container

Default Queue Mapping
What
• Admin can define a default queue per user or group
• If no queue is specified for an application, YARN Capacity Scheduler will use the user/group’s
default queue for application
Why
• Queues are required for enforcing SLAs, make it easy to utilize queues
• Users and Applications want to submit Yarn apps/jobs and not have to specify queue

ResourceManager Rest API
What
• Submit YARN applications through REST
• Get YARN application status through REST
• Kill YARN application status through REST
Why
• External Applications need to interact (schedule/monitor) with YARN without needing to
embed Java library
• Enable administration from a remote system

When Things go wrong
• Where to look
– yarn application –list (get the list of running application)
– yarn logs -applicationId <app_id>
– Check Spark Environment : http://<host>:8088/proxy/<job_id>/environment/
• Common Issues
– Submitted a job but nothing happens
– Job stays in accepted state when allocated more memory/cores than is available
– May need to kill unresponsive/stale jobs
– Insufficient HDFS access
– May lead to failure such as
“Loading data to table default.testtable
Failed with exception Unable to move sourcehdfs://red1:8020/tmp/hive-spark/hive_2015-03-04_12-45-
42_404_3643812080461575333-1/-ext-10000/kv1.txt to destination hdfs://red1:8020/apps/hive/warehouse/testtable/kv1.txt”
– Wrong host in Beeline, shows error as invalid URL
– “Error: Invalid URL: jdbc:hive2://localhost:10001 (state=08S01,code=0)”
– Error about closed SQLContext, restart Thrift Server
Grant user/group necessary HDF

Thank you!

Get most out of Spark on YARN

Related slideshows

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (15)

Similar to Get most out of Spark on YARN

Similar to Get most out of Spark on YARN (20)

More from DataWorks Summit

More from DataWorks Summit (20)

Recently uploaded

Recently uploaded (20)

Get most out of Spark on YARN