Innovations in Apache Hadoop MapReduce, Pig and Hive for improving query performance

Innovations In Apache Hadoop MapReduce,
Pig and Hive for improving query
performance

gopalv@apache.org
vinodkv@apache.org

Page 1

Operation Stinger

© Hortonworks Inc. 2013 Page 3

Performance at any cost

© Hortonworks Inc. 2013

• Scalability
– Already works great, just don’t break it for performance gains
• Isolation + Security
– Queries between different users run as different users
• Fault tolerance
– Keep all of MR’s safety nets to work around bad nodes in clusters
• UDFs
– Make sure they are “User” defined and not “Admin” defined


First things first
• How far can we push Hive as it exists today?


Benchmark spec
• The TPC-DS benchmark data+query set
• Query 27 (big joins small)
– For all items sold in stores located in specified states during a given
year, find the average quantity, average list price, average list sales
price, average coupon amount for a given gender, marital status,
education and customer demographic.
• Query 82 (big joins big)
– List all items and current prices sold through the store channel from
certain manufacturers in a given price range and consistently had a
quantity between 100 and 500 on hand in a 60-day period.


TL;DR
• TPC-DS Query 27, Scale=200, 10 EC2 nodes (40 disks)


TL;DR - II
• TPC-DS Query 82, Scale=200, 10 EC2 nodes (40 disks)


Forget the actual benchmark
• First of all, YMMV
– Software
– Hardware
– Setup
– Tuning
• Text formats seem to be the staple of all comparisons
– Really?
– Everybody’s using it but only for benchmarks!


What did the trick?
• Mapreduce?
• HDFS?
• Or is it just Hive?


Optional Advice


RCFile
• Binary RCFiles
• Hive pushes down column projections
• Less I/O, Less CPU
• Smaller files


Data organization
• No data system at scale is loaded once & left alone
• Partitions are essential
• Data flows into new partitions every day


A closer look
• Now revisiting the benchmark and its results


Query27 - Before


Before


Query 27 - After


After


Query 82 - Before


Query 82 - After


What changed?
• Job Count/Correct plan
• Correct data formats
• Correct data organization
• Correct configuration


What changed?
Data Formats
Data Organization

Query Plan


Is that all?
• NO!
• In Hive
– Metastore
– RCFile issues
– CPU intensive code
• In YARN+MR
– Parallelism
– Spin-up times
– Data locality
• In HDFS
– Bad disks/deteriorating nodes


In Hive
• NO!
• In Hive
– Metastore
– RCFile issues
• In YARN+MR
– Parallelism
– Spin-up times
– Data locality
• In HDFS


Hive Metastore
• 1+N Select problem
– SELECT partitions FROM tables;
– /* for each needed partition */ SELECT * FROM Partition ..
– For query 27 , generates > 5000 queries! 4-5 seconds lost on each call!
– Lazy loading or Include/Join are general solutions
• Datanucleus/ORM issues
– 100K NPEs try.. Catch.. Ignore..
• Metastore DB Schema revisit
– Denormalize some/all of it?


RCFile issues
• RCFiles do not split well
– Row groups and row group boundaries
• Small row groups vs big row groups
– Sync() vs min split
– Storage packing
• Run-length information is lost
– Unnecessary deserialization costs


ORC file format
• A single file as output of each task.
– Dramatically simplifies integration with Hive
– Lowers pressure on the NameNode
• Support for the Hive type model
– Complex types (struct, list, map, union)
– New types (datetime, decimal)
– Encoding specific to the column type
• Split files without scanning for markers
• Bound the amount of memory required for
reading or writing.


CPU intensive code


CPU intensive code
• Hive query engine processes one row at a time
– Very inefficient in terms of CPU usage
• Lazy deserialization: layers
• Object inspector calls
• Lots of virtual method calls


Tighten your loops


Vectorization to the rescue
• Process a row batch at a time instead of a single row
• Row batch to consist of column vectors
– The column vector will consist of array(s) of primitive types as far as
possible
• Each operator will process the whole column vector at a
time
• File formats to give out vectorized batches for processing
• Underlying research promises
– Better instruction pipelines and cache usage
– Mechanical sympathy


Vectorization: Prelim results
• Functionality
– Some arithmetic operators and filters using primitive type columns
– Have a basic integration benchmark to prove that the whole setup
works
• Performance
– Micro benchmark
– More than 30x improvement in the CPU time
– Disclaimer:
– Micro benchmark!
– Include io or deserialization costs or complex and string datatypes


In YARN+MR
• NO!
• In Hive
– Metastore
– RCFile issues
• In YARN+MR
– Data locality
– Parallelism
– Spin-up times
• In HDFS


Data Locality
• CombineInputFormat
• AM interaction with locality
• Short-circuit reads!
• Delay scheduling
– Good for throughput
– Bad for latency


Parallelism
• Can tune it (to some extent)
– Controlling splits/reducer count
• Hive doesn’t know dynamic cluster status
– Benchmarks max out clusters, real jobs may or may not
• Hive does not let you control parallelism
– particularly in case of multiple jobs in a query


Spin up times
• AM startup costs
• Task startup costs
• Multiple waves of map tasks


Apache Tez
• Generic DAG workflow
• Container re-use
• AM pool service


AM Pool Service
• Pre-launches a pool of AMs
• Jobs submitted to these pre-launched AMs
– Saves 3-5 seconds
• Pre-launched AMs can pre-allocate containers
• Tasks can be started as soon as the job is submitted
– Saves 2-3 seconds


Container reuse
• Tez MapReduce AM supports Container reuse
• Launched JVMs are re-used between tasks
– about 4-5 seconds saved in case of multiple waves
• Allows future enhancements
– re-using task data structures across splits


In HDFS
• NO!
• In Hive
– Metastore
– RCFile issues
• In YARN+MR
– Data locality
– Parallelism
– Spin-up times
• In HDFS


Speculation/bad disks
• No cluster remains at 100% forever
• Bad disks cause latency issues
– Speculation is one defense, but it is not enough
– Fault tolerance is a safety net
• Possible solutions:
– More feedback from HDFS about stale nodes, bad/slow disks
– Volume scheduling


General guidelines
• Benchmarking
– Be wary of benchmarks! Including ours!
– Algebra with X


General guidelines contd.
• Benchmarks: To repeat, YMMV.
• Benchmark *your* use-case.
• Decide your problem size
– If (smallData) {
Mysql/Postgres/Your smart phone
} else {
–Make it work
–Make it scale
–Make it faster
}
• If it is (seems to be) slow, file a bug, spend a little time!
• Replacing systems without understanding them
– Is an easy way to have an illusion of progress


Related talks
• “Optimizing Hive Queries” by Owen O’Malley
• “What’s New and What’s Next in Apache Hive” by Gunther
Hagleitner


Credits
• Arun C Murthy
• Bikas Saha
• Gopal Vijayaraghavan
• Hitesh Shah
• Siddharth Seth
• Vinod Kumar Vavilapalli
• Alan Gates
• Ashutosh Chauhan
• Vikram Dixit
• Gunther Hagleitner
• Owen O’Malley
• Jintendranath Pandey
• Yahoo!, Facebook, Twitter, SAP and Microsoft all contributing.


Q&A
• Thanks!


Innovations in Apache Hadoop MapReduce, Pig and Hive for improving query performance

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to Innovations in Apache Hadoop MapReduce, Pig and Hive for improving query performance (20)

Recently uploaded (20)

Innovations in Apache Hadoop MapReduce, Pig and Hive for improving query performance

Editor's Notes