![Darach
Ennis
(darach@streambase.com)
@darachennis
StreamBase
Systems
Erlang
Factory
London
-‐
June
10th
2011
Complex
Er[jl]ang
Processing
with
StreamBase:
A
DSL
for
Low
Latency
High
Frequency
CompuOng
+ + =
Copyright 2011 © StreamBase Systems, Inc. 1](https://image.slidesharecdn.com/complexerjlangprocessing-pptx-110611095407-phpapp01/85/Complex-Er-jl-ang-Processing-with-StreamBase-1-320.jpg)
![Expose
to
StreamBase
[
Call
by
Behaviour/Mod]
#1
Copyright 2011 © StreamBase Systems, Inc. 35](https://image.slidesharecdn.com/complexerjlangprocessing-pptx-110611095407-phpapp01/85/Complex-Er-jl-ang-Processing-with-StreamBase-35-320.jpg)
![Expose
To
StreamBase
[Call
by
Behaviour/Mod]
#2
Copyright 2011 © StreamBase Systems, Inc. 36](https://image.slidesharecdn.com/complexerjlangprocessing-pptx-110611095407-phpapp01/85/Complex-Er-jl-ang-Processing-with-StreamBase-36-320.jpg)
Complex Er[jl]ang Processing with StreamBase
- 1. Darach Ennis (darach@streambase.com) @darachennis StreamBase Systems Erlang Factory London -‐ June 10th 2011 Complex Er[jl]ang Processing with StreamBase: A DSL for Low Latency High Frequency CompuOng + + = Copyright 2011 © StreamBase Systems, Inc. 1
- 2. Agenda What is ‘Complex Event Processing’ • Specifically flow oriented event processing (there are others) • Streams & Operators. Windowing, Branching, Combining, Extending A day in the life of a flow programmer • RelaOvity -‐ Data parallelism, concurrency, latency & throughput • ConOnuity -‐ ConOnuous Streaming Map Reduce • Reliability – High availability, the low latency way • Flow, meets FuncOon. Embed Erlang in process via Erjang Integra?on. Erlang – the ecosystem. • Calling Erlang from StreamBase – Simple & windowed funcOons • Client/Server – Pushing events to/from StreamBase • RabbitMQ -‐ Messaging TheC. Erlang – the inspira?on. Paxos, in StreamBase Copyright 2011 © StreamBase Systems, Inc. 2
- 3. High Level DSLs : Myth Vs Reality Myth: High level domain specific languages are too slow for HFT. Reality: High level domain specific languages can deliver beLer performance than system programming languages when tailored to a specific task. Copyright 2011 © StreamBase Systems, Inc. 3
- 4. Complex Event Processing aka Event Processing SoCware organized by events (cf: object/func?on oriented) • What’s an event? What’s an object? • Something that can trigger processing, can include data. • Naturally but not usually represents a “real world” events & observaOons. Complex Event Processing PlaTorms • Soeware stack for event based systems, event driven architectures • Event Programming Language – SQL-‐based, Rules-‐based, or State-‐based • Commercial and open source: StreamBase, Progress, Microsoe, IBM, Oracle, SAP, Esper, Drools and many more Adopted in financial services and other markets • System monitoring, industrial process control, logisOcs, defense/ intelligence Other Event Processing Approaches: • Erlang, Scala/Akka, Actors, node.js, .NET Rx Copyright 2011 © StreamBase Systems, Inc. 4
- 5. What does a CEP DSL or Language offer? Con?nuously Observe, Orient, Decide Act (OODA) on event streams Wikipedia: • ConOnuous Incremental Query John Boyd, • Pajern matching within or across streams OODA Loop • Branch – Split, Causal Split, Filter. • Combine -‐ Semi-‐Join, Union, Gather, Merge, Join, Pajern • Windows – Process sets of streaming data • Sliding or Tumbling, Overlapping or Non-‐Overlapping, Gaps or No Gaps • Finite (1 second, 1000 tuples), Infinite • Emission Policies: On Close, Every odd message • Predicate based – Roll your own window type • State Management – In memory, CSV files, CSV sockets, RDBMS, Parallel DBMSs, Column Stores, KV stores, NoSQL, NewSQL… • Nice to have: • DeclaraOve concurrency, Interface Polymorphism, DistribuOon, Extensible Copyright 2011 © StreamBase Systems, Inc. 5
- 6. Challenges for CEP ‘Über’ Ultra Low Latency? • Sub-‐milli is standard, sub-‐100-‐micro is desirable. Less is more! Large Data Volumes • Hundreds of thousands of events, thousands of decisions, per thread. • Big Data. ~Hundreds of SMP CEP nodes. Demanding Opera?onal Environment • 24x7, 365 – in criOcal environments (trading, surveillance, uOliOes) Sophis?cated Data Processing (some?mes) • OpOons pricing, yield curves, risk metrics, smart grid capacity planning, fraud detecOon. How it’s done (QCON London 2011): • How LMAX did it? hjp://bit.ly/fUeS0P • How we did it? hjp://bit.ly/hM6NAP <-‐ Our CTO Richard Tibbejs talk Copyright 2011 © StreamBase Systems, Inc. 6
- 7. StreamBase Event Processing Plauorm Studio Integrated Development Environment Visualiza?on Developer Studio Graphical StreamSQL for developing, back tesOng and deploying applicaOons. StreamBase StreamBase Component Frameworks Exchange ApplicaOons Input Adapter(s) Inject streaming (market data) Adapters Adapters and staOc (reference data) StreamBase Server sources. Event Processing Server Output Adapter(s) High performance opOmized engine Send results to systems, users, user can process events at market data screens and databases. speeds. Copyright 2011 © StreamBase Systems, Inc. 7
- 8. How did we do it? Compila?on and Sta?c Analysis • Design the language for it Modular abstrac?on, interfaces • Quants and Developers Collaborate, share code Bytecode genera?on and the Janino compiler • OpOmized bytecodes, in-‐memory generaOon Garbage op?miza?on • Pooling, data class, invasive collecOons Integra?ons, C++ and Java plugins • Efficient naOve interfaces, Hardware acceleraOon Adapter API, FIX Messaging • Threading and API structure for ultra low latency Parallelism, Clustering, Lanes and Tiers • Scalability, with a latency bias. Modularity through Named Data Formats, Schemas • Sharing data and semanOcs between apps Copyright 2011 © StreamBase Systems, Inc. 8
- 9. StreamBase StreamSQL EventFlow Rapid Deployment & Unit TesOng Modularity & Polymorphism Interfaces & Off The Shelf Off The Shelf Extension Points Business Logic ConnecOvity Copyright 2011 © StreamBase Systems, Inc. 9
- 10. Operators – Hi Erlang. Hello StreamBase Did you just tell me to go flow myself? Copyright 2011 © StreamBase Systems, Inc. 10
- 11. Agenda What is ‘Complex Event Processing’ • Specifically flow oriented event processing (there are others) • Streams & Operators. Windowing, Branching, Combining, Extending A day in the life of a flow programmer • RelaOvity -‐ Data parallelism, concurrency, latency & throughput • ConOnuity -‐ ConOnuous Streaming Map Reduce • Reliability – High availability, the low latency way • Flow, meets FuncOon. Embed Erlang in process via Erjang Integra?on. Erlang – the ecosystem. • Calling Erlang from StreamBase – Simple & windowed funcOons • Client/Server – Pushing events to/from StreamBase • RabbitMQ -‐ Messaging TheC. Erlang – the inspira?on. Paxos, in StreamBase Copyright 2011 © StreamBase Systems, Inc. 11
- 12. A day in the life… Why a DSL? High level – Windowing, Combina?on & PaLern Matching Streams Graphical – ‘See’ the flow, dependencies, pathways Fast, Flexible SDLC – Deploy new algorithms, con?nuously Understandable – Rise to the abstrac?on Flexible Copyright 2011 © StreamBase Systems, Inc. 12
- 13. A day in the life.. More is more! Copyright 2011 © StreamBase Systems, Inc. 13
- 14. Make it work. Measure it. -‐ Baseline ‘Noop’ Performance WARNING: Micro-‐ benchmark. YMMV Copyright 2011 © StreamBase Systems, Inc. 14
- 15. Concurrency – Baseline -‐ Results Nice. 1.1 million events/sec Copyright 2011 © StreamBase Systems, Inc. 15
- 16. Parallelize it – The ‘Convenient, but incorrect’ way Execution Order & Concurrency: Rule 1 – Each event processed to completion left-right Rule 2 – Branches processed sequentially Rule 3 – Outputs are processed sequentially Rule 4 – Module output processed to completion immediately Rule 5 – One operator executed at a time Copyright 2011 © StreamBase Systems, Inc. 16
- 17. Gah! WTF?! Copyright 2011 © StreamBase Systems, Inc. 17
- 18. OpOmize it! ObservaOon #1 Measure the cost of thread context switches. Dodge & Burn! Copyright 2011 © StreamBase Systems, Inc. 18
- 19. OpOmize it! ObservaOon #2 – It’s Map/Combine/Reduce Copyright 2011 © StreamBase Systems, Inc. 19
- 20. CSMR – A ‘pajern’ for low latency high throughput? Copyright 2011 © StreamBase Systems, Inc. 20
- 21. CSMR – A ‘pajern’ for low latency high throughput? Yup! 25 20 20 20 18 Millions of tuples / second Throughput (Mtps) 15 Across Region 10 10 9.4 Within Region 8 8 5 4 4 4 0 Gather Merge Branch & Join Filter Even Random Baseline (Noop) Benchmark HW: Dell RS510, 8 core, 32 GB RAM, 2.4GHz IA-‐64, OS: RHEL 5u3, SB 7.0.1.2 Copyright 2011 © StreamBase Systems, Inc. 21
- 22. CSMR + GPUs / FPGAs? Sure! We can build on SMP and cluster wide distribution algorithms to further optimize interactions with accelerated compute technologies. We can exploit accelerated hardware messaging where regular network IO is insufficient for distribution results over 1Ge or regular networking technologies. The pattern above is a continuous streaming variant of Map/Reduce in EventFlow. Copyright 2011 © StreamBase Systems, Inc. 22
- 23. Similar OpOmizaOon’s apply to GPUs too! Copyright 2011 © StreamBase Systems, Inc. 23
- 24. A day in the life.. Reliability Copyright 2011 © StreamBase Systems, Inc. 24
- 25. A day in the life.. Reliability Hot/Hot, Hot/Warm, Hot/Cold, None?! Copyright 2011 © StreamBase Systems, Inc. 25
- 26. A day in the life.. Reliability Copyright 2011 © StreamBase Systems, Inc. 26
- 27. A day in the life.. Reliable CSMR Copyright 2011 © StreamBase Systems, Inc. 27
- 28. Agenda What is ‘Complex Event Processing’ • Specifically flow oriented event processing (there are others) • Streams & Operators. Windowing, Branching, Combining, Extending A day in the life of a flow programmer • RelaOvity -‐ Data parallelism, concurrency, latency & throughput • ConOnuity -‐ ConOnuous Streaming Map Reduce • Reliability – High availability, the low latency way Integra?on. Erlang – the ecosystem. • Calling Erlang from StreamBase – Simple & windowed funcOons • Client/Server – Pushing events to/from StreamBase • RabbitMQ -‐ Messaging TheC. Erlang – the inspira?on. Paxos, in StreamBase Copyright 2011 © StreamBase Systems, Inc. 28
- 29. Operators – Hi Erlang. Hello StreamBase Aggregate – A window of moving data • Time based – eg: the last 10 seconds, .001 seconds, day • Field based – eg: historical Ome (Omestamp, …) • Tuple based – eg: the last 1000 tuples • Predicate based – expressions determine when to open, close and/or emit interesOng results from a window Mul?-‐dimensional • Give me the last seconds worth of events or the last 10000, whichever happens first Grouping • ParOOon by Symbol implies a window per Symbol ‘concurrently – on the same thread’ Copyright 2011 © StreamBase Systems, Inc. 29
- 30. Windows in the Wild -‐ #1 Riak Riak Core: hLps://github.com/basho/riak_core • Mixes the ‘when’ window dimension (Ome) with the what ‘aggregaOon’ Copyright 2011 © StreamBase Systems, Inc. 30
- 31. Windows in the Wild -‐ #2 tlack on BitBucket Thomas Lackner -‐ tlack -‐ EMA: • hjps://bitbucket.org/tlack/erlang-‐exponenOal-‐moving-‐average/overview • Mixes the ‘when’ window dimension (Ome) with the what ‘aggregaOon’ and number of occurrences ‘how many’ Copyright 2011 © StreamBase Systems, Inc. 31
- 32. Embedding Erlang (Erjang) Prepend the dir defining our behaviors to the code path Copyright 2011 © StreamBase Systems, Inc. 32
- 33. Define Behavioral Contracts Copyright 2011 © StreamBase Systems, Inc. 33
- 34. Implement & Test in Erlang (or Erjang!) Duh! Typo! Copyright 2011 © StreamBase Systems, Inc. 34
- 35. Expose to StreamBase [ Call by Behaviour/Mod] #1 Copyright 2011 © StreamBase Systems, Inc. 35
- 36. Expose To StreamBase [Call by Behaviour/Mod] #2 Copyright 2011 © StreamBase Systems, Inc. 36
- 37. Use, Deploy & Run Wikipedia: MACD Example: Continuous Streaming Map/Reduce with 1-second MACD compression. Linearly SMP scalable. Just add boxes to scale! Copyright 2011 © StreamBase Systems, Inc. 37
- 38. Run, Rabbit, Run, Rabbit, … Copyright 2011 © StreamBase Systems, Inc. 38
- 39. Agenda What is ‘Complex Event Processing’ • Specifically flow oriented event processing (there are others) • Streams & Operators. Windowing, Branching, Combining, Extending A day in the life of a flow programmer • RelaOvity -‐ Data parallelism, concurrency, latency & throughput • ConOnuity -‐ ConOnuous Streaming Map Reduce • Reliability – High availability, the low latency way Integra?on. Erlang – the ecosystem. • Calling Erlang from StreamBase – Simple & windowed funcOons • Client/Server – Pushing events to/from StreamBase • RabbitMQ -‐ Messaging TheC. Erlang – the inspira?on. Paxos, in StreamBase Copyright 2011 © StreamBase Systems, Inc. 39
- 40. SB Paxos – Entrypoint (cc @kevsmith!) Copyright 2011 © StreamBase Systems, Inc. 40
- 41. SB Paxos – Autonomic, Self Healing Ring Copyright 2011 © StreamBase Systems, Inc. 41
- 42. SB Paxos – Generic FSM ‘Behaviour’ Copyright 2011 © StreamBase Systems, Inc. 42
- 43. SB Paxos – Generic FSM ‘Behaviour’ – Paxos FSM StateMap,State,TransiOon,NextState,AcOon,DescripOon, BasicPaxosMap,IniOal,Bootstrap,Proposer,Prepare,"..." BasicPaxosMap,IniOal,PromiseOk,Accepter,PromiseOk,"..." BasicPaxosMap,IniOal,PromiseNotOk,Done,PromiseNotOk,"..." BasicPaxosMap,IniOal,Accepted,Learner,Response,"When ..." BasicPaxosMap,Proposer,Accept,Proposer,Accept,"..." BasicPaxosMap,Proposer,PromiseNotOk,Proposer,DoNothing,"..." BasicPaxosMap,Proposer,Response,Done,Gone,DoStop,"..." BasicPaxosMap,Accepter,Accept,Learner,Accept,"..." BasicPaxosMap,Accepter,Accepted,Learner,Response,"..." BasicPaxosMap,Learner,Accepted,Done,Response,"..." BasicPaxosMap,Learner,Done,Gone,DoStop,"..." Copyright 2011 © StreamBase Systems, Inc. 43
- 44. Shameless Plugs StreamBase • You could build one of these yourself, or use ours… • Download and test out the full product hjp:/www.streambase.com • Build something and submit to the StreamBase Component Exchange hjp://sbx.streambase.com • Contact us to buy or to an OEM partner, offices London, Boston, New York • We’re hiring • We’re training • hjp://www.streambase.com/developers-‐training-‐events.htm DEBS – Distributed Event Based Systems • Academic (ACM) Conference outside NYC in July EPTS – Event Processing Technology Society • hjp://ep-‐ts.org industry consorOum Ques?ons? Copyright 2011 © StreamBase Systems, Inc. 44
- 45. Acknowledgements Erlang, Erlang Solu?ons, Erlang UG London, & Erlang Factory • hjp://www.erlang.org/ • hjp://www.erlang-‐soluOons.com/ • hjp://www.erlang-‐soluOons.com/etc/usergroup/london • hjp://www.erlang-‐factory.com/ Erjang – The Java based Erlang Virtual Machine • hjps://github.com/trifork/erjang/wiki/ erlIDE • hjp://erlide.sourceforge.net/ @?bbeLs – I stole borrowed some of his slides! Download StreamBase and tell us what you think: • hjp://www.streambase.com Copyright 2011 © StreamBase Systems, Inc. 45
- 46. Download StreamBase and More InformaOon hjp://www.streambase.com QuesOons? + + = Copyright 2011 © StreamBase Systems, Inc. 46