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-a-service from
cloud-based providers ?
225/05/14 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
4
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

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

CLOUD-BASED
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
8
[1] M. Stonebraker, U. Cetintemel and S. Zdonik. "The 8 Requirements of real-time Stream Processing", ACM
SIGMOD Record, 2005.

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 GENERATIONSECOND GENERATION
Aurora (2002)
Medusa (2004)
Borealis (2005)
TelegraphCQ (2003)
NiagaraCQ
(2001)
STREAM (2003)
StreamMapReduce(2011)
StreamMine3G (2013)
Yahoo! S4(2010)
Apache S4(2012)
Twitter Storm (2011) Trident (2013)
StreamCloud(2010)
Elastic StreamCloud (2012)
SPADE(2008)
Elastic 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)
Elastic 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
11
[2] S. Chandrasekaran et al. "TelegraphCQ: Continuous Dataflow Processing", SIGMOD, 2003.
Source: 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
12
[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
Source: D. J. Abadi et al. "The Design of the Borealis Stream Processing Engine", CIDR 2005.

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

CLOUD-BASED
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
1625/05/14 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[4]
 Incoming searches are organized in 1-second buckets
 Buckets are compared to historical data
 Useful to promptly detect anomalies (spikes/dips)
17
[4] Akidau, Balikov, Bekiroğlu, Chernyak, Haberman, Lax, McVeety, Mills, Nordstrom, Whittle. MillWheel: Fault-tolerant
Stream Processing at Internet Scale.VLDB, 2013.

POPULAR QUERY ANALYSIS AT TWITTER
 When a relevant event happens, an increase occurs in the
number of queries submitted to Twitter[5]
 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
18
[5] Improving Twitter Search with Real-time Human Computation, 2013 http://blog.echen.me/2013/01/08/improving-
twitter-search-with-real-time-human-computation

 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
19
how can we understand that
#bindersfullofwomen refers
to politics?

20
Search
Engine
Kafka
Queue
query Log
Spout
Bolt
(query, ts)
popular
queries
Human
Evaluation
categorizations of queries
engine tuning
Storm

USER PROFILING AT YAHOO!
 Queries submitted by users are evaluated (thousands
queries per second by millions of users)
 Run on S4[6]
 Useful to generate highly personalized advertising
21
[6] L. Neumeyer, B. Robbins, A. Nair, and A. Kesari. S4: Distributed Stream Computing Platform.ICDMW, 2010.

BUS ROUTING MANAGEMENT IN DUBLIN
 Tracking of bus locations (1000 buses) to improve
public transportation for 1.2 million citizens[7]
 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
22
[7] 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[8]
 Useful to continuously estimate the mood about specific
topics
23
[8] 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[9]
 Useful to reactively spot dishonest behaviors
24
[9] V. Gulisano, R. Jimenez-Peris, M. Patino-Martinez, C. Soriente, and P. Valduriez. Streamcloud: An Elastic and Scalable
Data Streaming System. IEEE TPDS, 2012.

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)

CLOUD-BASED
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:
Spout
Spout
Spout
Spout
Bolt Bolt
Bolt
BoltBolt
Bolt
Bolt
Bolt

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[10] offers a more high level programming
interface
[10] 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
Spout
Spout
Spout
Spout
Bolt Bolt
Bolt
BoltBolt
Bolt
Bolt
Bolt
x3
x8
x5

STORM INTERNALS
 A storm cluster is constituted by a Nimbus node
and n Worker nodes

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

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: http://wpcertification.blogspot.it/2014/02/helloworld-apache-storm-word-counter.html
H elloStorm
LineReader
Spout
W ordSplitter
Bolt
W ordCounter
Bolt
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 WordSplitterBolt 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"));
}
}

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());
}
}
}

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 WordSplitterBolt().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();
}
}

CLOUD-BASED
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?
28/05/2014 Cloud-based Data Stream Processing
42

RQUIREMENTS FOR DATA
PARALLELISM
 Transparent to the user
 correct results in correct order (identical to
sequential execution)
Filter
Filter
Round-
Robin Union
Aggr
Aggr
Hash
(groupID)
Merge&
Sort
43

DATA PARALLELISM
 First presented by FLUX[11] and Borealis[12]
 Explicitely done using partitioning operators
 User needs to decide:
 partitioning scheme
 merging scheme
 level of parallelism
44
[11] M. Shah, et al. "Flux: An adaptive partitioning operator for continuous query systems." In ICDE, 2003.
[12] 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
45

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
46

MAPREDUCE FOR STREAMING [13,14]
 Extend MapReduce model for streaming data:
 Break strict phases
 Introduce Stateful Reducer
Map
Map
Map
Reduce
Reduce
OutputInput
[13] T. Condie, et al. "MapReduce Online." In NSDI,2010.
[14] A. Brito, et al. "Scalable and low-latency data processing with StreamMapReduce." CloudCom, 2011.
47

STATEFUL REDUCER[14]
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);
}
[14] A. Brito, et al. "Scalable and low-latency data processing with StreamMapReduce." CloudCom, 2011.
28/05/2014 Cloud-based Data Stream Processing 48

AUTO PARALLELIZATION[15,16]
 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
[15] S Schneider, et al. "Auto-parallelizing stateful distributed streaming applications." In PACT, 2012.
[16] Wu et al. “Parallelizing Stateful Operators in a Distributed Stream Processing System: How, Should You and How
Much?” In DEBS, 2012.
49

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
50

ELASTICITY[17]
 Goal: React to unpredicted load peaks & reduce
the amount of ideling resources.
Workload
Load
Static Provisioning
Elastic Provisioning
Underprovisioning
Overprovisioning
[17] M. Armbrust, et al. "A view of cloud computing." Communications of the ACM, 2010.
51

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.
52

ELASTIC OPERATOR EXECUTION[18]
 Dynamically adapt number of threads for a
stateless operator based on the workload
 Limitations: works only on thread-level and for
stateless operators
[18] S. Schneider, et al. "Elastic scaling of data parallel operators in stream processing." In IPDPS, 2009.
Operator
Thread Pool
53

ELASTIC AUTO-PARALLELIZATION[19]
 Combines ideas of Elasticity and Auto-
Parallelization
 Adapt parallelization level using controller-based
approach
 Dynamic adaption of parallelization level based
on operator migration protocol
54
[19] B. Gedik, et al. “Elastic Scaling for Data Stream Processing." In TPDS, 2013.

Horizontal barrierVertical
barrier
ELASTIC AUTO-PARALLELIZATION
 Dynamic adaption of parallelization: lend &
borrow approach
55
Op1
Op2
Splitter Merger
Op2
Op2
1. Lend Phase2. 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[20]
 Optimization goal: What is optimized ?
 Execution mode: Centralized or Decentralized?
 Algorithm runtime: Offline, online or both?
56
[20] 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)
57

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
58

DECENTRALIZER EXECUTION
 Based on pairwise agreement[21]
Host 1 Host 2 Host 3 Host 4
Aggr4 Aggr2
Filter 2
Filter 3
Filter 5
Filter 6
Filter 4
Filter 1
Aggr 3Aggr 4
Aggr 3
Aggr 5Aggr 6
Source 1
Source 2
59
[21] M. Balazinska, et al. "Contract-Based Load Management in Federated Distributed Systems." NSDI. Vol. 4.
2004.

DECENTRALIZER EXECUTION
 Based on decentralized routing[22]
Host 2
Host 3
Host 4
Host 1
Src1
Src2
J
D
F F
[22] 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
61

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[23]
62
[23] 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
63

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
64

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.
65

RECENT OPERATOR PLACEMENT
APPROACHES
 SQPR[24]
 Query planner for data centers with heterogenes
resources
 MACE[25]
 Present san approach for precise latency estimation
[24] V. Kalyvianaki et al. "SQPR: Stream query planning with reuse“. In ICDE, 2011.
[25] B. Chandramouli et al. "Accurate latency estimation in a distributed event processing system“. In ICDE, 2011.
66

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
67

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[26]
Source
(spout)
Aggr
(bolt)
Sink
(bolt)
Parallelism
hint =3
Parallelism
hint =6
Parallelism
hint =3
Worker 1
Source Task
Sink Task
Aggr Task
Aggr Task
Worker 2
Source Task
Sink Task
Aggr Task
Aggr Task
Worker 3
Source Task
Sink Task
Aggr Task
Aggr Task
[26] 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[27]
 Load balancing adapted to Storm architecture
$ storm rebalance mytopo -n 4 -e src=8
70
[27] 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. [28]
Engine New engine instance is started and new queries are
employed on this engine
[29]
Query New query instance is started and the input data is
split
[30]
Operator New operator instance is created and the input data
is split
[31]
[28] T. Knauth, et al. "Scaling non-elastic applications using virtual machines." In Cloud, 2011.
[29] W. Kleiminger, et al."Balancing load in stream processing with the cloud." In ICDEW, 2011.
[30] M. Migliavacca, et al. "SEEP: scalable and elastic event processing.“ In Middleware, 2010.
[31] R. Fernandez, et al. „Integrating scale out and fault tolerance in stream processing using operator state
management”, SIGMOD 2013.
71

ELASTICS DSP SYSTEM ARCHITECTURE
Elasticity
Manager
Resource
Manager
Data Stream
Processing Engine
Data Stream
Processing Engine
Data Stream
Processing Engine
Virtual
Machine
Virtual
Machine
Virtual
Machine
...
...
72

STREAMCLOUD[9]
 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
[9] V. Gulisano, R. Jimenez-Peris, M. Patino-Martinez, C. Soriente, and P. Valduriez. Streamcloud: An Elastic and
Scalable Data Streaming System. IEEE TPDS, 2012.

SEEP[31]
 Present the state of the art mechanims for state
management
 combine fault tolerance & scale out using same
mechanism
 Highly scalable and fast recovery
[31] R. Fernandez, et al. "Integrating scale out and fault tolerance in stream processing using operator state
management", SIGMOD 2013.

MILLWHEEL[4]
 Similar mechanisms like SEEP or StreamCloud
 Support massive scale out
 Centralized manager
 Optimization of CPU Load
[4] T. Akidau, et al. „MillWheel: Fault-Tolerant Stream Processing at Internet Scale”, VLDB 2013.

CLOUD-BASED
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  at-least-once message processing
 Transactional topologies  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
Spout
Splitter
bolt
Counter
bolt
“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: https://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: https://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. Getting 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
 Initializes the vector and encodes 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

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 re-
compute 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
Operator
Operator
Previousstage
Nextstage
State
State
State
Data
State implicitly synchronized by
ordered evaluation of same data
Active replication
Operator
Operator
(dormant)
Operator
(dormant)
Previousstage
Nextstage
State
State
State
Data
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
Spout
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.

CLOUD-BASED
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 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.[32] 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.
[32] 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 promising results[33,34]
[33] M. Duller, J. S. Rellermeyer, G. Alonso, and N. Tatbul. Virtualizing stream processing. Middleware, 2011.
[34] H. Lim and S. Babu. Execution and optimization of continuous queries with cyclops. SIGMOD, 2013.

Storm
Performance
MapReduce
Performance
MULTI-DSP INTEGRATION
ESPER
Performance
Cyclops
New Query
With Window Size w
Choose DSP
1
2
[34] H. Lim and S. Babu. Execution and optimization of continuous queries with cyclops. SIGMOD, 2013.

REFERENCES
[1] M. Stonebraker, U. Cetintemel and S. Zdonik "The 8 Requirements of real-time Stream Processing", ACM
SIGMOD Record, 2005.
[2] S. Chandrasekaran et al. "TelegraphCQ: Continuous Dataflow Processing", SIGMOD, 2003.
[3] D. J. Abadi et al. "The Design of the Borealis Stream Processing Engine", CIDR 2005.
[4] T. Akidau, A. Balikov, K. Bekirog ̆lu, S. Chernyak, J. Haberman, R. Lax, S. McVeety, D. Mills, P. Nordstrom,
and S. Whittle. "MillWheel: Fault-tolerant Stream Processing at Internet Scale". VLDB, 2013.
[5] Improving Twitter Search with Real-time Human Computation, 2013 http://blog.echen.me/2013/01/08/improving-
twitter-search-with-real-time-human-computation
[6] L. Neumeyer, B. Robbins, A. Nair, and A. Kesari. S4: Distributed Stream Computing Platform.ICDMW,
2010.
[7] "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
[8] 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.
[9] V. Gulisano, R. Jimenez-Peris, M. Patino-Martinez, C. Soriente, and P. Valduriez. "Streamcloud: An Elastic
and Scalable Data Streaming System". IEEE TPDS, 2012.
[10] N. Marz. "Trident tutorial". https://github.com/nathanmarz/storm/wiki/Trident-tutorial.
[11] M. Shah, et al. "Flux: An adaptive partitioning operator for continuous query systems." In ICDE, 2003.
[12] D. Abadi, et al. "The Design of the Borealis Stream Processing Engine." In CIDR, 2005..
[13] T. Condie, et al. "MapReduce Online." In NSDI,2010.
[14] A. Brito, et al. "Scalable and low-latency data processing with stream mapreduce." CloudCom, 2011.
[15] S. Schneider, et al. "Auto-parallelizing stateful distributed streaming applications." In PACT, 2012.

REFERENCES
[16] Wu et al. “Parallelizing Stateful Operators in a Distributed Stream Processing System: How, Should You
and How Much?” In DEBS, 2012.
[17] M. Armbrust, et al. "A view of cloud computing." Communications of the ACM, 2010.
[18] S. Schneider, et al. "Elastic scaling of data parallel operators in stream processing." In IPDPS, 2009.
[19] B. Gedik, et al. “Elastic Scaling for Data Stream Processing." In TPDS, 2013.
[20] G.T. Lakshmanan, et al. "Placement strategies for internet-scale data stream systems." In IEEE Internet
Computing, 2008.
[21] M. Balazinska, et al. "Contract-Based Load Management in Federated Distributed Systems." NSDI. Vol. 4.
2004.
[22] Y.Ahmad, et al. "Network-aware query processing for stream-based applications. In VLDB, 2004.
[23] Y. Zhu, et al. "Dynamic plan migration for continuous queries over data streams." In SIGMOD, 2004.
[24] V. Kalyvianaki et al. "SQPR: Stream query planning with reuse“. In ICDE, 2011.
[25] B. Chandramouli et al. "Accurate latency estimation in a distributed event processing system“. In ICDE,
2011.
[26] Storm. http://storm.incubator.apache.org/documentation/Understanding-the-parallelism-of-a-Storm-
topology.html.
[27] L. Aniello, et al. "Adaptive online scheduling in storm.“ In DEBS, 2013.
[28] T. Knauth, et al. "Scaling non-elastic applications using virtual machines." In Cloud, 2011.
[29] W. Kleiminger, et al."Balancing load in stream processing with the cloud." In ICDEW, 2011.
[30] M. Migliavacca, et al. "SEEP: scalable and elastic event processing.“ In Middleware, 2010.

REFERENCES
[31] R. Fernandez, et al. "Integrating scale out and fault tolerance in stream processing using operator state
management", SIGMOD 2013.
[32] P. Bellavista, A. Corradi, S. Kotoulas, and A. Reale. "Adaptive fault-tolerance for dynamic resource
provisioning in distributed stream processing systems". In EDBT, 2014.
[33] M. Duller, J. S. Rellermeyer, G. Alonso, and N. Tatbul. "Virtualizing stream processing". Middleware, 2011.
[34] H. Lim and S. Babu. "Execution and optimization of continuous queries with cyclops". SIGMOD, 2013.

Cloud-based Data Stream Processing

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (20)

Similar to Cloud-based Data Stream Processing

Similar to Cloud-based Data Stream Processing (20)

More from Zbigniew Jerzak

More from Zbigniew Jerzak (14)

Recently uploaded

Recently uploaded (20)

Cloud-based Data Stream Processing