Download as PDF, PPTX
![DataSet<String>
§ Think of it as a PCollection<String>, or a
Spark RDD[String]
§ With a major difference: it can be produced/
recovered in several ways
• … like a Java collection
• … like an RDD
• … perhaps it is never fully materialized (because
the program does not need it to)
• … implicitly updated in an iteration
§ And this is transparent to the user
15](https://image.slidesharecdn.com/flinkinternalsweb-141119104252-conversion-gate02/75/Apache-Flink-internals-15-2048.jpg)
![Two execution plans
34
GroupRed
sort
Combine
Map DataSource
Filter
DataSource
orders.tbl
lineitem.tbl
Join
Hybrid Hash
buildHT probe
broadcast forward
Map DataSource
Filter
DataSource
orders.tbl
lineitem.tbl
Join
Hybrid Hash
buildHT probe
hash-part [0] hash-part [0]
hash-part [0,1]
GroupRed
sort
Best plan forward
depends on
relative sizes
of input files](https://image.slidesharecdn.com/flinkinternalsweb-141119104252-conversion-gate02/75/Apache-Flink-internals-34-2048.jpg)
![Memory in Flink (2)
§ Internal memory management
• Flink initially allocates 70% of the free heap as byte[]
segments
• Internal operators allocate() and release() these
segments
§ Flink has its own serialization stack
• All accepted data types serialized to data segments
§ Easy to reason about memory, (almost) no
OutOfMemory errors, reduces the pressure to
the GC (smooth performance)
40](https://image.slidesharecdn.com/flinkinternalsweb-141119104252-conversion-gate02/75/Apache-Flink-internals-40-2048.jpg)
![DataSet<String>
§ Think of it as a PCollection<String>, or a
Spark RDD[String]
§ With a major difference: it can be produced/
recovered in several ways
• … like a Java collection
• … like an RDD
• … perhaps it is never fully materialized (because
the program does not need it to)
• … implicitly updated in an iteration
§ And this is transparent to the user
15](https://crownmelresort.com/image.slidesharecdn.com/flinkinternalsweb-141119104252-conversion-gate02/75/Apache-Flink-internals-15-2048.jpg)
![Two execution plans
34
GroupRed
sort
Combine
Map DataSource
Filter
DataSource
orders.tbl
lineitem.tbl
Join
Hybrid Hash
buildHT probe
broadcast forward
Map DataSource
Filter
DataSource
orders.tbl
lineitem.tbl
Join
Hybrid Hash
buildHT probe
hash-part [0] hash-part [0]
hash-part [0,1]
GroupRed
sort
Best plan forward
depends on
relative sizes
of input files](https://crownmelresort.com/image.slidesharecdn.com/flinkinternalsweb-141119104252-conversion-gate02/75/Apache-Flink-internals-34-2048.jpg)
![Memory in Flink (2)
§ Internal memory management
• Flink initially allocates 70% of the free heap as byte[]
segments
• Internal operators allocate() and release() these
segments
§ Flink has its own serialization stack
• All accepted data types serialized to data segments
§ Easy to reason about memory, (almost) no
OutOfMemory errors, reduces the pressure to
the GC (smooth performance)
40](https://crownmelresort.com/image.slidesharecdn.com/flinkinternalsweb-141119104252-conversion-gate02/75/Apache-Flink-internals-40-2048.jpg)
Uploaded byKostas Tzoumas
Apache Flink internals
This document provides an overview of Apache Flink internals. It begins with an introduction and recap of Flink programming concepts. It then discusses how Flink programs are compiled into execution plans and executed in a pipelined fashion, as opposed to being executed eagerly like regular code. The document outlines Flink's architecture including the optimizer, runtime environment, and data storage integrations. It also covers iterative processing and how Flink handles iterations both by unrolling loops and with native iterative datasets.
More Related Content
Similar to Apache Flink internals
Apache Flink internals
- 1. Flink internals KostasTzoumas Flink committer & Co-founder, data Artisans ktzoumas@apache.org @kostas_tzoumas
- 2. Welcome § Lasttalk: how to program PageRank in Flink, and Flink programming model § This talk: how Flink works internally § Again, a big bravo to the Flink community 2
- 3.
- 4. DataSet and transformations Input X First Y Second Operator X Operator Y ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment(); DataSet<String> input = env.readTextFile(input); DataSet<String> first = input .filter (str -‐> str.contains(“Apache Flink“)); DataSet<String> second = first .filter (str -‐> str.length() > 40); second.print() env.execute(); 4
- 5. Available transformations §map § flatMap § filter § reduce § reduceGroup § join § coGroup § aggregate § cross § project § distinct § union § iterate § iterateDelta § repartition § … 5
- 6. Other API elements& tools § Accumulators and counters • Int, Long, Double counters • Histogram accumulator • Define your own § Broadcast variables § Plan visualization § Local debugging/testing mode 6
- 7. Data types andgrouping public static class Access { public int userId; public String url; ... } public static class User { public int userId; public int region; public Date customerSince; ... } DataSet<Tuple2<Access,User>> campaign = access.join(users) .where(“userId“).equalTo(“userId“) DataSet<Tuple3<Integer,String,String> someLog; someLog.groupBy(0,1).reduceGroup(...); § Bean-style Java classes & field names § Tuples and position addressing § Any data type with key selector function 7
- 8. Other API elements § Hadoop compatibility • Supports all Hadoop data types, input/output formats, Hadoop mappers and reducers § Data streaming API • DataStream instead of DataSet • Similar set of operators • Currently in alpha but moving very fast § Scala and Java APIs (mirrored) § Graph API (Spargel) 8
- 9.
- 10. for (String token : value.split("W")) { out.collect(new Tuple2<>(token, 1)); Task Manager DataSet<String> text = env.readTextFile(input); DataSet<Tuple2<String, Integer>> result = text Job Manager Task Manager .flatMap((str, out) -‐> { }) .groupBy(0) .aggregate(SUM, 1); Flink Client & Optimizer O Romeo, Romeo, wherefore art thou Romeo? O, 1 Romeo, 3 wherefore, 1 art, 1 thou, 1 Apache Flink 10 Nor arm, nor face, nor any other part nor, 3 arm, 1 face, 1, any, 1, other, 1 part, 1
- 11. If you wantto know one thing about Flink is that you don’t need to know the internals of Flink. 11
- 12. Philosophy § Flink“hides” its internal workings from the user § This is good • User does not worry about how jobs are executed • Internals can be changed without breaking changes § … and bad • Execution model more complicated to explain compared to MapReduce or Spark RDD 12
- 13. Recap: DataSet InputX First Y Second Operator X Operator Y 13 ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment(); DataSet<String> input = env.readTextFile(input); DataSet<String> first = input .filter (str -‐> str.contains(“Apache Flink“)); DataSet<String> second = first .filter (str -‐> str.length() > 40); second.print() env.execute();
- 14. Common misconception InputX First Y Second § Programs are not executed eagerly § Instead, system compiles program to an execution plan and executes that plan 14
- 15. DataSet<String> § Thinkof it as a PCollection<String>, or a Spark RDD[String] § With a major difference: it can be produced/ recovered in several ways • … like a Java collection • … like an RDD • … perhaps it is never fully materialized (because the program does not need it to) • … implicitly updated in an iteration § And this is transparent to the user 15
- 16. Example: grep Romeo, Romeo, where art thou Romeo? Load Log Search for str1 Search for str2 Search for str3 Grep 1 Grep 2 Grep 3 16
- 17. Staged (batch) execution Romeo, Romeo, where art thou Romeo? Load Log Load Log Search for str1 Search for str2 Search for str3 Grep 1 Grep 2 Grep 3 Stage 1: Create/cache Log Subseqent stages: Grep log for matches Caching in-memory and disk if needed Search for str1 Search for str2 Search for str2 Grep 1 Grep 2 Grep 2 Load Log Search for str1 Search for str2 Search for str2 Grep 1 Grep 2 Grep 2 17
- 18. Load Log Search for str1 Search for str2 Search for str2 Grep 1 Grep 2 Grep 2 Pipelined execution Romeo, Romeo, where art thou Romeo? Load Log Load Log Search for str1 Search for str2 Search for str3 Grep 1 Grep 2 Grep 3 000000111111000000111111 Stage 1: Deploy and start operators Data transfer in-memory and disk if needed Search for str1 Search for str2 Search for str2 Grep 1 Grep 2 Grep 2 18 Note: Log DataSet is never “created”!
- 19. Benefits of pipelining § 25 node cluster § Grep log for 3 terms § Scale data size from 100GB to 1TB 2500 Time to complete grep (sec) Data size (GB) 2250 2000 1750 1500 1250 1000 750 500 250 0 Pipelined with Flink 0 100 200 300 400 500 600 700 800 900 1000 Cluster memory exceeded 19
- 20.
- 21. Drawbacks of pipelining § Long pipelines may be active at the same time leading to memory fragmentation • FLINK-1101: Changes memory allocation from static to adaptive § Fault-tolerance harder to get right • FLINK-986: Adds intermediate data sets (similar to RDDS) as first-class citizen to Flink Runtime. Will lead to fine-grained fault-tolerance among other features. 21
- 22. Example: Iterative processing DataSet<Page> pages = ... DataSet<Neighborhood> edges = ... DataSet<Page> oldRanks = pages; DataSet<Page> newRanks; for (i = 0; i < maxIterations; i++) { newRanks = update(oldRanks, edges) oldRanks = newRanks } DataSet<Page> result = newRanks; DataSet<Page> update (DataSet<Page> ranks, DataSet<Neighborhood> adjacency) { return oldRanks .join(adjacency) .where(“id“).equalTo(“id“) .with ( (page, adj, out) -‐> { for (long n : adj.neighbors) out.collect(new Page(n, df * page.rank / adj.neighbors.length)) }) .groupBy(“id“) .reduce ( (a, b) -‐> new Page(a.id, a.rank + b.rank) ); 22
- 23. Iterate by unrolling Client Step Step Step Step Step § for/while loop in client submits one job per iteration step § Data reuse by caching in memory and/or disk 23
- 24. Iterate natively Yinitial solution DataSet<Page> pages = ... DataSet<Neighborhood> edges = ... IterativeDataSet<Page> pagesIter = pages.iterate(maxIterations); DataSet<Page> newRanks = update (pagesIter, edges); DataSet<Page> result = pagesIter.closeWith(newRanks) 24 partial solution partial X solution other datasets iteration result Replace Step function
- 25. Iterate natively withdeltas Replace workset A B workset initial workset initial partial solution solution Y delta X set other datasets Merge deltas DeltaIteration<...> pagesIter = pages.iterateDelta(initialDeltas, iteration result maxIterations, 0); DataSet<...> newRanks = update (pagesIter, edges); DataSet<...> newRanks = ... DataSet<...> result = pagesIter.closeWith(newRanks, deltas) See http://data-artisans.com/data-analysis-with-flink.html 25
- 26. Native, unrolling, anddelta 26
- 27.
- 28.
- 29. The growing Flinkstack 29 Python API (upcoming) Graph API Apache Common API Flink Optimizer Flink Stream Builder Scala API (batch) Java API (streaming) Java API (batch) MRQL Flink Local Runtime Embedded environment (Java collections) Local Environment (for debugging) Remote environment (Regular cluster execution) Apache Tez Single node execution Standalone or YARN cluster Data storage Files HDFS S3 JDBC Redis Rabbit Kafka MQ Azure tables …
- 30. Stack without FlinkStreaming 30 30 Python API (upcoming) Graph API Apache Focus on regular (batch) processing… Scala API Java API Common API Flink Optimizer MRQL Embedded Flink Local Runtime environment (Java collections) Local Environment (for debugging) Remote environment (Regular cluster execution) Apache Tez Standalone or YARN cluster Data storage Files HDFS S3 JDBC Azure tables … Single node execution
- 31. Program lifecycle 30 30 Python API (upcoming) Graph API Apache Scala API Java API Common API Flink Optimizer MRQL Embedded Flink Local Runtime environment (Java collections) Local Environment (for debugging) Remote environment (Regular cluster execution) Apache Tez Standalone or YARN cluster Data storage Files HDFS S3 JDBC Azure tables … Single node execution 31 val source1 = … val source2 = … maxed = source1 .map(v => (v._1,v._2, val math.max(v._1,v._2)) val filtered = source2 .filter(v => (v._1 > 4)) val result = maxed .join(filtered).where(0).equalTo(0) .filter(_1 > 3) .groupBy(0) .reduceGroup {……} 1 3 4 5 2
- 32. 30 30 PythonAPI (upcoming) Graph API Apache Scala API Java API Common API Flink Optimizer MRQL Embedded Flink Local Runtime environment (Java collections) Local Environment (for debugging) Remote environment (Regular cluster execution) Apache Tez Standalone or YARN cluster Data storage Files HDFS S3 JDBC Azure tables … Single node execution § The optimizer is the component that selects an execution plan for a Common API program § Think of an AI system manipulating your program for you J § But don’t be scared – it works • Relational databases have been doing this for decades – Flink ports the technology to API-based systems Flink Optimizer 32
- 33. A simple program 33 DataSet<Tuple5<Integer, String, String, String, Integer>> orders = … DataSet<Tuple2<Integer, Double>> lineitems = … DataSet<Tuple2<Integer, Integer>> filteredOrders = orders .filter(. . .) .project(0,4).types(Integer.class, Integer.class); DataSet<Tuple3<Integer, Integer, Double>> lineitemsOfOrders = filteredOrders .join(lineitems) .where(0).equalTo(0) .projectFirst(0,1).projectSecond(1) .types(Integer.class, Integer.class, Double.class); DataSet<Tuple3<Integer, Integer, Double>> priceSums = lineitemsOfOrders .groupBy(0,1).aggregate(Aggregations.SUM, 2); priceSums.writeAsCsv(outputPath);
- 34. Two execution plans 34 GroupRed sort Combine Map DataSource Filter DataSource orders.tbl lineitem.tbl Join Hybrid Hash buildHT probe broadcast forward Map DataSource Filter DataSource orders.tbl lineitem.tbl Join Hybrid Hash buildHT probe hash-part [0] hash-part [0] hash-part [0,1] GroupRed sort Best plan forward depends on relative sizes of input files
- 35. Flink Local Runtime 30 30 Python API (upcoming) Graph API Apache Scala API Java API Common API Flink Optimizer MRQL Embedded Flink Local Runtime environment (Java collections) Local Environment (for debugging) Remote environment (Regular cluster execution) Apache Tez Standalone or YARN cluster Data storage Files HDFS S3 JDBC Azure tables … Single node execution § Local runtime, not the distributed execution engine § Aka: what happens inside every parallel task 35
- 36. Flink runtime operators § Sorting and hashing data • Necessary for grouping, aggregation, reduce, join, cogroup, delta iterations § Flink contains tailored implementations of hybrid hashing and external sorting in Java • Scale well with both abundant and restricted memory sizes 36
- 37. Internal data representation 37 JVM Heap map JVM Heap reduce O Romeo, Romeo, wherefore art thou Romeo? 00110011 art, 1 O, 1 Romeo, 1 Romeo, 1 00110011 00010111 01110001 01111010 00010111 00110011 Network transfer Local sort How is intermediate data internally represented?
- 38. Internal data representation § Two options: Java objects or raw bytes § Java objects • Easier to program • Can suffer from GC overhead • Hard to de-stage data to disk, may suffer from “out of memory exceptions” § Raw bytes • Harder to program (customer serialization stack, more involved runtime operators) • Solves most of memory and GC problems • Overhead from object (de)serialization § Flink follows the raw byte approach 38
- 39. Memory in Flink public class WC { public String word; public int count; } empty page Pool of Memory Pages JVM Heap User code objects Sorting, hashing, caching Shuffling, broadcasts Unmanaged heap Managed heap Network buffers 39
- 40. Memory in Flink(2) § Internal memory management • Flink initially allocates 70% of the free heap as byte[] segments • Internal operators allocate() and release() these segments § Flink has its own serialization stack • All accepted data types serialized to data segments § Easy to reason about memory, (almost) no OutOfMemory errors, reduces the pressure to the GC (smooth performance) 40
- 41. Operating on serializeddata Microbenchmark § Sorting 1GB worth of (long, double) tuples § 67,108,864 elements § Simple quicksort 41
- 42. Flink distributed execution 30 30 Python API (upcoming) Graph API Apache Scala API Java API Common API Flink Optimizer MRQL Embedded Flink Local Runtime environment (Java collections) Local Environment (for debugging) Remote environment (Regular cluster execution) Apache Tez Standalone or YARN cluster Data storage Files HDFS S3 JDBC Azure tables … Single node execution 42 § Pipelined • Same engine for Flink and Flink streaming § Pluggable • Local runtime can be executed on other engines • E.g., Java collections and Apache Tez
- 43.
- 44. Summary § Flinkdecouples API from execution • Same program can be executed in many different ways • Hopefully users do not need to care about this and still get very good performance § Unique Flink internal features • Pipelined execution, native iterations, optimizer, serialized data manipulation, good disk destaging § Very good performance • Known issues currently worked on actively 44
- 45. Stay informed §flink.incubator.apache.org • Subscribe to the mailing lists! • http://flink.incubator.apache.org/community.html#mailing-lists § Blogs • flink.incubator.apache.org/blog • data-artisans.com/blog § Twitter • follow @ApacheFlink 45
- 46.
- 47. That’s it, timefor beer 47