Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

Apache Spark Architecture

•
•

This is the presentation I made on JavaDay Kiev 2015 regarding the architecture of Apache Spark. It covers the memory model, the shuffle implementations, data frames and some other high-level staff and can be used as an introduction to Apache Spark

1 of 114

1

1Pivotal Confidential–Internal Use Only 1Pivotal Confidential–Internal Use Only Spark Architecture A.Grishchenko

2

About me Enterprise Architect @ Pivotal  7 years in data processing  5 years with MPP  4 years with Hadoop  Spark contributor  http://0x0fff.com

3

Outline  Spark Motivation  Spark Pillars  Spark Architecture  Spark Shuffle  Spark DataFrame

4

Outline  Spark Motivation  Spark Pillars  Spark Architecture  Spark Shuffle  Spark DataFrame

5

Spark Motivation  Difficultly of programming directly in Hadoop MapReduce

6

Spark Motivation  Difficultly of programming directly in Hadoop MapReduce  Performance bottlenecks, or batch not fitting use cases

7

Spark Motivation  Difficultly of programming directly in Hadoop MapReduce  Performance bottlenecks, or batch not fitting use cases  Better support iterative jobs typical for machine learning

8

Difficulty of Programming in MR Word Count implementations  Hadoop MR – 61 lines in Java  Spark – 1 line in interactive shell sc.textFile('...').flatMap(lambda x: x.split()) .map(lambda x: (x, 1)).reduceByKey(lambda x, y: x+y) .saveAsTextFile('...') VS

9

Performance Bottlenecks How many times the data is put to the HDD during a single MapReduce Job?  One  Two  Three  More

10

Performance Bottlenecks How many times the data is put to the HDD during a single MapReduce Job?  One  Two  Three  More

11

Performance Bottlenecks Consider Hive as main SQL tool

12

Performance Bottlenecks Consider Hive as main SQL tool  Typical Hive query is translated to 3-5 MR jobs

13

Performance Bottlenecks Consider Hive as main SQL tool  Typical Hive query is translated to 3-5 MR jobs  Each MR would scan put data to HDD 3+ times

14

Performance Bottlenecks Consider Hive as main SQL tool  Typical Hive query is translated to 3-5 MR jobs  Each MR would scan put data to HDD 3+ times  Each put to HDD – write followed by read

15

Performance Bottlenecks Consider Hive as main SQL tool  Typical Hive query is translated to 3-5 MR jobs  Each MR would scan put data to HDD 3+ times  Each put to HDD – write followed by read  Sums up to 18-30 scans of data during a single Hive query

16

Performance Bottlenecks Spark offers you  Lazy Computations – Optimize the job before executing

17

Performance Bottlenecks Spark offers you  Lazy Computations – Optimize the job before executing  In-memory data caching – Scan HDD only once, then scan your RAM

18

Performance Bottlenecks Spark offers you  Lazy Computations – Optimize the job before executing  In-memory data caching – Scan HDD only once, then scan your RAM  Efficient pipelining – Avoids the data hitting the HDD by all means

19

Outline  Spark Motivation  Spark Pillars  Spark Architecture  Spark Shuffle  Spark DataFrame

20

Spark Pillars Two main abstractions of Spark

21

Spark Pillars Two main abstractions of Spark  RDD – Resilient Distributed Dataset

22

Spark Pillars Two main abstractions of Spark  RDD – Resilient Distributed Dataset  DAG – Direct Acyclic Graph

23

RDD  Simple view – RDD is collection of data items split into partitions and stored in memory on worker nodes of the cluster

24

RDD  Simple view – RDD is collection of data items split into partitions and stored in memory on worker nodes of the cluster  Complex view – RDD is an interface for data transformation

25

RDD  Simple view – RDD is collection of data items split into partitions and stored in memory on worker nodes of the cluster  Complex view – RDD is an interface for data transformation – RDD refers to the data stored either in persisted store (HDFS, Cassandra, HBase, etc.) or in cache (memory, memory+disks, disk only, etc.) or in another RDD

26

RDD  Complex view (cont’d) – Partitions are recomputed on failure or cache eviction

27

RDD  Complex view (cont’d) – Partitions are recomputed on failure or cache eviction – Metadata stored for interface ▪ Partitions – set of data splits associated with this RDD

28

RDD  Complex view (cont’d) – Partitions are recomputed on failure or cache eviction – Metadata stored for interface ▪ Partitions – set of data splits associated with this RDD ▪ Dependencies – list of parent RDDs involved in computation

29

RDD  Complex view (cont’d) – Partitions are recomputed on failure or cache eviction – Metadata stored for interface ▪ Partitions – set of data splits associated with this RDD ▪ Dependencies – list of parent RDDs involved in computation ▪ Compute – function to compute partition of the RDD given the parent partitions from the Dependencies

30

RDD  Complex view (cont’d) – Partitions are recomputed on failure or cache eviction – Metadata stored for interface ▪ Partitions – set of data splits associated with this RDD ▪ Dependencies – list of parent RDDs involved in computation ▪ Compute – function to compute partition of the RDD given the parent partitions from the Dependencies ▪ Preferred Locations – where is the best place to put computations on this partition (data locality)

31

RDD  Complex view (cont’d) – Partitions are recomputed on failure or cache eviction – Metadata stored for interface ▪ Partitions – set of data splits associated with this RDD ▪ Dependencies – list of parent RDDs involved in computation ▪ Compute – function to compute partition of the RDD given the parent partitions from the Dependencies ▪ Preferred Locations – where is the best place to put computations on this partition (data locality) ▪ Partitioner – how the data is split into partitions

32

RDD  RDD is the main and only tool for data manipulation in Spark  Two classes of operations – Transformations – Actions

33

RDD Lazy computations model Transformation cause only metadata change

34

DAG Direct Acyclic Graph – sequence of computations performed on data

35

DAG Direct Acyclic Graph – sequence of computations performed on data  Node – RDD partition

36

DAG Direct Acyclic Graph – sequence of computations performed on data  Node – RDD partition  Edge – transformation on top of data

37

DAG Direct Acyclic Graph – sequence of computations performed on data  Node – RDD partition  Edge – transformation on top of data  Acyclic – graph cannot return to the older partition

38

DAG Direct Acyclic Graph – sequence of computations performed on data  Node – RDD partition  Edge – transformation on top of data  Acyclic – graph cannot return to the older partition  Direct – transformation is an action that transitions data partition state (from A to B)

39

DAG WordCount example

40

DAG WordCount example HDFSInputSplits HDFS RDDPartitions RDD RDD RDD RDD RDD sc.textFile(‘hdfs://…’) flatMap map reduceByKey foreach

41

Outline  Spark Motivation  Spark Pillars  Spark Architecture  Spark Shuffle  Spark DataFrames

42

Spark Cluster Driver Node … Worker Node Worker Node Worker Node

43

Spark Cluster Driver Node Driver … Worker Node … Executor Executor Worker Node … Executor Executor Worker Node … Executor Executor

44

Spark Cluster Driver Node Driver Spark Context … Worker Node … Executor Cache Executor Cache Worker Node … Executor Cache Executor Cache Worker Node … Executor Cache Executor Cache

45

Spark Cluster Driver Node Driver Spark Context … Worker Node … Executor Cache Task … Task Task Executor Cache Task … Task Task Worker Node … Executor Cache Task … Task Task Executor Cache Task … Task Task Worker Node … Executor Cache Task … Task Task Executor Cache Task … Task Task

46

Spark Cluster  Driver – Entry point of the Spark Shell (Scala, Python, R)

47

Spark Cluster  Driver – Entry point of the Spark Shell (Scala, Python, R) – The place where SparkContext is created

48

Spark Cluster  Driver – Entry point of the Spark Shell (Scala, Python, R) – The place where SparkContext is created – Translates RDD into the execution graph

49

Spark Cluster  Driver – Entry point of the Spark Shell (Scala, Python, R) – The place where SparkContext is created – Translates RDD into the execution graph – Splits graph into stages

50

Spark Cluster  Driver – Entry point of the Spark Shell (Scala, Python, R) – The place where SparkContext is created – Translates RDD into the execution graph – Splits graph into stages – Schedules tasks and controls their execution

51

Spark Cluster  Driver – Entry point of the Spark Shell (Scala, Python, R) – The place where SparkContext is created – Translates RDD into the execution graph – Splits graph into stages – Schedules tasks and controls their execution – Stores metadata about all the RDDs and their partitions

52

Spark Cluster  Driver – Entry point of the Spark Shell (Scala, Python, R) – The place where SparkContext is created – Translates RDD into the execution graph – Splits graph into stages – Schedules tasks and controls their execution – Stores metadata about all the RDDs and their partitions – Brings up Spark WebUI with job information

53

Spark Cluster  Executor – Stores the data in cache in JVM heap or on HDDs

54

Spark Cluster  Executor – Stores the data in cache in JVM heap or on HDDs – Reads data from external sources

55

Spark Cluster  Executor – Stores the data in cache in JVM heap or on HDDs – Reads data from external sources – Writes data to external sources

56

Spark Cluster  Executor – Stores the data in cache in JVM heap or on HDDs – Reads data from external sources – Writes data to external sources – Performs all the data processing

57

Executor Memory

58

Spark Cluster – Detailed

59

Spark Cluster – PySpark

60

Application Decomposition  Application – Single instance of SparkContext that stores some data processing logic and can schedule series of jobs, sequentially or in parallel (SparkContext is thread-safe)

61

Application Decomposition  Application – Single instance of SparkContext that stores some data processing logic and can schedule series of jobs, sequentially or in parallel (SparkContext is thread-safe)  Job – Complete set of transformations on RDD that finishes with action or data saving, triggered by the driver application

62

Application Decomposition  Stage – Set of transformations that can be pipelined and executed by a single independent worker. Usually it is app the transformations between “read”, “shuffle”, “action”, “save”

63

Application Decomposition  Stage – Set of transformations that can be pipelined and executed by a single independent worker. Usually it is app the transformations between “read”, “shuffle”, “action”, “save”  Task – Execution of the stage on a single data partition. Basic unit of scheduling

64

WordCount ExampleHDFSInputSplits HDFS RDDPartitions RDD RDD RDD RDD RDD sc.textFile(‘hdfs://…’) flatMap map reduceByKey foreach

65

Stage 1 WordCount ExampleHDFSInputSplits HDFS RDDPartitions RDD RDD RDD RDD RDD sc.textFile(‘hdfs://…’) flatMap map reduceByKey foreach

66

Stage 2Stage 1 WordCount ExampleHDFSInputSplits HDFS RDDPartitions RDD RDD RDD RDD RDD sc.textFile(‘hdfs://…’) flatMap map reduceByKey foreach

67

Stage 2Stage 1 WordCount ExampleHDFSInputSplits HDFS RDDPartitions RDD RDD RDD RDD RDD sc.textFile(‘hdfs://…’) flatMap map reduceByKey foreach Task 1 Task 2 Task 3 Task 4

68

Stage 2Stage 1 WordCount ExampleHDFSInputSplits HDFS RDDPartitions RDD RDD RDD RDD RDD sc.textFile(‘hdfs://…’) flatMap map reduceByKey foreach Task 1 Task 2 Task 3 Task 4 Task 5 Task 6 Task 7 Task 8

69

Stage 2Stage 1 WordCount ExampleHDFSInputSplits HDFS pipeline Task 1 Task 2 Task 3 Task 4 Task 5 Task 6 Task 7 Task 8 partition shuffle pipeline

70

Persistence in Spark Persistence Level Description MEMORY_ONLY Store RDD as deserialized Java objects in the JVM. If the RDD does not fit in memory, some partitions will not be cached and will be recomputed on the fly each time they're needed. This is the default level. MEMORY_AND_DISK Store RDD as deserialized Java objects in the JVM. If the RDD does not fit in memory, store the partitions that don't fit on disk, and read them from there when they're needed. MEMORY_ONLY_SER Store RDD as serialized Java objects (one byte array per partition). This is generally more space-efficient than deserialized objects, especially when using a fast serializer, but more CPU-intensive to read. MEMORY_AND_DISK_SER Similar to MEMORY_ONLY_SER, but spill partitions that don't fit in memory to disk instead of recomputing them on the fly each time they're needed. DISK_ONLY Store the RDD partitions only on disk. MEMORY_ONLY_2, DISK_ONLY_2, etc. Same as the levels above, but replicate each partition on two cluster nodes.

71

Persistence in Spark  Spark considers memory as a cache with LRU eviction rules  If “Disk” is involved, data is evicted to disks rdd = sc.parallelize(xrange(1000)) rdd.cache().count() rdd.persist(StorageLevel.MEMORY_AND_DISK_SER).count() rdd.unpersist()

72

Outline  Spark Motivation  Spark Pillars  Spark Architecture  Spark Shuffle  Spark DataFrame

73

Shuffles in Spark  Hash Shuffle – default prior to 1.2.0

74

Shuffles in Spark  Hash Shuffle – default prior to 1.2.0  Sort Shuffle – default now

75

Shuffles in Spark  Hash Shuffle – default prior to 1.2.0  Sort Shuffle – default now  Tungsten Sort – new optimized one!

76

Hash Shuffle

77

Executor JVM Partition Partition Partition Partition Partition Partition Partition Partition Hash Shuffle

78

Executor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task spark.executor.cores/ spark.task.cpus … Hash Shuffle

79

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task spark.executor.cores/ spark.task.cpus … Output File Output File Output File … Numberof “Reducers” spark.local.dir Hash Shuffle

80

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task spark.executor.cores/ spark.task.cpus … Output File Output File Output File … Numberof “Reducers” Output File Output File Output File … Numberof “Reducers” … spark.local.dir Hash Shuffle Numberof“map”tasks executedbythisexecutor

81

Hash Shuffle with Consolidation

82

Executor JVM Partition Partition Partition Partition Partition Partition Partition Partition Hash Shuffle With Consolidation

83

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task … Numberof “Reducers” Output File Output File Output File spark.local.dir Hash Shuffle With Consolidationspark.executor.cores/ spark.task.cpus

84

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … … … Numberof “Reducers” spark.local.dir Hash Shuffle With Consolidationspark.executor.cores/ spark.task.cpus spark.executor.cores/ spark.task.cpus … Numberof “Reducers” Output File Output File Output File Output File Output File Output File

85

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task … … … Numberof “Reducers” spark.local.dir Hash Shuffle With Consolidationspark.executor.cores/ spark.task.cpus spark.executor.cores/ spark.task.cpus … Numberof “Reducers” Output File Output File Output File Output File Output File Output File

86

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … … … Numberof “Reducers” spark.local.dir Hash Shuffle With Consolidationspark.executor.cores/ spark.task.cpus spark.executor.cores/ spark.task.cpus Output File Output File Output File … Numberof “Reducers” Output File Output File Output File

87

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task … … … Numberof “Reducers” spark.local.dir Hash Shuffle With Consolidationspark.executor.cores/ spark.task.cpus spark.executor.cores/ spark.task.cpus Output File Output File Output File … Numberof “Reducers” Output File Output File Output File

88

Local DirectoryExecutor JVM Partition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … … … Numberof “Reducers” spark.local.dir Hash Shuffle With Consolidationspark.executor.cores/ spark.task.cpus spark.executor.cores/ spark.task.cpus … Numberof “Reducers” Output File Output File Output File Output File Output File Output File

89

Sort Shuffle

90

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition Sort Shuffle spark.storage.[safetyFraction * memoryFraction] Partition Partition

91

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Sort Shuffle spark.storage.[safetyFraction * memoryFraction] Partition Partition spark.executor.cores/spark.task.cpus

92

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition AppendOnlyMap … AppendOnlyMap spark.executor.cores/spark.task.cpus

93

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition AppendOnlyMap … AppendOnlyMap sort & spill spark.executor.cores/spark.task.cpus

94

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition AppendOnlyMap … AppendOnlyMap sort & spill Local Directory Output File index spark.executor.cores/spark.task.cpus

95

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition AppendOnlyMap … AppendOnlyMap sort & spill sort & spill Local Directory Output File index Output File index spark.executor.cores/spark.task.cpus

96

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition AppendOnlyMap … AppendOnlyMap sort & spill sort & spill sort & spill … Local Directory … Output File index Output File index Output File index spark.executor.cores/spark.task.cpus

97

MinHeap Merge MinHeap Merge Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition AppendOnlyMap … AppendOnlyMap sort & spill sort & spill sort & spill … Local Directory … “reduce”task“reduce”task …Output File index Output File index Output File index spark.executor.cores/spark.task.cpus

98

Tungsten Sort Shuffle

99

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] Partition Partition

100

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] Partition Partition spark.executor.cores/spark.task.cpus

101

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> Array of data pointers and Partition IDs, long[] spark.executor.cores/spark.task.cpus

102

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> Array of data pointers and Partition IDs, long[] sort & spill spark.executor.cores/spark.task.cpus

103

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> Local Directory Array of data pointers and Partition IDs, long[] sort & spill spark.local.dir Output File partition partition partition partition index spark.executor.cores/spark.task.cpus

104

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> Local Directory Array of data pointers and Partition IDs, long[] sort & spill sort & spill … spark.local.dir Output File partition partition partition partition index Output File partition partition partition partition index spark.executor.cores/spark.task.cpus

105

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> Local Directory Array of data pointers and Partition IDs, long[] sort & spill sort & spill … spark.local.dir Output File partition partition partition partition index Output File partition partition partition partition index spark.executor.cores/spark.task.cpus

106

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> … Local Directory Array of data pointers and Partition IDs, long[] Serialized Data LinkedList<MemoryBlock> Array of data pointers and Partition IDs, long[] sort & spill sort & spill … spark.local.dir Output File partition partition partition partition index Output File partition partition partition partition index spark.executor.cores/spark.task.cpus

107

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> … Local Directory … Array of data pointers and Partition IDs, long[] Serialized Data LinkedList<MemoryBlock> Array of data pointers and Partition IDs, long[] sort & spill sort & spill … sort & spill spark.local.dir Output File partition partition partition partition index Output File partition partition partition partition index Output File partition partition partition partition index spark.executor.cores/spark.task.cpus

108

Executor JVMPartition Partition Partition Partition Partition Partition Partition Partition “map” task “map” task … Tungsten Sort Shuffle spark.storage.[safetyFraction * memoryFraction] spark.shuffle. [safetyFraction * memoryFraction] Partition Partition Serialized Data LinkedList<MemoryBlock> … Local Directory … Array of data pointers and Partition IDs, long[] Serialized Data LinkedList<MemoryBlock> Array of data pointers and Partition IDs, long[] sort & spill sort & spill … sort & spill spark.local.dir Output File partition partition partition partition index Output File partition partition partition partition index Output File partition partition partition partition index Output File partition partition partition partition index merge spark.executor.cores/spark.task.cpus

109

Outline  Spark Motivation  Spark Pillars  Spark Architecture  Spark Shuffle  Spark DataFrame

110

DataFrame Idea

111

DataFrame Implementation  Interface – DataFrame is an RDD with schema – field names, field data types and statistics – Unified transformation interface in all languages, all the transformations are passed to JVM – Can be accessed as RDD, in this case transformed to the RDD of Row objects

112

DataFrame Implementation  Internals – Internally it is the same RDD – Data is stored in row-columnar format, row chunk size is set by spark.sql.inMemoryColumnarStorage.batchSize – Each column in each partition stores min-max values for partition pruning – Allows better compression ratio than standard RDD – Delivers faster performance for small subsets of columns

113

113Pivotal Confidential–Internal Use Only 113Pivotal Confidential–Internal Use Only Questions?

114

BUILT FOR THE SPEED OF BUSINESS

More Related Content

Apache Spark Architecture

Editor's Notes

  1. def printfunc (x): print 'Word "%s" occurs %d times' % (x[0], x[1]) infile = sc.textFile('hdfs://sparkdemo:8020/sparkdemo/textfiles/README.md', 4) rdd1 = infile.flatMap(lambda x: x.split()) rdd2 = rdd1.map(lambda x: (x, 1)).reduceByKey(lambda x, y: x+y) print rdd2.toDebugString() rdd2.foreach(printfunc)