Apache: Big Data North America 2017 参加報告 #streamctjp

>100 Views

July 28, 17

streamctjp

スライド概要

https://connpass.com/event/61546/

Yahoo!デベロッパーネットワーク

@ydnjp

スライド一覧

2023年10月からSpeaker Deckに移行しました。最新情報はこちらをご覧ください。 https://speakerdeck.com/lycorptech_jp

またはPlayer版

埋め込む »CMSなどでJSが使えない場合

関連スライド

深層学習による自然言語処理入門: word2vecからBERT, GPT-3まで

Yahoo!デベロッパーネットワーク 135.6K

ゼロから始める転移学習

Yahoo!デベロッパーネットワーク 75.7K

ヤフーにおける WebAuthn と Passkey の UX の紹介と考察 #idcon #fidcon

idcon fidcon

Yahoo!デベロッパーネットワーク 65.2K

OpenID Connectとネイティブアプリを取り巻く仕様と動向 Yahoo! JAPANの取り組み #openid #openid_tokyo

openid openid_tokyo

Yahoo!デベロッパーネットワーク 52.6K

運用業務とスクラムは本当に組み合わせにくいのか︖運用業務が大半を占めるプロダクト開発での試行錯誤

devsumi

Yahoo!デベロッパーネットワーク 28.4K

ヤフーのオンプレ機械学習基盤AIPFについて #ml_kubernetes

ml_kubernetes

Yahoo!デベロッパーネットワーク 26.1K

各ページのテキスト

Apache Big Data 2017 Miami • • • • • 5 Apache: Big Data North America 2017 May 16 – 18@Miami, Florida Apache Projects Developers, operators and users working in Big Data http://events.linuxfoundation.org/events/a pache-big-data-north-america Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

http://events.linuxfoundation.org/events/a

セッション数 session type Use Cases Ops Streaming SQL Hadoop Beam/Zeppelin Big Data Cassandra Deep Learning/GPU Machine Learning/Natural Language Processing Spark 8 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved. 件数 16 11 9 9 7 6 6 5 5 4 4

ストリーム処理エンジンへの言及各ストリームエンジンが何件の概要で言及されているか (/82件) Engine count Apex 6 Flink 6 Samza 2 Spark Streaming 2 Gearpump 1 Google Cloud Dataflow 1 Storm 0 9 参考 Spark Kafka Beam 14件 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved. count 28 12 4

10.

https://www.slideshare.net/techblogyahoo/hadoop-summit-2016-san-jose-streamctjp

11.

各ストリーム処理エンジンの動向 Apache BigDataに参加して感じたこと → 各ストリーム処理エンジンが基本機能を抑えてきてる • Beam Model対応 • unified API • event time • late data • exactly once • state management • High level API 11 • SQL Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

12.

各ストリーム処理が基本機能を抑えてきてる Spark 2.x Samza Apex (Flink) (Storm) unified API ◯ ◯ ◯ ◯ stream only event time ◯ coming soon ◯ ◯ ◯ late data ◯ coming soon ◯ ◯ ― state management ◯ ◯ ◯ ◯ ◯ SQL ◯ coming soon ◯ ◯ ◯ BEAM Runner 開発中 coming soon ◯ ◯ ― 基本機能を抑えてる → 各ストリーム処理エンジンの特徴は？ 12 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

13.

引用: Advantage of Dstreams (Spark 2.0) • Processing with event-time, dealing with late data • Exactly same API for batch, streaming, and interactive • End-to-end exactly-once guarantees for the system • Performance through SQL optimizations • Logical plan optimizations, Tungsten, Codegen, etc. • Faster state management for stateful stream management “Continuous Applications with Apache Spark 2.0” 13 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

14.

引用: Key differentiators for Apache Samza • Stream processing both as a multi-tenant service and a light-weight embedded library • No micro batching(first class streaming) • Unified processing of batch and streaming data • Efficient remote calls I/O using build-in async mode • World-class support for scalable local state • Incremental check-pointing • Instant recover with zero down time • Low level API and Stream based high Level API(DSL) “What It Takes to Process a Trillion Events a Day: Case-Studies in Scaling Stream Processing at LinkedIn” 14 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

15.

引用: Why Apex • State management • Processing based on event time • Native streaming • Scalability • Library of connectors and transformations • Exactly-once. Fault tolerance, windowing • Fine grained recovery, low-latency SLA support • Queryable state • Correctness, repeatable/replay • Low latency + high throughput, efficient resource utilization • Pipelined processing (data in motion) • Process more data by adding compute resources, no platform/architecture limit. • Dynamic scaling and resource allocation, elasticity • Time to value "From Batch to Streaming ET(L) with Apache Apex" 15 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

16.

17.

Apache Beam • 発表日(5/17)にfirst stable release • 参加セッション • “Using Apache Beam for Batch, Streaming, and Everything in Between” • Daniel Halperin: Apache Beam PMC, Google • “Apache Beam: Integrating the Big Data Ecosystem Up, Down, and Sideways” • Davor Bonaci: Apache Beam PMC chair, Google • Jean-Baptiste Onofré: Apache Beam PMC, Talend 17 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

18.

http://events.linuxfoundation.org/events/apache-big-data-north-america

19.

20.

http://events.linuxfoundation.org/sites/events/files/slides/Beam 2017 ApacheCon.pdf

21.

22.

23.

24.

25.

26.

Ecosystem SDK Runner DSLs Libraries transformation IO Connector File System 26 Javaに加えてPythonもサポート Apache Gearpumpが追加された今後: Spark 2, JStorm 今後: Scala(Scio), Calcite 例としてTensorFlow Kafka, elasticsearch, HBase,… など約20 今後: Cassandra, Redis HDFS, Amazon S3など Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

27.

28.

29.

Stateful Streaming Data Pipelines with Apache Apex • Chandni Singh • Apache Apex PMC and Committer, Simplifi.it • Timothy Farkas • Apache Apex Committer, Simplifi.it • Apache Apexの紹介 • Stateful managementの問題点と対応 29 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

30.

Apache Apex概要 • • • • 30 Distributed data processing engine Runs on Hadoop Real-time streaming Fault-tolerant • N window毎にチェックポイント • Committed window: 全てのオペレータがcheckpoint されときの最新のwindowをcommitted windowと呼ぶ Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

31.

Stateful managementの問題点 • Checkpointの時間はoperatorのstate sizeに相関 • Stateの増加に伴い、operatorはクラッシュしてしまう • それ以前に応答の遅さからYARNにkillされるだろう • → Managed State • key/valueの状態を差分更新 • 閾値に達した段階でメモリからWALに出力 • Keyはユーザ定義バケットで分割 • 最終的にHDFSに保持 31 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

32.

33.

よもやま話 • セッションでは性能改善について触れてない • リリース時のブログには以下のように記載 • we improved greatly how objects are serialized and reduced garbage collection considerably in the Managed State layer. 33 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

34.

35.

まとめ • ストリーム処理 • Stormの話がなくて心配 • 差別化要因がどこになるか • Apache Beamの期待感 • Unified API • Dynamic Work Rebalancing • I/O ConnectorやLibrariesの拡充が気になる • Stateful management • ステートが巨大化した時にどう対処するか 35 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

36.

reference • Using Apache Beam for Batch, Streaming, and Everything in Between • http://events.linuxfoundation.org/sites/events/files/slides/ 2017-0517%20dhalperi%20%40ApacheCon%20Big%20Data.pdf • Apache Beam: Integrating the Big Data Ecosystem Up, Down, and Sideways • http://events.linuxfoundation.org/sites/events/files/slides/ Beam%202017%20ApacheCon.pdf • Stateful Streaming Data Pipelines with Apache Apex • http://events.linuxfoundation.org/sites/events/files/slides/ Stateful%20streaming%20data%20pipelines.pdf#search= %27Spillable+Data+Structures%27 36 Copyright © 2017 Yahoo Japan Corporation. All Rights Reserved.

http://events.linuxfoundation.org/sites/events/files/slides/