cloud based data stream processing -...

CLOUD-BASED �DATA STREAM PROCESSING

Thomas Heinze, Leonardo Aniello, �Leonardo Querzoni, Zbigniew Jerzak

�DEBS 2014

Mumbai 26/5/2014

TUTORIAL GOALS Data stream processing (DSP) was in the past considered a solution for very specific problems.

§  Financial trading §  Logistics tracking §  Factory monitoring

Today the potentialities of DSPs start to be used in more general settings.

§ DSP : online processing = MR : batch processing

DSPs will possibly be offered as services from cloud-based providers ?

2 25/05/14 CLoud-Based Data Stream Processing

TUTORIAL GOALS

Here we present an overview about current research trends within data streaming systems

§ How they consider the requirements imposed by recent use-cases

§ How are they moving toward cloud platforms

Two focus areas: §  Scalability §  Fault Tolerance

THE AUTHORS

Thomas Heinze !Phd Student @ SAP Phd Topic: Elastic Data Stream Processing

Leonardo Aniello !Researcher fellow @ Sapienza Univ. of Rome

Leonardo Querzoni!Assistant professor @ Sapienza Univ. of Rome

Zbigniew Jerzak !Senior Researcher @ SAP

Co-Organizer of DEBS Grand Challenge

25/05/14 CLoud-Based Data Stream Processing

OUTLINE §  Historical overview of data stream processing §  Use cases for cloud-based data stream

processing §  How DSPs work: Storm as example §  Scalability methods §  Fault tolerance integrated in modern DSPs §  Future research directions and conclusions

Historical Overview 6

DATA STREAM PROCESSING ENGINE §  It is a piece of software that

§  continuously calculates results for long-standing queries

§  over potentially infinite data streams §  using operators

§ algebraic (filters, join, aggregation)

§ user defined

§  That can be stateless or stateful

25/05/14 CLoud-Based Data Stream Processing 7

Source Aggr Sink

REQUIREMENTS[1]

Rule 1 - Keep the data moving Rule 2 - Query using SQL on stream (StreamSQL)

Rule 3 - Handle stream imperfections (delayed, missing and out-of-order data)

Rule 4 - Generate predictable outcomes

Rule 5 - Integrate stored and streaming data Rule 6 - Guarantee data safety and availability

Rule 7 - Partition and scale applications automatically

Rule 8 - Process and respond instantaneously

[1] M. Stonebraker, U. Cetintemel and S. Zdonik. "The 8 Requirements of real-time Stream Processing", ACM SIGMOD Record, 2005. 25/05/14 CLoud-Based Data Stream Processing

THREE GENERATIONS §  First Generation

§  Extensions to existing database engines or simplistic engines

§  Dedicated to specific use cases

§  Second Generation §  Enhanced methods regarding language expressiveness, load

balancing, fault tolerance

§  Third Generation §  Dedicated towards trend of cloud computing; designed

towards massive parallelization

GENEALOGY

10 FIRST GENERATION THIRD GENERATION SECOND GENERATION

Aurora (2002) Medusa (2004)

Borealis (2005)

TelegraphCQ (2003)

NiagaraCQ (2001)

STREAM (2003)

StreamMapReduce(2011)

StreamMine3G (2013)

Yahoo! S4(2010)

Apache S4(2012)

TwiOer Storm (2011) Trident (2013)

StreamCloud(2010)

ElasPc StreamCloud (2012)

SPADE(2008)

ElasPc Operator(2009)

InfoSphere Streams(2010)

FLUX (2002)

GigaScope (2002)

Spark Streaming (2010) D-‐Streams(2013)

Cayuga (2007)

SASE (2008)

ZStream (2009)

CAPE (2004)

D-‐CAPE (2005)

SGuard (2008)

PIPES (2004) HybMig (2006)

StreamInsight (2010)

TimeStream (2013) CEDR (2005)

SEEP (2013)

Esper (2006)

Truviso (2006)

StreamBase (2003)

Coral8 (2005)

Aleri (2006) Sybase CEP (2010)

Oracle CQL (2006) Oracle CEP (2009)

SAP ESP (2012)

DejaVu (2009)

System S(2005) LAAR(2014)

StreamIt (2002)

Flextream(2007)

ElasPc System S(2013)

Cisco Prime (2012)

RTM Analyzer (2008)

webMethods Business Events(2011)

TIBCO StreamBase (2013)

TIBCO BusinessEvents(2008)

NILE (2004)

NILE-‐PDT (2005)

RLD (2013)

MillWheel(2013)

Johka (2007)

FIRST GENERATION - TELEGRAPH CQ[2]

§  Data stream processing engine built on top of Postgres DB

[2] S. Chandrasekaran et al. "TelegraphCQ: Continuous Dataflow Processing", SIGMOD, 2003.

SECOND GENERATION - BOREALIS[3]

§  Joint research project of Brandeis University, Brown University and MIT

§  Duration: ~3 years §  Allowed the experimentation of several

techniques

[3] D. J. Abadi et al. "The Design of the Borealis Stream Processing Engine", CIDR 2005.

BOREALIS MAIN NOVELTIES

§  Load-aware Distribution

§  Fine-grained High-availability §  Load Shedding mechanisms

THIRD GENERATION

§  Novel use cases are driving toward §  Uprecedeted levels of parallelism §  Efficient fault tolerance § Dynamic scalability §  Etc.

Use Cases for Cloud-Based Data Stream Processing

SCENARIOS FOR THIRD GENERATION DSPS

§  Tracking of query trend evolution in Google §  Analysis of popular queries submitted to Twitter §  User profiling at Yahoo! based on submitted queries §  Bus routing monitoring and management in Dublin §  Sentiment analysis on multiple tweet streams §  Fraud monitoring in cellular telephony

16 25/05/14 Cloud-Based Data Stream Processing

QUERY MONITORING AT GOOGLE

§  Analyze queries submitted to Google search engine to create a query historical model

§  Run on Zeitgeist on top of MillWheel §  Incoming searches are organized in 1-second buckets §  Buckets are compared to historical data §  Useful to promptly detect anomalies (spikes/dips)

Akidau, Balikov, Bekiroğlu, Chernyak, Haberman, Lax, McVeety, Mills, Nordstrom, Whittle. MillWheel: Fault-tolerant Stream Processing at Internet Scale.VLDB, 2013. 25/05/14 Cloud-Based Data Stream Processing

POPULAR QUERY ANALYSIS AT TWITTER

§  When a relevant event happens, an increase occurs in the number of queries submitted to Twitter

§  These queries have to be correlated in real-time with tweets (2.1 billion tweets per day)

§  Runs on Storm §  These spikes are likely to fade away in a limited amount of

time §  Useful to improve the accuracy of popular queries

Improving Twitter Search with Real-time Human Computation, 2013 http://blog.echen.me/2013/01/08/improving-twitter-search-with-real-time-human-computation

25/05/14 Cloud-Based Data Stream Processing

POPULAR QUERY ANALYSIS AT TWITTER §  Problem

§  Sudden peak of queries about a new (never seen before) event

§  How to properly assign the right semantic (categorization) to the queries? §  Example

§  Solution §  Employ Human Evaluation (Amazon’s Mechanical Turk Service) to produce

categorizations to queries unseen so far

§  incorporate such categorizations into backend models

how can we understand that #bindersfullofwomen refers to politics?

POPULAR QUERY ANALYSIS AT TWITTER

Search Engine

Kaea Queue

query Log

(query, ts)

popular queries

Human EvaluaPon

categoriza5ons of queries

engine tuning

USER PROFILING AT YAHOO! §  Queries submitted by users are evaluated (thousands

queries per second by millions of users) §  Run on S4[1]

§  Useful to generate highly personalized advertising

[1] L. Neumeyer, B. Robbins, A. Nair, and A. Kesari. S4: Distributed Stream Computing Platform.ICDMW, 2010. 25/05/14 Cloud-Based Data Stream Processing

BUS ROUTING MANAGEMENT IN DUBLIN

§  Tracking of bus locations (1000 buses) to improve public transportation for 1.2 million citizens

§  Run on System S §  Position tracking by GPS signals to provide real-time

traffic information monitoring §  Useful to predict arrival times and suggest better routes

Dublin City Council - Traffic Flow Improved by Big Data Analytics Used to Predict Bus Arrival and Transit Times. IBM, 2013. http://www-03.ibm.com/software/businesscasestudies/en/us/corp?docid=RNAE-9C9PN5

SENTIMENT ANALYSIS ON TWEET STREAMS §  Tweet streams flowing at high rates (10K tweet/s) §  Sentiment analysis in real-time §  Limited computation time (latency up to 2 seconds) §  Run on Timestream[1]

§  Useful to continuously estimate the mood about specific topics

[1] Z. Qian, Y. He, C. Su, Z. Wu, H. Zhu, T. Zhang, L. Zhou, Y. Yu, Z. Zhang. Timestream: Reliable Stream Computation in the Cloud. Eurosys, 2013.

FRAUD MONITORING FOR MOBILE CALLS §  Fraud detection by real-time processing of Call Detail

Records (10k-50k CDR/s) §  Requires self-joins over large time windows (queries on

millions of CDRs) §  Run on StreamCloud[1]

§  Useful to reactively spot dishonest behaviors

[1] V. Gulisano, R. Jimenez-Peris, M. Patino-Martinez, C. Soriente, and P. Valduriez. Streamcloud: An Elastic and Scalable Data Streaming System. IEEE TPDS, 2012. 25/05/14 Cloud-Based Data Stream Processing

REQUIREMENTS §  Real-time and continuous complex analysis of heavy

data streams §  More than 10k event/s

§  Limited computation latency §  Up to few seconds

§  Correlation of new and historical data §  Computations have a state to be kept

§  Input load varies considerably over time §  Computations have to adapt dynamically (Scalability)

§  Hardware and Software failures can occur §  Computations have to transparently tolerate faults with limited

performance degradation (Fault Tolerance)

25 25/05/14 Cloud-Based Data Stream Processing

How DSPs work: Storm as example

STORM §  Storm is an open source distributed realtime

computation system §  Provides abstractions for implementing event-based

computations over a cluster of physical nodes § Manages high throughput data streams §  Performs parallel computations on them

§  It can be effectively used to design complex event-driven applications on intense streams of data

25/05/14 Cloud-Based Data Stream Processing 27

STORM §  Originally developed by Nathan Marz and team

at BackType, then acquired by Twitter, now an Apache Incubator project.

§  Currently used by Twitter, Groupon, The Weather Channel, Taobao, etc.

§  Design goals: § Guaranteed data processing § Horizontal scalability §  Fault Tolerance

STORM An application is represented by a topology:

Bolt Bolt

BoltBolt

OPERATOR EXPRESSIVENESS §  Storm is designed for custom operator definition

§  Bolts can be designed as POJOs adhering to a specific interface

§  Implementations in other languages are feasible

§  Trident[1] offers a more high level programming interface

[1] N. Marz. Trident tutorial. https://github.com/nathanmarz/storm/wiki/Trident-tutorial, 2013.

OPERATOR EXPRESSIVENESS §  Operators are connected through grouping:

§  Shuffle grouping §  Fields grouping §  All grouping § Global grouping § None grouping § Direct grouping §  Local or shuffle grouping

PARALLELIZATION §  Storm asks the developer to provide “parallelism

hints” in the topology

Bolt Bolt

BoltBolt

STORM INTERNALS §  A storm cluster is constituted by a Nimbus node

and n Worker nodes

Master node Worker node 1

Worker (process)Executor

Supervisor Supervisor

Worker node n

Storm cluster

Scheduler

Nimbus (process)25/05/14 Cloud-Based Data Stream Processing 33

TOPOLOGY EXECUTION §  A topology is run by submitting it to Nimbus

§  Nimbus allocates the execution of components (spouts and bolts) to the worker nodes using a scheduler §  Each component has multiple instances (parallelism) §  Each instance is mapped to an executor

§  A worker is instantiated whenever the hosting node must run executors for the submitted topology

§  Each worker node locally manages incoming/outgoing streams and local computation §  The local surpervisor takes care that everything runs as

prescribed §  Nimbus monitors worker nodes during the execution to

manage potential failures and the current resource usage

TOPOLOGY EXECUTION §  Storm’s default scheduler (EvenScheduler)

applies a simple round robin strategy

Executor A.1Executor A.2Executor A.3Executor A.4Executor B.1Executor B.2Executor C.1Executor D.1Executor D.2Executor D.3Executor D.4

Component A

Component B

Component C

Component D

Executors

Configuredworkers

WN4 WN5

Available worker nodes

A PRACTICAL EXAMPLE §  Word count: the HelloWorld for DSPs §  Input: stream of text (e.g. from documents) §  Output: number of appearance for each word

Source: hOp://wpcerPficaPon.blogspot.it/2014/02/helloworld-‐apache-‐storm-‐word-‐counter.html

HelloStorm

LineReaderSpout

WordSplitterBolt

WordCounterBolt

text lines words

TXT docs

A PRACTICAL EXAMPLE §  LineReaderSpout: reads docs and creates tuples

public class LineReaderSpout implements IRichSpout { public void open(Map conf, TopologyContext context, SpoutOutputCollector collector) { this.context = context; this.fileReader = new FileReader(conf.get("inputFile").toString()); this.collector = collector; } public void nextTuple() { if (completed) { Thread.sleep(1000); } String str; BufferedReader reader = new BufferedReader(fileReader); while ((str = reader.readLine()) != null) { this.collector.emit(new Values(str), str); completed = true; } public void declareOutputFields(OutputFieldsDeclarer declarer) { declarer.declare(new Fields("line")); } public void close() { fileReader.close(); } }

A PRACTICAL EXAMPLE §  WordSplitterBolt: cuts lines in words

public class WordSpitterBolt implements IRichBolt { public void prepare(Map stormConf, TopologyContext context,OutputCollector c) { this.collector = c; } public void execute(Tuple input) { String sentence = input.getString(0); String[] words = sentence.split(" "); for(String word: words){ word = word.trim(); if(!word.isEmpty()){ word = word.toLowerCase(); collector.emit(new Values(word)); } } collector.ack(input); } public void declareOutputFields(OutputFieldsDeclarer declarer) { declarer.declare(new Fields("word")); } } 25/05/14 Cloud-Based Data Stream Processing 38

A PRACTICAL EXAMPLE §  WordCounterBolt: counts word occurrences

public class WordCounterBolt implements IRichBolt { public void prepare(Map stormConf, TopologyContext context, OutputCollector c) { this.counters = new HashMap<String, Integer>(); this.collector = c; } public void execute(Tuple input) { String str = input.getString(0); if(!counters.containsKey(str)){ counters.put(str, 1); } else { Integer c = counters.get(str) +1; counters.put(str, c); } collector.ack(input); } public void cleanup() { for (Map.Entry<String, Integer> entry : counters.entrySet()){ System.out.println (entry.getKey()+" : " + entry.getValue()); } } } 25/05/14 Cloud-Based Data Stream Processing 39

A PRACTICAL EXAMPLE §  HelloStorm: contains the topology definition

public class HelloStorm { public static void main (String[] args) throws Exception{ Config config = new Config(); config.put("inputFile", args[0]); config.setDebug(true); config.put(Config.TOPOLOGY_MAX_SPOUT_PENDING, 1); TopologyBuilder builder = new TopologyBuilder(); builder.setSpout("line-reader-spout", new LineReaderSpout()); builder.setBolt("word-spitter", new WordSpitterBolt().shuffleGrouping( "line-reader-spout"); builder.setBolt("word-counter", new WordCounterBolt()).shuffleGrouping( "word-spitter"); LocalCluster cluster = new LocalCluster(); cluster.submitTopology("HelloStorm", config, builder.createTopology()); Thread.sleep(10000); cluster.shutdown(); } }

Scalability

PARTITIONING SCHEMES §  Data Parallelism:

§ How to parallelize the execution of an operator? § How to detect the optimal level of parallelization?

§  Operator Distribution: § How to distribute the load across available hosts? § How to achieve a load balance between these

machines?

25/05/14 Cloud-based Data Stream Processing 42

RQUIREMENTS FOR DATA PARALLELISM §  Transparent to the user

§  correct results in correct order (identical to sequential execution)

25/05/14 Cloud-based Data Stream Processing

Filter

Round-Robin Union

Hash (groupID)

Merge& Sort

DATA PARALLELISM §  First presented by FLUX[1] and Borealis[2]

§  Explicitely done using partitioning operators §  User needs to decide:

§ partitioning scheme

§ merging scheme

§  level of parallelism

[1] M. Shah, et al. "Flux: An adaptive partitioning operator for continuous query systems." In ICDE, 2003. [2] D. Abadi, et al. "The Design of the Borealis Stream Processing Engine." In CIDR, 2005..

PARALLELISM FOR CLOUD-BASED DSP §  Massive parallelization (>100 partitions) §  Support custom operators §  Adapt parallelization level to the workload

without user interaction

DATA PARALLELISM IN STORM §  User defines number of parallel task §  Storm support different partitioning schemes

(aka grouping): §  Shuffle grouping, Fields grouping, All grouping, Global

grouping, None grouping, Direct grouping, Local or shuffle grouping, Custom

MAPREDUCE FOR STREAMING [3,4]

§  Extend MapReduce model for streaming data: §  Break strict phases §  Introduce Stateful Reducer

Reduce

Output Input

[3] T. Condie, et al. "MapReduce Online." In NSDI,2010. [4] A. Brito, et al. "Scalable and low-latency data processing with stream mapreduce." CloudCom, 2011.

STATEFUL REDUCER[4]

reduceInit(k1) {

// custom user class

S = new State();S.sum = 0; S.count = 0;

// object S is now associated with key k1

return S;

reduce(Key k, <List new, List expired, List window, UserState S>) {

For v in expired do:

// Remove contribution of expired events

S.sum -= v; S.count--;

For v in new do:

// Add contribution of new events

S.sum += v; S.count++;

send(k1, S.sum/S.count);

[4] A. Brito, et al. "Scalable and low-latency data processing with StreamMapReduce." CloudCom, 2011.

AUTO PARALLELIZATION[5,6]

§  Goal: §  detect parallelizable regions in operator graph with

custom operators for user-defined operators §  Runtime support for enforcing safety conditions

§  Safety Conditions: §  For an operator: no state or partitioned state,

selectivity < 1, at most one pre/successor §  For an parallel region: compatible keys, forwarded

keys, region-local fusion dependencies

[5] S Schneider, et al. "Auto-parallelizing stateful distributed streaming applications." In PACT, 2012. [6] Wu et al. “Parallelizing Stateful Operators in a Distributed Stream Processing System: How, Should You and How Much?” In DEBS, 2012.

AUTO PARALLELIZATION §  Compiler-based Approach:

§  Characterize each operator based on a set of criterias (state type, selectivity, forwarding)

§ Merge different operators together into parallel regions (based on left-first approach)

§  Best parallelization strategy is applied automatically

§  Level of parallelism is decided on job admission

ELASTICITY[7] §  Goal: React to unpredicted load peaks & reduce

the amount of ideling resources.

Workload Load

Static Provisioning

Elastic Provisioning

Underprovisioning

Overprovisioning

[7] M. Armbrust, et al. "A view of cloud computing." Communications of the ACM, 2010.

DIFFERENT TYPES OF ELASTICITY §  Vertical Elasticity (Scale up)

§  Adapt the number of threads per operator based on the workload.

§  Horizontal Elasticity (Scale out) §  Adapt the number of hosts based on the workload.

ELASTIC OPERATOR EXECUTION[8]

§  Dynamically adapt number of threads for a stateless operator based on the workload

§  Limitations: works only on thread-level and for stateless operators

[8] S. Schneider, et al. "Elastic scaling of data parallel operators in stream processing." In IPDPS, 2009.

Operator

Thread Pool

ELASTIC AUTO-PARALLELIZATION[9]

§  Combines ideas of Elasticity and Auto-Parallelization

§  Adapt parallelization level using controller-based approach

§  Dynamic adaption of parallelization level based on operator migration protocol

[9] B. Gedik, et al. “Elastic Scaling for Data Stream Processing." In TPDS, 2013.

Horizontal barrier Vertical barrier

ELASTIC AUTO-PARALLELIZATION §  Dynamic adaption of parallelization: lend &

borrow approach

Splitter Merger

1.  Lend Phase 2. Borrow Phase

OPERATOR DISTRIBUTION §  Well-studied problem since first DSP prototypes §  Typically referred to as „Operator Placement“ §  Examples: Borealis, SODA, Pietzuch et al.

§  Design decisions[10] § Optimization goal: What is optimized ? §  Execution mode: Centralized or Decentralized? §  Algorithm runtime: Offline, online or both?

[10] G.T. Lakshmanan, et al. "Placement strategies for internet-scale data stream systems." In IEEE Internet Computing, 2008.

OPTIMIZATION GOAL §  Different objectives:

§  Balance CPU load (handle load variations) § Minimize Network latency (minimize latency for

sensor networks) §  Latency optimization (predictable quality of service)

EXECUTION MODE §  Centralized Execution:

§  All decision done by centralized management component

§ Major disadvantage: manager becomes scalability bottleneck

§  Decentralized Execution: § Different hosts try to agree on an operator

placement §  Two examples: operator routing and commutative

execution

DECENTRALIZER EXECUTION §  Based on pairwise agreement[11]

Host 1 Host 2 Host 3 Host 4

Aggr4 Aggr2

Filter 2 Filter 3

Filter 5

Filter 6

Filter 4

Filter 1

Aggr 3 Aggr 4 Aggr 3

Aggr 5 Aggr 6

Source 1

Source 2

[11] M. Balazinska, et al. "Contract-Based Load Management in Federated Distributed Systems." NSDI. Vol. 4. 2004.

DECENTRALIZER EXECUTION §  Based on decentralized routing[12]

Host 2

Host 3

Host 4 Host 1

[12] Y.Ahmad, et al. "Network-aware query processing for stream-based applications. In VLDB, 2004.

ALGORITHM RUNTIME §  Offline

§  Based on estimation of input rates, selectivities, etc.

§  Online § Mostly simplistic initial estimation (e.g. round-robin or

random placement) §  Continously measurements and adaption during

runtime

MOVEMENT PROTOCOLS §  How to ensure loss-free movement of

operators? § No input event shall be lost §  State need to be restored on new host

§  Movement strategies: §  Pause & Resume vs. Parallel track §  large overlap with research on adaptive query

processing[13]

[13] Y. Zhu, et al. "Dynamic plan migration for continuous queries over data streams." In SIGMOD, 2004.

PAUSE & RESUME §  Approach:

1.  Stop execution 2.  Move state to new host 3.  Restart execution on new hosts

§  Properties: §  Simple & generic §  Latency Peak can be observed due to pausing

PARALLEL TRACK §  Approach:

1.  Start new instance 2.  Move or create up to date state 3.  Stop old instance as soon as instances are in sync

§  Properties: § No latency peak §  Requires duplicate detection, detection of „sync“

status §  Events processed twice

OPERATOR PLACEMENT ITHIN CLOUD-BASED DSP

§  Different setup (highly location-wise distributed vs. single cluster)

§  Larger scale (100 …. 1000 hosts) §  New objectives: Elasticity, Monetary Cost, Energy

Efficiency

Mostly centralized, adaptive solutions optimizing CPU utilization or monetary cost with Pause &

Resume Operator Movement.

RECENT OPERATOR PLACEMENT APPROACHES

§  SQPR[14]

§ Query planner for data centers with heterogenes resources

§  MACE[15]

§  Present san approach for precise latency estimation

[14] V. Kalyvianaki et al. „SQPR: Stream query planning with reuse“. In ICDE, 2011. [15] B. Chandramouli et al. „Accurate latency estimation in a distributed event processing system.“ In ICDE, 2011.

STORM LOAD MODEL §  Worker: physical JVM and executes subset of all

the tasks of the topology §  Task: Parallel instance of an operator §  Executor: Thread of an worker, executes one or

more tasks of the same operator

[16] Storm. http://storm.incubator.apache.org/documentation/Understanding-the-parallelism-of-a-Storm-topology.html.

EXAMPLE FOR STORM LOAD MODEL Config conf = new Config(); conf.setNumWorkers(3); // use two worker

processes topologyBuilder.setSpout(„src", new Source(), 3);

topologyBuilder.setBolt(„aggr", new Aggregation(), 6).shuffleGrouping(„src");

topologyBuilder.setBolt(„sink", new Sink(), 6).shuffleGrouping(„aggr");

StormSubmitter.submitTopology( "mytopology", conf, topologyBuilder.createTopology() );

http://storm.incubator.apache.org/documentation/Understanding-the-parallelism-of-a-Storm-topology.html.

STORM: DISTRIBUTION MODEL[16]

Source (spout)

Aggr (bolt)

Sink (bolt)

Parallelism hint =3

Parallelism hint =6

Parallelism hint =3

Worker 1

Source Task

Sink Task

Aggr Task

Worker 2

Source Task

Sink Task

Aggr Task

Worker 3

Source Task

Sink Task

Aggr Task

[16] Storm. http://storm.incubator.apache.org/documentation/Understanding-the-parallelism-of-a-Storm-topology.html.

ADAPTIVE PLACEMENT IN STORM §  Manual reconfiguration (pause & resume

complete topology)

§  Enhanced Load scheduler for Twitter Storm[17] §  Load balancing adapted to Storm architecture

$ storm rebalance mytopo -‐n 4 -‐e src=8

[17] L. Aniello, et al. "Adaptive online scheduling in storm.“ In DEBS, 2013.

DIFFERENT LEVELS OF ELASTICITY

Elastic Action Taken System

Virtual Machine Move virtual machines transparent to the user. [18]

Engine New engine instance is started and new queries are employed on this engine

Query New query instance is started and the input data is split

Operator New operator instance is created and the input data is split

[18] T. Knauth, et al. "Scaling non-elastic applications using virtual machines." In Cloud, 2011. [19] W. Kleiminger, et al."Balancing load in stream processing with the cloud." In ICDEW, 2011. [20] Matteo Migliavacca, et al. "SEEP: scalable and elastic event processing.“ In Middleware, 2010. [21] R. Fernandez, et al. „Integrating scale out and fault tolerance in stream processing using operator state management”, SIGMOD 2013.

ELASTICS DSP SYSTEM ARCHITECTURE

Elasticity Manager

Resource Manager

Data Stream Processing Engine

Virtual Machine

. . . . . .

STREAMCLOUD[22]

§  Realized different level of elasticity on a highly scalable DSP

§  Studied different major design decisions: § Optimal Level of Elasticity § Migration strategy §  Load balancing based on user-defined thresholds

[22] V. Guilsano, et al. „Streamcloud: An elastic and scalable data streaming system”, TPDS, 2012.

SEEP[21]

§  Present the state of the art mechanims for state management

§  combine fault tolerance & scale out using same mechanism

§  Highly scalable and fast recovery

[21] R. Fernandez, et al. "Integrating scale out and fault tolerance in stream processing using operator state management", SIGMOD 2013.

MILLWHEEL[23]

§  Similar mechanisms like SEEP or StreamCloud §  Support massive scale out §  Centralized manager § Optimization of CPU Load

[23] T. Akidau, et al. „MillWheel: Fault-Tolerant Stream Processing at Internet Scale”, VLDB 2013.

Fault tolerance

FAULT TOLERANCE IN DSPS §  Small scale stream processing

§  Faults are an exception § Optimize for the lucky case §  Catch faults at execution time and start recovery

procedures (possibly expensive)

§  Large scale DSPs (e.g. cloud based) §  Faults are likely to affect every execution §  Consider them in the design process

FAULT TOLERANCE IN DSPS §  Two main fault causes

§ Message losses §  Computational element failures

§  Event tracking § Makes sure each injected event is correctly processed

with well-defined semantics

§  State management § Makes sure that the failure of a computational

element will not affect the system‘s correct execution

EVENT TRACKING §  When a fault occurs, events may need to be

replayed §  Typical approach:

§  Request acks from downstream operators § Detect losses by setting timeouts on acks §  Replay lost events from upstream operators

§  Tracking and managing information on processed events may be needed to guarantee specific processing semantics

STORM – ET §  Event tracking in storm is guaranteed by two

different techniques: §  Acker processes è at-least-once message processing §  Transactional topologies è exactly-once message

processing

STORM – ET §  A tuple injected in the system can cause the

production of multiple tuples in the topology §  This production partially maps the underlying

application topology §  Storm keeps track of the processing DAG

stemming from each input tuple

STORM – ET §  Example: word-count topology

SpoutSplitterbolt

Counterbolt

“stay hungry, stay foolish”

text lines words

[“stay”, “hungry”, “stay”, “foolish”] [“stay”, 2][“hungry”, 1][“foolish”, 1]

“stay hungry, stay foolish”

“stay” [“stay”, 1]

“hungry”

“stay”

“foolish”

[“hungry”, 1]

[“foolish”, 1]

[“stay”, 2]

STORM – ET §  Acker tasks are responsible for monitoring the

flow of tuples in the DAG §  Each bolt “acks” the correct processing of a tuple §  The processing of a tuple can be committed when it

has fully traversed the DAG §  In this case the acker notifies the original spout

§  The spout must implement an ack(Object msgId) method §  It can be used to garbage collect tuple state

STORM – ET §  If a tuple does not reach the end of the DAG

§  The acker timeouts and invoke fail(Object msgId) on the spout

§  The spout method implementation is in charge of replaying failed tuples

§  Notice: the original data source must be able to reliably reply events (e.g. a reliable MQ)

STORM – ET §  Developer perspective:

§  Correctly connect bolts with tuple sources (anchoring)

§  Anchoring is how you specify the tuple tree

public void execute(Tuple tuple) { String sentence = tuple.getString(0); for(String word: sentence.split(" ")) { _collector.emit(tuple, new Values(word)); } _collector.ack(tuple); }

Anchor

Source: hOps://github.com/nathanmarz/storm/wiki/Guaranteeing-‐message-‐processing

STORM – ET §  Developer perspective:

§  Explicitly ack processed tuples

§ Or extend BaseBasicBolt

public void execute(Tuple tuple) { String sentence = tuple.getString(0); for(String word: sentence.split(" ")) { _collector.emit(tuple, new Values(word)); } _collector.ack(tuple); }

Explicit ack

Source: hOps://github.com/nathanmarz/storm/wiki/Guaranteeing-‐message-‐processing

STORM – ET §  Implement ack() and fail() on the spout(s)

public void nextTuple() { if(!toSend.isEmpty()){ for(Map.Entry<Integer, String> transactionEntry: toSend.entrySet()){ Integer transactionId = transactionEntry.getKey(); String transactionMessage = transactionEntry.getValue(); collector.emit(new Values(transactionMessage),transactionId); } toSend.clear(); } } public void ack(Object msgId) { messages.remove(msgId); } public void fail(Object msgId) { Integer transactionId = (Integer) msgId; Integer failures = transactionFailureCount.get(transactionId) + 1; if(failures >= MAX_FAILS){ throw new RuntimeException(”Too many failures on Tx [”+transactionId+"]”); } transactionFailureCount.put(transactionId, failures); toSend.put(transactionId,messages.get(transactionId)); }

Source: J. Leibiusky, G. Eisbruch and D. Simonassi. Gekng Started with Storm. O’Reilly, 2012

re-‐schedule

Failed topology Completed processing

STORM – ET §  How does the acker task work?

§  It is a standard bolt §  The acker tracks for each tuple emitted by a spout

the corresponding DAG §  It acks the spout whenever the DAG is complete

§  You can instantiate parallel ackers to improve performance §  Tuples are randomly assigned to ackers to improve

load balancing (uses mod hashing)

STORM – ET §  How can ackers keep track of DAGs?

§  Each tuple is identified by a unique 64bit ID §  The acker stores in a map for each tuple IP

§ The ID of the emitter task

§ An ack val

§  An ack val is a bit-vector that encodes §  IDs of tuples stemmed from the initial one §  IDs of acked tuples

STORM – ET §  Ack val management

§ When a tuple is emitted by a spout §  Initialize the vector an encode in it the tuple ID

§ When a tuple is acked § XOR its ID in the ack val

§ When an anchored tuple is emitted § XOR its ID in the ack val

§ When the ack val is empty, all tuples in the DAG have been acked (with high probability)

STORM – ET §  If exactly-once semantics is required use a

transactional topology §  Transaction = processing + committing §  Processing is heavily parallelized §  Committing is strictly sequential

§  Storm takes care of §  State management (through Zookeeper) §  Transaction coordination §  Fault detection §  Provides a batch processing API

§  Note: requires a source able to reply data batches

Source: hOps://github.com/nathanmarz/storm/wiki/TransacPonal-‐topologies

EVENT TRACKING §  In other systems the event tracking functionality

is strongly coupled with state management §  events cannot be garbage collected when they are

acknowledged § Need to wait for the checkpointing of a state

updated with such events

§  Timestream does not store all the events and recompute those to be replayed by tracking their dependencies with input events, similarly to Storm

EVENT TRACKING §  SEEP stores non-checkpointed events on the

upstream operators §  Millwheel persists all intermediate results to an

underlying Distributed File System. §  It also provides exactly-once semantics

§  D-Streams stores data to be processed in immutable partitioned datasets §  These are implemented as resilient distributed

datasets (RDD)

STATE MANAGEMENT §  Stateful operators require their state to be

persisted in case of failures. §  Two classic approaches

§  Active replication §  Passive replication

STATE MANAGEMENT

Operator

State implicitly synchronized byordered evaluation of same data

Active replication

Operator

Operator(dormant)

State periodically persisted on stable storage and recovered on demand

Passive replication

Checkpoints

STORM - SM §  What happen when tasks fail ?

§  If a worker dies its supervisor restarts it §  If it fails on startup Nimbus will reassign it on a different

machine

§  If a machine fails its assigned tasks will timeout and Nimbus will reassign them

§  If Nimbus/Supervisors die they are simply restarted § Behave like fail-fast processes

§ They’re stateless in practice

§ Their state is safely maintained in in a Zookeeper cluster

STORM - SM §  There is no explicit state management for

operators in Storm §  Trident builds automatic SM on top of it

§  Batch of tuples have unique Tx id §  If a batch is retried it will have the exact same Tx id §  State updates are ordered among batches

§ A new Tx is not committed if an old one is still pending

§  Transactional state guarantees transparently exactly-once tuple processing

STORM - SM §  Transactional state is possible only if supported

by the data source §  As an alternative

§ Opaque transactional state § Each tuple is guaranteed to be executed exactly in one

transaction

§ But the set of transactions for a given Tx id can change in case of failures

§ Non transactional state

STORM - SM §  Few combinations guarantee exactly-once sem.

State Non

transactional Transactional Opaque

transactional

Non transactional No No No

Transactional No Yes ! Yes

Opaque transactional No No Yes

APACHE S4 - SM §  Lightweight approach

§  Assumes that lossy failovers are acceptable. §  PEs hosted on failed PNs are automatically moved to a

standby server §  Running PEs periodically perform uncoordinated and

asynchronous checkpointing of their internal state § Distinct PE instances can checkpoint their state autonomously

without synchronization § No global consistency

§  Executed asynchronously by first serializing the operator state and then saving it to stable storage through a pluggable adapter

§  Can be overridden by a user implementation

MILLWHEEL - SM §  Guarantees strongly consistent processing

§  Checkpoints on persistent storage every single state change incurred after a computation

§  Can be executed § before emitting results downstream

§ operator implementations are automatically rendered idempotent with respect to the execution

§ after emitting results downstream §  it’s up to the developer to implement idempotent operators if

needed

§  Produced results are checkpointed with state (strong productions)

TIMESTREAM- SM §  Takes a similar approach

§  State information checkpointed to stable storage §  For each operator the state includes

§  state dependency: the list of input events that made the operator reach a specific state

§  output dependency: the list of input events that made an operator produce a specific output starting from a specific state.

§  Allow to correctly recover from a fault without having to store all the intermediate events produced by the operators and their states.

§  Is it possibile to periodically checkpoint a full operator state in order to avoid re-emitting the whole history of input events in order to recompute it.

SEEP - SM §  Takes a different route allowing state to be

stored on upstream operators §  allows SEEP to treat operator recovery as a special

case of a standard operator scale-out procedure §  state in SEEP is characterized by three elements

§  internal state

§ output buffers

§  routing state

§  treated differently to reduce the state management impact on system performance.

D-STREAMS- SM §  SM depends strictly on the computation model

§  A computation is structured as a sequence of deterministic batch computations on small time intervals

§  The input and output of each batch, as well as the state of each computation, are stored as reliable distributed datasets (RDDs)

§  For each RDD, the graph of operations used to compute (its lineage) it is tracked and reliably stored

§  The recovery of an RDD can be performed in parallel on separate nodes in order to speed up recovery operation

§  Operator state can be optionally checkpointed on stable storage to limit the number of operations required to restore it.

Open research directions

CONCLUSIONS §  A third generation of DPSs is coming out that

promise §  Unprecedented computational power through

horizontal scalability § On-demand dynamic load adaptation §  Simplified programming models through powerful

event management semantics § Graceful performance degradation in in case of faults

§  A few issues remain to be solved to make DSP ready for the cloud-era

INFRASTRUCTURE AWARENESS §  Most existing DSP systems are infrastructure

oblivious §  deployment strategies do not take into account the

peculiar characteristics of the available hardware §  The physical connection and relationship among

infrastructural element is known to be a key factor to both improve system performance and fault tolerance. §  E.g. Hadoop's “rack awareness”

§  We think infrastructure awareness is an open research field for data stream processing systems that could possibly bring important improvements.

COST-EFFICIENCY §  Users in these days are not only interested in the

performance of the system, but also the monetary cost §  Cloud-based DSP systems should consider running cost as

a variable for performance optimization §  efficient scaling behavior maximizing the system utilization §  efficient fault tolerance mechanisms

§  Bellavista et al. [1] proposed a first prototype, which allows the user to trade-off monetary cost and fault tolerance. §  Their prototype only selects a subset of the operators for

replication based on a user-defined value for the expected information completeness.

[1] P. Bellavista, A. Corradi, S. Kotoulas, and A. Reale. Adaptive fault-tolerance for dynamic resource provisioning in distributed stream processing systems. In EDBT, 2014.

ENERGY-EFFICIENCY §  Most large-scale datacenters are striving to "go

green" by making energy consumption more effective.

§  DSP systems should consider energy consumption a yet-another-variable in their performance optimization process

§  Note: this aspect is possibly strictly linked to the infrastructure awareness theme

ADVANCED ELASTICITY §  Most of the existing elastic scaling solutions for

DSPs apply simplistic schemes for load balancing. § Operator placement algorithms only optimize system

utilization § Other metrics (end to end latency, network

bandwidth, etc.) are only partially considered

§  However these metric are often used to sign contract-binding SLAs

§  Would it be possible to design DSP systems able to probabilistically guarantee performance ?

MULTI-DSP INTEGRATION §  Most DSPs are particularly well suited for specific

use cases §  No one-size-fits-all solution §  Components automatically selecting the best engine

for each give use case would significantly improve the applicability of these systems. §  No need to know in advance which is the best solution

for given use case §  No need to deploy and maintain different solutions

§  First results are promising [2,3]

[2] M. Duller, J. S. Rellermeyer, G. Alonso, and N. Tatbul. Virtualizing stream processing. Middleware, 2011. [3] H. Lim and S. Babu. Execution and optimization of continuous queries with cyclops. SIGMOD, 2013.

Performance

MapReduce

Performance

MULTI-DSP INTEGRATION

Performance

Cyclops New Query

With Window Size w

Choose DSP

[3] H. Lim and S. Babu. Execution and optimization of continuous queries with cyclops. SIGMOD, 2013.

THANK YOU !��

cloud based data stream processing -...

Documents

grid-based data stream processing in e-science€¦ ·...

survey stream processing - indiana university...

distributed real-time stream processing: why and...

dancing with stream processing

oracle stream explorer - simplifying event/stream processing

high performance stream processing

scalable distributed stream...

stream processing frameworks

xml stream processing

introduction to stream processing

twitter stream processing

big data stream processing - scads · big data stream...

scalable stream processing spark streaming and flink stream

stream processing application

stream processing platforms data

concurrent stream processing

data management 13 stream processing - mboehm7.github.io ·...

stateful distributed stream processing

continuous data stream processing

stream processing blocks