Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Real Time Analytics: Algorithms and Systems

16,802 views

Published on

In this tutorial, an in-depth overview of streaming analytics -- applications, algorithms and platforms -- landscape is presented. We walk through how the field has evolved over the last decade and then discuss the current challenges -- the impact of the other three Vs, viz., Volume, Variety and Veracity, on Big Data streaming analytics.

Published in: Data & Analytics

Real Time Analytics: Algorithms and Systems

  1. 1. Real-time Analytics Algorithms and Systems Arun  Kejariwal*,  Sanjeev  Kulkarni+,  Karthik  Ramasamy☨   *Machine  Zone,  +PeerNova,  ☨Twi@er @arun_kejariwal,  @sanjeevrk,  @karthikz
  2. 2. 2 A look at our presentation agenda Outline Motivation Why  bother? Emerging Applications IoT,  Health  Care,  Machine  data   Connected  vehicles
  3. 3. 3 Algorithms: I ClassificaAon Systems: II 3rd  GeneraAon Systems: I 1st  &  2nd  GeneraAon Algorithms: II Deep  Dive
  4. 4. 4 The Road Ahead Challenges Closing Q&A
  5. 5. 5 Real-time is key Information Age Ká !
  6. 6. 6 Large  variety  of  media   Blogs,  reviews,  news  arAcles,    streaming  content   > 500M Tweets  everyday Challenge: Surfacing Relevant Content Explosive Content Creation [1]  hPp://www.kpcb.com/blog/2014-­‐internet-­‐trends   > 300 hrs Video  uploaded  every  minute > 1.8 B Photos  uploaded  online  in  2014  [1]
  7. 7. 7 High Volume Content Consumption WhatsApp Messages  per  day  [1] Pandora Listener  hours     (Q2  2015)  [3] Skype Calls  per  month E-mails Per  second Google Searches  /year  [2] Netflix Hours  streamed     per  month >30B 5.3B 4.76B >  1T >2.2M >  1B ! É [1]  hPps://www.facebook.com/jan.koum/posts/10152994719980011?pnref=story   [2]  hPp://searchengineland.com/google-­‐1-­‐trillion-­‐searches-­‐per-­‐year-­‐212940   [3]  hPp://press.pandora.com/phoenix.zhtml?c=251764&p=irol-­‐newsArAcle&ID=2070623 ] 9
  8. 8. 8 A New World Mobile, Mobile, Mobile 5.4  B  Mobile  Phone  Users  [1] 69%  Y/Y  Growth  Data  Traffic  55%  Mobile  Video  Traffic 34%  Global  e-­‐Commerce  [2] AVAILABILITY PERFORMANCE RELIABILITY Anywhere, Anytime, Any Device Smartphone  Subscrip`ons   in  2014  [1] 2.1B [1]  hPp://www.kpcb.com/blog/2015-­‐internet-­‐trends     [2]  hPp://www.criteo.com/media/1894/criteo-­‐state-­‐of-­‐mobile-­‐commerce-­‐q1-­‐2015-­‐ppt.pdf f K .
  9. 9. 9 Market pulse Finance/Investing [1]  Image  borrowed  from  hPp://www.bloomberg.com/bw/arAcles/2013-­‐06-­‐06/how-­‐the-­‐robots-­‐lost-­‐high-­‐frequency-­‐tradings-­‐rise-­‐and-­‐fall   [2]  hPp://arAcles.economicAmes.indiaAmes.com/2014-­‐12-­‐26/news/57420480_1_ravi-­‐varanasi-­‐mobile-­‐plaeorm-­‐nse 1  minute  bids  and  offers   March  8,  2011 [1] Mobile  trading  on  the  rise  [2]    NSE      48%  increase  in  turnover,  Jan’14  -­‐>  Dec’14    BSE     0.25%  (Jan’14)  -­‐>  0.5%  (Nov’14)  of  total   volume
  10. 10. 10 Entertainment: MMOs Game of War Largest single world concurrent mobile game in the world “Real-­‐`me        Many-­‐to-­‐Many  is        Tomorrow's  Internet”          -­‐  Francois  Orsini  -­‐ Global scale CollaboraAve:  make  alliances Real-time messaging Chat  translaAon  in  mulAple   languages
  11. 11. 11 On  the rise Cybersecurity 2014 Staples Dec’14 JP  Morgan Oct’14 New  York July’14 Michaels Jan’14 PF  Changs June’14 Home  Depot Sept’14 UPS Aug’14 Sony Nov’14 OPM,  Anthem,  UCLA   2015 2015 [1]  hPp://www.mcafee.com/us/resources/reports/rp-­‐economic-­‐impact-­‐cybercrime2.pdf 400 B [1]
  12. 12. 12 Supporting higher volume and speed Hardware Innovations Massively parallel Intel’s “Knights Landing” Xeon Phi - 72 cores [1] High speed Low Power “…   quickly   idenAfy   fraud   detecAon   paPerns   in   financial   transacAons;   healthcare   researchers   could   process   and   analyze   larger   data   sets   in   real   Ame,   acceleraAng   complex   tasks   such   as   geneAc  analysis  and  disease  tracking.”  [3] Intel and Micron’s 3D XPoint Technology 1000x faster than NAND [1]  hPp://www.anandtech.com/show/9436/quick-­‐note-­‐intel-­‐knights-­‐landing-­‐xeon-­‐phi-­‐omnipath-­‐100-­‐isc-­‐2015   [2]  Intel  IDS’15   [3]  hPp://newsroom.intel.com/community/intel_newsroom/blog/2015/07/28/intel-­‐and-­‐micron-­‐produce-­‐breakthrough-­‐memory-­‐technology [2] Q
  13. 13. 13 Hardware support for apps Hardware Innovations [1]  Images  borrowed  from  Julius  Madelblat’s    and  Andy  Vargas,  Rajeev  Nalawadi  and  Shane  Abreu’s  Technology  Insight  at  IDF’15. Image and Touch processing support in Intel’s Skylake [1]
  14. 14. Emerging  Applica`ons Overview
  15. 15. 15 Real time User Experience, Productivity Real-time Video Streams N E W S Drones Robotics I N D U S T R Y   $ 4 0   B   b y   2 0 2 0   [ 3 ] D E L I V E R Y / M O N i T O R I N G   $ 1 . 7 B   f o r   2 0 1 5 [ 1 ] [1]    hPp://www.kpcb.com/blog/2015-­‐internet-­‐trends   [2]  hPp://www.bostondynamics.com/robot_Atlas.html   [3]  hPp://www.marketsandmarkets.com/Market-­‐Reports/Industrial-­‐RoboAcs-­‐Market-­‐643.html [2]
  16. 16. 16 $1.9  T  in  value  by  2020  -­‐  Mfg  (15%),  Health  Care  (15%),  Insurance  (11%)   26  B  -­‐  75  B  units  [2,  3,  4,  5] [1]    Background  image  taken  from  hPps://www.uspsoig.gov/sites/default/files/document-­‐library-­‐files/2015/rarc-­‐wp-­‐15-­‐013.pdf   [2]  hPp://www.gartner.com/newsroom/id/2636073   [3]  hPps://www.abiresearch.com/press/more-­‐than-­‐30-­‐billion-­‐devices-­‐will-­‐wirelessly-­‐conne   [4]  hPp://newsroom.cisco.com/feature-­‐content?type=webcontent&arAcleId=1208342     [5]  hPp://www.businessinsider.com/75-­‐billion-­‐devices-­‐will-­‐be-­‐connected-­‐to-­‐the-­‐internet-­‐by-­‐2020-­‐2013-­‐10   [6]  hPps://www.abiresearch.com/press/ibeaconble-­‐beacon-­‐shipments-­‐to-­‐break-­‐60-­‐million-­‐by/ Improve  operaAonal  efficiencies,  customer  experience,  new  business  modelsY Beacons:  Retailers  and  bank  branches   60M  units  market  by  2019  [6] Smart  buildings:    Reduce  energy  costs,  cut  maintenance  costs   Increase  safety  &  security Large Market Potential Internet of Things
  17. 17. 17 The Future Biostamps [2] Mobile Sensor Network Exponential growth [1] [1]  hPp://opensignal.com/assets/pdf/reports/2015_08_fragmentaAon_report.pdf   [2]  hPp://www.ericsson.com/thinkingahead/networked_society/stories/#/film/mc10-­‐biostamp
  18. 18. 18 Continuous Monitoring Intelligent Health Care Tracking Movements Measure  effect  of  social   influences Google Lens Measure  glucose  level  in   tears Watch/Wristband Smart Textiles Skin  temperature   PerspiraAon Ingestible Sensors MedicaAon  compliance  [1] Heart  funcAon [1]  hPp://www.hhnmag.com/Magazine/2015/Apr/cover-­‐medical-­‐technology ! !
  19. 19. 19 Connected World Internet of Things 30  B  connected  devices  by  2020 Health Care 153  Exabytes  (2013)  -­‐>  2314  Exabytes  (2020) Machine Data 40%  of  digital  universe  by  2020 Connected Vehicles Data  transferred  per  vehicle  per  month   4  MB  -­‐>  5  GB Digital Assistants (Predictive Analytics) $2B  (2012)  -­‐>  $6.5B  (2019)  [1]   Siri/Cortana/Google  Now Augmented/Virtual Reality $150B  by  2020  [2]   Oculus/HoloLens/Magic  Leap Ñ !+ > [1]  hPp://www.siemens.com/innovaAon/en/home/pictures-­‐of-­‐the-­‐future/digitalizaAon-­‐and-­‐so{ware/digital-­‐assistants-­‐trends.html     [2]  hPp://techcrunch.com/2015/04/06/augmented-­‐and-­‐virtual-­‐reality-­‐to-­‐hit-­‐150-­‐billion-­‐by-­‐2020/#.7q0heh:oABw
  20. 20. ANALYTICS What is Real-Time Analytics?
  21. 21. 21 What is Analytics? According to wikipedia DISCOVERY Ability  to  idenAfy  paPerns  in  data   COMMUNICATION Provide  insights  in  a  meaningful  way " "
  22. 22. 22 Types of Analytics " E CUBE ANALYTICS Business  Intelligence PREDICTIVE ANALYTICS StaAsAcs  and  Machine  learning
  23. 23. 23 What is Real-Time Analytics? BATCH high throughput > 1 hour monthly active users relevance for ads adhoc queries NEAR REAL TIME low latency < 1 ms Financial Trading ad impressions count hash tag trends approximate > 1 sec Online Non-Transactional latency sensitive < 500 ms fanout Tweets search for Tweets deterministic workflows Online Transactional It’s contextual
  24. 24. 24 What is Real-Time Analytics?It’s contextual Value&of&Data&to&Decision/Making& Time& Preven8ve/& Predic8ve& Ac8onable& Reac8ve& Historical& Real%& Time& Seconds& Minutes& Hours& Days& Tradi8onal&“Batch”&&&&&&&&&&&&&&& Business&&Intelligence& Informa9on&Half%Life& In&Decision%Making& Months& Time/cri8cal& Decisions& [1]  Courtesy  Michael  Franklin,  BIRTE,  2015.  
  25. 25. 25 Real Time Analytics STREAMING Analyze  data  as  it  is  being   produced INTERACTIVE Store  data  and  provide  results   instantly   when   a   query   is   posed H C
  26. 26. ALGORITHMS Mining Streaming Data
  27. 27. 27 It’s different Key Characteristics APPROXIMATE H I G H   V E L O C I T Y ONE PASS L O W   L A T E N C Y DISTRIBUTED H I G H   V O L U M E
  28. 28. 28 It’s different Key Characteristics FAULT TOLERANCE [1] A V A I L A B I L I T Y SCALE OUT H I G H   P E R F O R M A N C E ROBUST I N C O M P L E T E   D A T A [1]  ByzanAne  failures  are  described  in  the  following  journal  paper:  J.  Driscoll,  Kevin;  Hall,  Brendan;  Sivencrona,  Håkan;  Zumsteg,  Phil  (2003).  "ByzanAne  Fault  Tolerance,  from  Theory  to  Reality"  2788.  pp.  235–248.
  29. 29. 29 Categorization Sampling A/B  TesAng Filtering Set  Membership Correlation Fraud  DetecAon "
  30. 30. 30 Estimating Cardinality Site  audience  analysis Estimating Quantiles Network  analysis Estimating Moments Databases Frequent Elements Trending  hashtags E
  31. 31. 31 Counting Inversions Measure  sortedness  of  data Finding Subsequences Traffic  analysis Path Analysis Web  graph  analysis Clustering Medical  imaging
  32. 32. 32 Data Prediction Financial  trading Anomaly Detection Sensor  networks
  33. 33. 33 Sampling Obtain  a  representaAve  sample  from  a  data  stream    Maintain  dynamic  sample    A  data  stream  is  a  conAnuous  process    Not  known  in  advance  how  many  points  may  elapse  before  an  analyst  may  need  to  use  a  representaAve  sample    Reservoir  sampling  [1]    ProbabilisAc  inserAons  and  deleAons  on  arrival  of  new  stream  points    Probability  of  successive  inserAon  of  new  points  reduces  with  progression  of  the  stream    An  unbiased  sample  contains  a  larger  and  larger  fracAon  of  points  from  the  distant  history  of  the  stream    PracAcal  perspecAve    Data  stream  may  evolve  and  hence,  the  majority  of  the  points  in  the  sample  may  represent  the  stale  history [1]  J.  S.  ViPer.  Random  Sampling  with  a  Reservoir.  ACM  TransacAons  on  MathemaAcal  So{ware,  Vol.  11(1):37–57,  March  1985.
  34. 34. 34 Sampling  Sliding  window  approach  (sample  size  k,  window  width  n)    Sequence-­‐based      Replace  expired  element  with  newly  arrived  element      Disadvantage:  highly  periodic    Chain-­‐sample  approach      Select  element  ith  with  probability  Min(i,n)/n    Select  uniformly  at  random  an  index  from  [i+1,  i+n]  of  the  element                      which  will  replace  the  ith  item    Maintain  k  independent  chain  samples    Timestamp-­‐based      #  elements  in  a  moving  window  may  vary  over  Ame    Priority-­‐sample  approach [1]  B.  Babcock.  Sampling  From  a  Moving  Window  Over  Streaming  Data.  In  Proceedings  of  SODA,  2002. 3 5 1 4 6 2 8 5 2 3 5 4 2 2 5 0 9 8 4 6 7 3 3 5 1 4 6 2 8 5 2 3 5 4 2 2 5 0 9 8 4 6 7 3 3 5 1 4 6 2 8 5 2 3 5 4 2 2 5 0 9 8 4 6 7 3 3 5 1 4 6 2 8 5 2 3 5 4 2 2 5 0 9 8 4 6 7 3
  35. 35. 35 Sampling  Biased  Reservoir  Sampling  [1]    Use  a  temporal  bias  funcAon  -­‐  recent  points  have  higher  probability  of  being  represented  in  the  sample  reservoir    Memory-­‐less  bias  funcAons    Future  probability  of  retaining  a  current  point  in  the  reservoir  is  independent  of  its  past  history  or  arrival  Ame      Probability  of  an  rth  point  belonging  to  the  reservoir  at  the  Ame  t  is  proporAonal  to  the  bias  funcAon          ExponenAal  bias  funcAons  for  rth  data  point  at  Ame  t,                                                                                  where,  r  ≤  t,    λ        [0,  1]  is  the  bias  rate    Maximum  reservoir  requirement  R(t)  is  bounded [1]  C.  C.  Aggarwal.On  Biased  Reservoir  Sampling  in  the  presence  of  Stream  EvoluAon.  in  Proceedings  of  VLDB,  2006.
  36. 36. 36 Sampling General problem  Input:  Tuples  of  n  components    Subset  are  key  components  -­‐  basis  for  sampling    Sample  of  size  a/b    Hash  key  to  b  buckets    Accept  a  tuple  if  hash  value  <  a    Space  constraint    a  <-­‐  a  -­‐  1    Remove  tuples  whose  keys  hash  to  a
  37. 37. 37 Set Membership Filtering Determine,  with  some  false  probability,  if  an  item  in  a  data  stream  has  been  seen  before    Databases  (e.g.,  speed  up  semi-­‐join  operaAons),  Caches,  Routers,  Storage  Systems    Reduce  space  requirement  in  probabilisAc  rouAng  tables    Speedup  longest-­‐prefix  matching  of  IP  addresses    Encode  mulAcast  forwarding  informaAon  in  packets    Summarize  content  to  aid  collaboraAons  in  overlay  and  peer-­‐to-­‐peer  networks    Improve  network  state  management  and  monitoring  
  38. 38. 38 Set Membership Filtering [1]  IllustraAon  borrowed  from  hPp://www.eecs.harvard.edu/~michaelm/postscripts/im2005b.pdf [1] ApplicaAon  to  hyphenaAon  programs   Early  UNIX  spell  checkers
  39. 39. 39 Set Membership Filtering  Natural  generalizaAon  of  hashing      False  posiAves  are  possible    No  false  negaAves    No  deleAons  allowed    For  false  posiAve  rate  ε,  #  hash  funcAons  =  log2(1/ε) where,  n  =  #  elements,  k  =  #  hash  funcAons                            m  =  #  bits  in  the  array
  40. 40. 40 Set Membership Filtering  Minimizing  false  posiAve  rate  ε  w.r.t.  k  [1]    k  =  ln  2  *  (m/n)    ε  =  (1/2)k  ≈  (0.6185)m/n    1.44  *  log2(1/ε)  bits  per  item    Independent  of  item  size  or  #  items    InformaAon-­‐theoreAc  minimum:  log2(1/ε)  bits  per  item    44%  overhead      X  =  #  0  bits   where [1]  A.  Broder  and  M.  Mitzenmacher.  Network  ApplicaAons  of  Bloom  Filters:  A  Survey.  In  Internet  MathemaAcs  Vol.  1,  No.  4,  2005.
  41. 41. 41 Set Membership Filtering DerivaAves    CounAng  Bloom  filters:  Support  deleAon      Bit  -­‐>  small  counter                          Typically,  4  bits  per  counter  suffice    Increment,  Decrement    Blocked  Bloom  filters    d-­‐le{  CounAng  Bloom  filters    QuoAent  filters    Rank-­‐Indexed  Hashing
  42. 42. 42 Set Membership Filtering Cuckoo Filter [1]  Key  Highlights    Add  and  remove  items  dynamically      For  false  posiAve  rate  ε  <  3%,  more  space  efficient  than  Bloom  filter    Higher  performance  than  Bloom  filter  for  many  real  workloads    AsymptoAcally  worse  performance  than  Bloom  filter    Min  fingerprint  size  α  log  (#  entries  in  table)    Overview      Stores  only  a  fingerprint  of  an  item  inserted    Original  key  and  value  bits  of  each  item  not  retrievable      Set  membership  query  for  item  x:  search  hash  table  for  fingerprint  of  x [1]  Fan  et  al.,  Cuckoo  Filter:  PracAcally  BePer  Than  Bloom.  In  Proceedings  of  the  10th  ACM  InternaAonal  on  Conference  on  Emerging  Networking  Experiments  and  Technologies,  2014.
  43. 43. 43 Set Membership Filtering [1]  R.  Pagh  and  F.  Rodler.  Cuckoo  hashing.  Journal  of  Algorithms,  51(2):122-­‐144,  2004.   [2]  IllustraAon  borrowed  from  “Fan  et  al.,  Cuckoo  Filter:  PracAcally  BePer  Than  Bloom.  In  Proceedings  of  the  10th  ACM  InternaAonal  on  Conference  on  Emerging  Networking  Experiments  and  Technologies,  2014.” [2] IllustraAon  of  Cuckoo  hashing  [2] Cuckoo Hashing [1]  High  space  occupancy    PracAcal  implementaAons:  mulAple  items/bucket    Example  uses:  So{ware-­‐based  Ethernet  switches   Cuckoo Filter  Uses  a  mulA-­‐way  associaAve  Cuckoo  hash  table    Employs  parAal-­‐key  cuckoo  hashing    Relocate  exisAng  fingerprints  to  their  alternaAve   locaAons [2]
  44. 44. 44 Set Membership Filtering Cuckoo Filter  ParAal-­‐key  cuckoo  hashing    Fingerprint  hashing  ensures  uniform  distribuAon  of   items  in  the  table    Length  of  fingerprint  <<  Size  of  h1  or  h2    Possible  to  have  mulAple  entries  of  a  fingerprint  in   a  bucket    DeleAon   Item  must  have  been  previously  inserted Comparison
  45. 45. 45 Estimating Cardinality Large  set  of  real-­‐world  applica`ons    Database  systems/Search  engines    #  disAnct  queries    Network  monitoring  applicaAons    Natural  language  processing    #  disAnct  moAfs  in  a  DNA  sequence    #  disAnct  elements  of  RFID/sensor  networks # Distinct Elements
  46. 46. 46 Estimating Cardinality Historical  context    ProbabilisAc  counAng  [Flajolet  and  MarAn,  1983]    LogLog  counAng  [Durand  and  Flajolet,  2003]    HyperLogLog  [Flajolet  et  al.,  2007]    Sliding  HyperLogLog  [Chabchoub  and  Hebrail,  2010]    HyperLogLog  in  PracAce  [Heule  et  al.,  2013]    Self-­‐Organizing  Bitmap  [Chen  and  Cao,  2009]    Discrete  Max-­‐Count  [Ting,  2014]    Sequence  of  sketches  forms  a  Markov  chain  when  h  is  a  strong  universal  hash    EsAmate  cardinality  using  a  marAngale # Distinct Elements N  ≤  109
  47. 47. 47 Estimating Cardinality Hyperloglog    Apply  hash  funcAon  h  to  every  element  in  a  mulAset      Cardinality  of  mulAset  is  2max(ϱ)  where  0ϱ-­‐11  is  the  bit  paPern  observed  at  the  beginning  of  a  hash  value    Above  suffers  with  high  variance    Employ  stochasAc  averaging    ParAAon  input  stream  into  m  sub-­‐streams  Si  using  first  p  bits  of  hash  values  (m  =  2p) # Distinct Elements where
  48. 48. 48 Estimating Cardinality Hyperloglog  in  Prac`ce:  Op`miza`ons    Use  of  64-­‐bit  hash  funcAon      Total  memory  requirement  5  *  2p  -­‐>  6  *  2p,  where  p  is  the  precision    Empirical  bias  correcAon    Uses  empirically  determined  data  for  cardinaliAes  smaller  than  5m  and  uses  the  unmodified  raw  esAmate  otherwise    Sparse  representaAon    For  n≪m,  store  an  integer  obtained  by  concatenaAng  the  bit  paPerns  for  idx  and  ϱ(w)    Use  variable  length  encoding  for  integers  that  uses  variable  number  of  bytes  to  represent  integers    Use  difference  encoding  -­‐  store  the  difference  between  successive  elements    Other  opAmizaAons  [1,  2] # Distinct Elements [1]  hPp://druid.io/blog/2014/02/18/hyperloglog-­‐opAmizaAons-­‐for-­‐real-­‐world-­‐systems.html   [2]  hPp://anArez.com/news/75
  49. 49. 49 Estimating Cardinality Self-­‐Learning  Bitmap  (S-­‐bitmap)  [1]    Achieve  constant  relaAve  esAmaAon  errors  for  unknown  cardinaliAes  in  a  wide  range,  say  from  10s  to  >106    Bitmap  obtained  via  adapAve  sampling  process    Bits  corresponding  to  the  sampled  items  are  set  to  1    Sampling  rates  are  learned  from  #  disAnct  items  already  passed  and  reduced  sequenAally  as  more  bits  are  set  to  1    For  given  input  parameters  Nmax  and  esAmaAon  precision  ε,  size  of  bit  mask    For  r  =  1  -­‐2ε2(1+ε2)-­‐1  and  sampling  probability  pk  =  m  (m+1-­‐k)-­‐1(1+ε2)rk,  where  k  ∈  [1,m]                RelaAve  error  ≣  ε # Distinct Elements [1]  Chen  et  al.  “DisAnct  counAng  with  a  self-­‐learning  bitmap”.  Journal  of  the  American  StaAsAcal  AssociaAon,  106(495):879–890,  2011.
  50. 50. 50 Estimating Quantiles Large  set  of  real-­‐world  applica`ons    Database  applicaAons    Sensor  networks    OperaAons   ProperAes      Provide  tunable  and  explicit  guarantees  on  the  precision  of  approximaAon    Single  pass   Early  work    [Greenwald  and  Khanna,  2001]  -­‐  worst  case  space  requirement      [Arasu  and  Manku,  2004]  -­‐  sliding  window  based  model,  worst  case  space  requirement   Quantiles, Histograms, Icebergs
  51. 51. 51 Estimating Quantiles q-­‐digest  [1]    Groups  values  in  variable  size  buckets  of  almost  equal  weights    Unlike  a  tradiAonal  histogram,  buckets  can  overlap    Key  features    Detailed  informaAon  about  frequent  values  preserved    Less  frequent  values  lumped  into  larger  buckets    Using  message  of  size  m,  answer  within  an  error  of      Except  root  and  leaf  nodes,  a  node  v  ∈  q-­‐digest  iff Quantiles, Histograms, Icebergs [1]  Shrivastava  et  al.,  Medians  and  Beyond:  New  AggregaAon  Techniques  for  Sensor  Networks.  In  Proceedings  of  SenSys,  2004. Max  signal   value #  Elements Compression   Factor Complete  binary  tree
  52. 52. 52 Estimating Quantiles q-­‐digest    Building  a  q-­‐digest    q-­‐digests  can  be  constructed  in  a  distributed  fashion    Merge  q-­‐digests Quantiles, Histograms, Icebergs
  53. 53. Applica`ons    Track  bandwidth  hogs    Determine  popular  tourist  desAnaAons    Itemset  mining    Entropy  esAmaAon      Compressed  sensing      Search  log  mining    Network  data  analysis    DBMS  opAmizaAon   53 Frequent Elements A core streaming problem
  54. 54. Count-­‐min  Sketch  [1]    A  two-­‐dimensional  array  counts  with  w  columns  and  d  rows    Each  entry  of  the  array  is  iniAally  zero    d  hash  funcAons  are  chosen  uniformly  at  random  from  a  pairwise  independent  family    Update    For  a  new  element  i,  for  each  row  j  and  k  =  hj(i),  increment  the  kth  column  by  one    Point  query                                                                                                          where,  sketch  is  the  table    Parameters 54 Frequent Elements A core streaming problem [1]  Cormode,  Graham;  S.  Muthukrishnan  (2005).  "An  Improved  Data  Stream  Summary:  The  Count-­‐Min  Sketch  and  its  ApplicaAons".  J.  Algorithms  55:  29–38. ),( δε }1{}1{:,,1 wnhh d ……… → ! ! " # # $ = ε e w ! ! " # # $ = δ 1 lnd sketch
  55. 55. Variants  of  Count-­‐min  Sketch  [1]    Count-­‐Min  sketch  with  conservaAve  update  (CU  sketch)    Update  an  item  with  frequency  c    Avoid  unnecessary  updaAng  of  counter  values  =>  Reduce  over-­‐esAmaAon  error    Prone  to  over-­‐esAmaAon  error  on  low-­‐frequency  items      Lossy  ConservaAve  Update  (LCU)  -­‐  SWS    Divide  stream  into  windows    At  window  boundaries,  ∀  1  ≤  i  ≤  w,  1  ≤  j  ≤  d,  decrement  sketch[i,j]  if  0  <  sketch[i,j]  ≤   55 Frequent Elements A core streaming problem [1]  Cormode,  G.  2009.  Encyclopedia  entry  on  ’Count-­‐MinSketch’.  In  Encyclopedia  of  Database  Systems.  Springer.,  511–516.
  56. 56. 56 Anomaly Detection Large  set  of  real-­‐world  applica`ons    Social  media:  Trending  analysis    Fraud  detecAon:  Insurance,  E-­‐commerce,  MarkeAng    Network  intrusion  detecAon    Health  care    Sensor  networks    Anomalous  state  detecAon  (e.g.,  wind  turbines)    OperaAons    Metric  space:  System,  ApplicaAon,  Data  Center      PotenAally  impact  performance,  availability,  reliability Researched over > 50 yrs
  57. 57. 57 Anomaly Detection Anomaly  is  contextual    Manufacturing      StaAsAcs    Econometrics,  Financial  engineering    Signal  processing    Control  systems,  Autonomous  systems  -­‐  fault  detecAon  [1]    Networking    ComputaAonal  biology  (e.g.,  microarray  analysis)    Computer  vision Researched over > 50 yrs [1]  A.  S.  Willsky,  “A  survey  of  design  methods  for  failure  detecAon  systems,”  AutomaAca,  vol.  12,  pp.  601–611,  1976.
  58. 58. 58 Anomaly Detection Characteriza`on    Magnitude    Width    Frequency    DirecAon   Flavors    Global    Local Researched over > 50 yrs Global Local
  59. 59. 59 Anomaly Detection Tradi`onal  Approaches    Rule  based:  μ  ±  σ    Manufacturing,  StaAsAcal  Process  Control  [1]      Moving  averages    SMA    EWMA    PEWMA    AssumpAon:  Normal  distribuAon    Mostly  does  not  hold  in  real  life Researched over > 50 yrs [1]  W.  A.  Shewhart.  Economic  Quality  Control  of  Manufactured  Product,  The  Bell  Labs  Technical  Journal,  9(2):364-­‐389,  1930. [1]
  60. 60. 60 Anomaly Detection In  Prac`ce    Robustness    μ  and  σ  are  not  robust  in  presence  of  anomalies    Use  median  and  MAD  (Median  Absolute  DeviaAon)      Seasonality    Trend    MulA-­‐modal  distribuAon    Time  series  decomposiAon    AnomalyDetecAon  R  package  [1]   Researched over > 50 yrs [1]  hPps://github.com/twiPer/AnomalyDetecAon
  61. 61. Marrying  Time  Series  Decomposi`on  and  Robust  Sta`s`cs   61 Anomaly Detection Researched over > 50 yrs Trend Smoothing Distortion Creates “Phantom” Anomalies Median is Free from Distortion
  62. 62. 62 Anomaly Detection Real-­‐Time    Challenges    AdapAve  learning    Automated  modeling    Marrying  theory  with  contextual  relevance    OperaAons    Large  set  of  different  services  in  a  technology  stack      Different  stacks  use  different  services    Promising  products  such  as  Opsclarity Researched over > 50 yrs
  63. 63. 63 Anomaly Detection Researched over > 50 yrs Anomalies  in  opera`onal  data:  Challenges Contextual Application Topology Map Hierarchical Datacenter ! Applications ! Services ! Hosts •  Automatically discover Developer / Architect’s view of the application - for the Operations team -  Framework for system config and context •  Real-time, streaming architecture -  Keeps up with today’s elastic infrastructure •  Scale to 1000s of hosts, 100s of (micro) services •  Present evolution of system state over time -  DVR-like replay of health, system changes, failures Evolving Needs of Modern Operations
  64. 64. 64 Anomaly Detection Researched over > 50 yrs Anomalies  in  opera`onal  data:  Challenges    AutomaAcally  learn  base-­‐lines  for  metrics    Data  variety  requires  advanced  staAsAcal  approaches    Detect  issues  earlier,  proacAve  alerAng Example: Detecting Disk Full Issues Early
  65. 65. SYSTEMS Overview "
  66. 66. 66 The Key Aspects Requirements of Stream Processing In-stream Handle imperfections Predictable Performance Process  data  as  it  is   passes  by Delayed,  missing  and   out-­‐of-­‐order  data and  Repeatable and  Scalability I 8  Requirements  of  Stream  Processing,  Mike  Stonebraker  et.  al,  SIGMOD  Record  2005
  67. 67. 67 The Key Aspects Requirements of Stream Processing High level languages Integrate stored and streaming data Data safety and availability Process and respond SQL  or  DSL for  comparing  present   with  the  past and  Repeatable ApplicaAon  should  keep   at  high  volumes 8  Requirements  of  Stream  Processing,  Mike  Stonebraker  et.  al,  SIGMOD  Record  2005 # # $ %
  68. 68. 68 Window Processing Stream Processing T.  Akidau  et  al.,  The  Dataflow  Model:  A  PracAcal  Approach  to  Balancing  Correctness,  Latency,  and  Cost  in  Massive-­‐Scale,  Unbounded,  Out-­‐of-­‐Order  Data  Processing,  In  VLDB,  2015. & # $
  69. 69. 69 Three Generations First Generation Extensions  to  exisAng  database  engines  or  simplisAc  engines   Dedicated  to  specific  applicaAons  or  use  cases Second Generation Enhanced  methods  regarding  language  expressiveness   Distributed  processing,  load  balancing  and  fault  tolerance Third Generation Massive  parallelizaAon  for  processing  large  data  sets   Dedicated  towards  cloud  compuAng , % hPp://www.slideshare.net/zbigniew.jerzak/cloudbased-­‐data-­‐stream-­‐processing
  70. 70. 1st generation - Active Database Systems SYSTEMS "
  71. 71. 71 Late 1980s Late 1990s 1st Generation Systems HiPAC [Dayal  et  al.,  1988] Starbust [Widom/Finkelstein  et  al.,  1990] !
  72. 72. 72 Postgres [Stonebraker/Kemnitz  et  al.,  1991] ODE [Gehani/Jagadish  et  al.,  1991]
  73. 73. 73 Notable features 1st Generation Systems Early: Active DBs, ECA rules, triggers, publish-subscribe Event-Condition-Action ) ' Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs Event   Occurrences Triggered   Rules Evaluated   Rules Selected    Rules Event   Source Signaling Triggering EvaluaAon SchedulingExecuAon G Systems - HiPAC, Starbust, Postgres, ODE “AcAve  Database  Systems”,  Paton  and  Diaz,  ACM  CompuAng  Surveys,  1999
  74. 74. 74 Notable features 1st Generation Applications Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs Actuation (also IoT?) Finance Enforcing database integrity constraints Monitoring the physical world (IoT?) Supply chain News and update dissemination ( #) # Battlefield awarenessHealth monitoring - d
  75. 75. 75 Issues 1st Generation Systems Rules were (are) hard to program or understand Smart engineering of traditional approaches can get you close enough?! Little commercial activity Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs #
  76. 76. 2nd generation - Streaming Database Systems SYSTEMS "
  77. 77. 77 Early 2000s Late 2000s 2nd Generation Systems Niagara CQ [Jianjun  Chun  et  al.,  2000] Telegraph, Telegraph CQ [Hellerstein  et  al.,  2000]   [Chandrasekaran  et  al.,  2003] !
  78. 78. 78 STREAM [Arasu  et  al.,  2003] Aurora [Abadi  et  al.,  2003] Borealis [Abadi  et  al.,  2005] ✉ (
  79. 79. 79 Cayuga [Demeres  et  al.,  2007] MCOPE [Park  et  al.,  2009]
  80. 80. Repeatedly apply generic SQL to the results of window operators 80 The basic idea Stream Query Processing Support full SQL language and eco system A table is a set of records and a stream is an unbounded sequence of records SQL g Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs Each window outputs a set of records Window operators convert streams to tablesÄ Rstream  semanAcs  in  CQL,  Arvind  Arasu  et  al.  VLDB  Journal  2006 Streams Tables Window  Operators 3 # $
  81. 81. 81 Telegraph CQ Data  stream  query  processor Con`nuous  and  adap`ve     query  processing Built  by  modifying  PostgreSQL 01 02 03 Developed at University of California, Berkeley Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs
  82. 82. 82 Niagara CQ Incremental    group  opAmizaAon  strategy   Incremental  evaluaAon  of  conAnuous  queries A   distributed   database   system   for   conAnuous   queries   using   a   query   language   like   XML-­‐QL   for   changing   data   sets Query  Grouping Allows  for  sharing  common  parts  of   two  or  more  queries Caching For  performance Push/Pull  data  inges`on for  detected  changes  in  data Change  based  and  Timer  CQ ConAnuous  queries  to  trigger  on  data   changes  and  regular  Amed  based 01 02 03 04 Developed at UW-Madison
  83. 83. 83 Niagara CQ Query grouping and sharing quotes.xml Select   Symbol  =  INTC Trigger  AcAon  1 quotes.xml Select   Symbol  =  MSFT Trigger  AcAon  2 Select Constant   Table   INTC/MSFT quotes.xml Split Trigger  AcAon  1 Trigger  AcAon  2
  84. 84. 84 Borealis Load  aware  distribuAon   Fine  grained  high  availability   Load  shredding  mechanisms A   low   latency   stream   processing   engine   with   a   focus   on   fault   tolerance   and   distribuAon Distributed  stream  engine Allows  for  sharing  common  parts  of   two  or  more  queries Dynamic  query  modifica`on For  performance Dynamic  system  op`miza`on for  detected  changes  in  data Dynamic  revision  of  results ConAnuous  queries  to  trigger  on  data   changes  and  regular  Amed  based 01 02 03 04 Developed at MIT, Brown and Brandeis
  85. 85. 85 Summary 2nd Generation Systems Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs Can reuse many of relational operators Historical comparison becomes a join of a stream and its history table Views on streams can be created Streams can be processed using relational operators Can leverage an RDMS system Stream and stream results can be stored in tables for later querying + (, g$ G
  86. 86. 86 Issues 2nd Generation Systems Despite significant commercial activity, no real breakout No standardization and comprehensive benchmarks 6 % Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs & Value proposition for learning new concepts was not clear
  87. 87. SYSTEMS 3rd generation "
  88. 88. 88 The last decade Streaming Platforms S4 Yahoo! Flink Apache Storm TwiPer Spark Databricks Samza LinkedIn Heron TwiPer MillWheel Google Pulsar eBay %% S-Store ISTC,  Intel,  MIT,  Brown,  CMU,  Portland  State S Trill Microso{ T
  89. 89. 89 Earliest distributed stream system Apache S4 Scalable Throughput  is  linear  as  addiAonal   nodes  are  added Cluster management Hides  managements  using  a  layer   in  ZooKeeper Decentralized All  nodes  are  symmetric  and  no   centralized  service Extensible Building  blocks  of  plaeorm  can  be  replaced   by  custom  implementaAons Fault tolerance Standby  servers  take  over  when  a     node  fails $ (, g# G Proven Deployed  in  Yahoo  processing  thousands  of   search  queries  per  second
  90. 90. 90 Twitter Storm Guaranteed Message Passing Horizontal Scalability Robust Fault Tolerance Concise Code-Focus on Logic b Ñ /
  91. 91. 91 Storm Terminology Topology Directed  acyclic  graph     verAces  =  computaAon,  and     edges  =  streams  of  data  tuples Spouts Sources  of  data  tuples  for  the  topology   Examples  -­‐  Ka•a/Kestrel/MySQL/Postgres Bolts Process  incoming  tuples,  and  emit  outgoing  tuples   Examples  -­‐  filtering/aggregaAon/join/any  funcAon , %
  92. 92. 92 Storm Topology % % % % % Spout 1 Spout 2 Bolt 1 Bolt 2 Bolt 3 Bolt 4 Bolt 5
  93. 93. 93 Tweet Word Count Topology % % Tweet Spout Parse Tweet Bolt Word Count Bolt Live stream of Tweets #worldcup : 1M soccer: 400K ….
  94. 94. 94 Tweet Word Count Topology % % Tweet Spout Parse Tweet Bolt Word Count Bolt When  a  parse  tweet  bolt  task  emits  a  tuple   which  word  count  bolt  task  should  it  send  to? % %% %% %% %
  95. 95. 95 Storm Groupings 01 02 03 04 Shuffle Grouping Random distribution of tuples Fields Grouping Group tuples by a field or multiple fields All Grouping Replicates tuples to all tasks Global Grouping Send the entire stream to one task / . - ,
  96. 96. 96 Tweet Word Count Topology % % Tweet Spout Parse Tweet Bolt Word Count Bolt % %% %% %% % Shuffle Grouping Fields Grouping
  97. 97. 97 Storm Architecture Nimbus ZK Cluster Supervisor W1 W2 W3 W4 Supervisor W1 W2 W3 W4 Topology Submission Assignment Maps Sync Code Slave Node Slave Node Master Node
  98. 98. 98 Storm Worker TASK TASKTASK TASK EXECUTOR TASKTASK EXECUTORTASK TASK EXECUTORTASK
  99. 99. 99 Data Flow in Storm Workers Global  Receive   Thread Global  Send   Thread In  Queue User  Logic     Thread Out  Queue Send   Thread Outgoing   Message  Buffer
  100. 100. 100 Storm Metrics Support and trouble shooting Continuous performance Cluster availability# g G
  101. 101. 101 Collecting Topology Metrics % % Tweet Spout Parse Tweet Bolt Word Count Bolt % Scribe Metrics Bolt
  102. 102. 102 Topology Dashboard
  103. 103. 103 Overloaded Zookeeper S1 S2 S3W W W STORM zk SERVICES
  104. 104. 104 Overloaded Zookeeper S1 S2 S3W W W STORM zk SERVICES zk
  105. 105. 105 Overloaded Zookeeper zk S1 S2 S3W W W STORM zk SERVICES
  106. 106. 106 Analyzing Zookeeper Traffic Overloaded Zookeeper 67 % 33 % Offset/ParAAon  is   wriPen  every  2   secs Kafka Spout Workers  write   heart  beats  every   3  secs Storm Runtime
  107. 107. W 107 Heartbeat Daemons Overloaded Zookeeper zk S1 S2 S3W W STORM zk SERVICES Heartbeat     Cluster   Key  Value   Store
  108. 108. 108 Some experiments Storm Overheads Read  from  Ka•a  cluster  and  serialize  in  a  loop   Sustain  input  rates  of  300K  msgs/sec  from  Ka•a  topic Java program No  acks  to  achieve  at  least  once  semanAcs   Storm  processes  were  co-­‐located    using  isolaAon  scheduler 1-stage topology Enable  acks  for  at  least  once  semanAcs 1-stage topology with acks
  109. 109. 109 Performance comparison Storm Overheads AverageCPUUtilization 0% 20% 40% 60% 80% MachinesUsed 0 1 2 3 JAVA 1-STAGE 1-STAGE-ACK Machines Avg. CPU 77% 58.2%58.3% 3 11
  110. 110. 110 Storm Deployment shared pool storm cluster
  111. 111. 111 Storm Deployment shared pool storm cluster joe’s topology isolated pools
  112. 112. 112 Storm Deployment shared pool storm cluster joe’s topology isolated pools jane’s topology
  113. 113. 113 Storm Deployment shared pool storm cluster joe’s topology isolated pools jane’s topology dave’s topology
  114. 114. 114 MillWheel DAG Processing Streams   ComputaAons . Cloud DataFlow  Uses  MillWheel (From Google Not  OpenSource ⛔ Exactly Once Checkpoint  User  State 4
  115. 115. 115 MillWheel Computations Arbitrary  User  Logic   Per  Key  OperaAon Persistent State Key/Value  API   Backed  by  BigTable Streams IdenAfied  By  Names   Unbounded Keys Per  Key  OperaAon  Serial   Different  Keys  Parallel Core Concepts L f ⚿ t
  116. 116. 116 MillWheel Caught up Time Defined  per  computaAon Discard Late Data ~0.001%  at  Google Seeded by Injectors Input  Sources Monotonic Makes  life  easy  for  users Low Watermark: The Concept of Time Ê 4 6 u
  117. 117. 117 MillWheel Checkpoint Same  Ame  as  User  State DoubleCount No  Dedup Seeded by Injectors Input  Sources No checkpoint Simpler  API Strong And Week: Productions ' 4 ( q
  118. 118. 118 MillWheel Key/Value Abstractions ComputaAons Persistance Layer BigTable Idempotent No  Side  Effects Batched Efficient Computation State: Exactly Once Semantics ó a t $
  119. 119. 119 PubSub weds Processing Exactly  Once  Processing 4 Tightly  Integrated  with  Kasaq Open  Sourced  by  LinkedIn K Durability  via  YarnV
  120. 120. 120 Samza ParAAon  1ParAAon  0 ParAAon  2 Streams: Partitioned
  121. 121. 121 Samza ParAAon  0 Task Task: Work on a single partition
  122. 122. 122 Samza Stream  A Stream  B Task  1 Task  2 Task  3 Stream  C Job  1 Job: Collection of Tasks
  123. 123. 123 Samza Samza State API key  value  store State As a Stream persist  on  Ka•a ó f Stateful Tasks: Exactly Once Semantics
  124. 124. 124 Samza Kafka based Streams Persistence t Simple API Single  Node  Job 2 Stateful Exactly  Once 4 Yarn Friendly Durability K Tight Coupling: Queue and Processing
  125. 125. 125 One Size Fits All Apache Flink General  Purpose  Analy`cs  Engine Open  Source  and  Community  Driven Works  well  with  Hadoop  Ecosystem K Came  out  of  Stratosphere n
  126. 126. 126 Apache Flink Fast RunTime Complex  DAG  Operators   Streamed  Data  to  Op Iterative Algorithms Much  Faster  In-­‐ Memory  OperaAons Intuitive APIs Java/Scala/Python       Concise Query Coming  from  OLTP   World % ! 2 b Ambitious Goal: One Size Fits All
  127. 127. 127 Apache Flink Data Streamed between  operators . Master Submission  and   Scheduling L Workers Do  Actual  Work K Distributed Runtime: Scale
  128. 128. 128 Apache Flink Stack: Co-Exist with Hadoop
  129. 129. 129 One system to replace them all!  General  purpose  Compute  Engine Open  Source/Big  Community K MapReduce,  Streaming,  SQL,  …! Integrates  well  with  Hadoop  Ecosystem(
  130. 130. 130 Lots Huge  CollecAon  with   Lineage  info Resilient Lost  DataSets  are  re-­‐ computed Distributed Across  the  cluster Core Concept: Lots of RDDS t ( )DataSet Input  Data  divided  into   Batches $ Streaming
  131. 131. 131 W1 W2 W1 W3 W2 W1 W2 W1 W3 W1  W4  W3   W1  W5  W4 W6  W2  W7   W4  W7  W3 W5  W8  W2   W1  W4  W8 FlatMap Map reduceByKey W1:1 W2:1 W1:1 W4:1 W1:1 W5:1 W1:3 W2:4 W3:1 W4:1 W5:4 W6:2 RDDs In Action:- WordCount Streaming
  132. 132. 132 Scala: Functional and Concise Streaming
  133. 133. 133 Streaming: Fits Naturally              Spark        Streaming              Spark              Engine W3 W2 W4 W1W2W1 DStream W2 W4 W1W3W2W1 Streaming
  134. 134. 134 T0  to  T1 T1  to  T2 T2  to  T3 T0  to  T1 T1  to  T2 T2  to  T3 lines words flatMap Series of RDDs 5 Window FunctionsA Can Create other Dstreamsq Streaming: With Dstreams Streaming
  135. 135. 135 DStream: Operators Regular Spark Operators map,  flatMap,  filter,  … Y Transform RDD  -­‐>  RDD $ Window Operators countByWindow,   reduceByWindow A Join join  mulAple   Dstreams , Streaming
  136. 136. 136 Basic Sources HDFS,  S3,  … É Reliability ack  vs  noAck  sources VCustom Implement  Interface J ^ Advanced Ka•a,  TwiPerUAls u Input DStreams: Sources of Data Streaming
  137. 137. 137 Exaclty Once Confident  about  results 4 Ecosystem Hadoop,Yarn,  Ka•a,  … K Scalable RDDs  as  scale  unit Single System Batch  +  Streaming v Basic Premise: One Size Fits All Streaming
  138. 138. 138 Annota`on  plugin  framework  to  extend  SQL Stream Processing: With SQL Processing  logic  in  SQL % Clustering  with  elas`c  scaling No  down`me  during  upgrades(
  139. 139. 139 Channels Key/Value  API É Processor SQL,  Custom J Core Concept: CEP Cell Inbound   Channel Outbound   Channel Processor CEP  Cell
  140. 140. 140 Example Pipeline: Stitching Cells
  141. 141. 141 Messaging Models Used  for  low  latency.   Producer  pushes  data  to  consumer.   Write  to  Kakfla  if  consumer  down  or   unable  to  keep  up  for  replay  later Push Atmost once / Producer  writes  events  to  Ka•a   Consumer  consumes  Ka•a   Storing  to  Ka•a  allows  for  replay   Pull Atleast once /
  142. 142. 142 Deployment Architecture Events are partitioned All  events  with  the  same  key  are  routed  to  the   same  cell   Scaling More  cells  are  added  to  the  pipeline  for  scaling   Pulsar   automaAcally   detects   new   cells   and   rebalances  traffic
  143. 143. 143 SQL:  Event filtering and routing
  144. 144. 144 SQL:  Top N items
  145. 145. 145 Better Storm Twitter Heron Container  Based  Architecture Separate  Monitoring  and  Scheduling - Simplified  Execu`on  Model 2 Much  Be@er  Performance%
  146. 146. 146 Storm: Issues Heron Poor Performance Queue  ContenAons   MulAple  Languages &Lack of BackPressure Unpredictable  Drops ! Complex Execution Env Hard  to  tune ! SPOF Overloaded  Nimbus "
  147. 147. 147 Heron Batching of tuples AmorAzing  the  cost  of  transferring  tuples $ Task isolation Ease  of  debug-­‐ability/isolaAon/profiling (Fully API compatible with Storm Directed  acyclic  graph      Topologies,  Spouts  and  Bolts , Support for back pressure Topologies  should  self  adjusAng gUse of main stream languages C++,  Java  and  Python # Efficiency Reduce resource consumption G Design: Goals
  148. 148. 148 Heron Topology 1 Topology Submission Scheduler Topology 2 Topology N Architecture: High Level
  149. 149. 149 Heron Topology Master ZK Cluster Stream Manager I1 I2 I3 I4 Stream Manager I1 I2 I3 I4 Logical Plan, Physical Plan and Execution State Sync Physical Plan CONTAINER CONTAINER Metrics Manager Metrics Manager Architecture: Topology
  150. 150. 150 Heron Gateway for metrics G Assigns role# Monitoring of containers g Topology Master
  151. 151. 151 Heron Topology Master ZK Cluster Logical Plan, Physical Plan and Execution State Prevent  mul`ple  TM  becoming     masters Allows  other  process  to  discover  TM 01 02 Topology Master
  152. 152. 152 Heron % % S1 B2 B3 % B4 Stream Manager: BackPressure
  153. 153. 153 Stream Manager S1 B2 B3 Stream Manager Stream Manager Stream Manager Stream Manager S1 B2 B3 B4 S1 B2 B3 S1 B2 B3 B4 B4 Stream Manager: BackPressure
  154. 154. 154 Heron Slows upstream and downstream instances S1 B2 B3 Stream Manager Stream Manager Stream Manager Stream Manager S1 B2 B3 B4 S1 B2 B3 S1 B2 B3 B4 B4 Stream Manager: TCP BackPressure
  155. 155. S1 S1 S1S1S1 S1 S1S1 155 Heron B2 B3 Stream Manager Stream Manager Stream Manager Stream Manager B2 B3 B4 B2 B3 B2 B3 B4 B4 Stream Manager: Spout BackPressure
  156. 156. 156 Heron Exposes Storm and Heron APIAPI Collects several metricsG Runs only one task (spout/bolt) g Instance: Worker Bee
  157. 157. 157 Heron Stream Manager Metrics Manager Gateway Thread Task Execution Thread data-in queue data-out queue metrics-out queue Instance: Worker Bee
  158. 158. 158 Heron Topology 1 Topology 2 Topology N Heron Tracker Heron VIZ Heron Web ZK Cluster Aurora Services Observability Deployment
  159. 159. 159 Heron Sample Topologies
  160. 160. 160 Heron Visualization
  161. 161. 161 Heron COMPONENTS EXPT #1 EXPT #2 EXPT #3 EXPT #4 Spout 25 100 200 300 Bolt 25 100 200 300 # Heron containers 25 100 200 300 # Storm workers 25 100 200 300 Performance: Settings
  162. 162. 162 Heron milliontuples/min 0 350 700 1050 1400 Spout Parallelism 25 100 200 500 Storm Heron latency(ms) 0 625 1250 1875 2500 Spout Parallelism 25 100 200 500 Storm Heron Throughput Latency 10 -14x 5 -15x Performance: Atleast Once
  163. 163. 163 Heron #coresused 0 625 1250 1875 2500 Spout Parallelism 25 100 200 500 Storm Heron 2 -3x Performance: CPU Usage
  164. 164. 164 Heron Throughput CPU usage milliontuples/min 0 1250 2500 3750 5000 Spout Parallelism 25 100 200 500 Storm Heron #coresused 0 625 1250 1875 2500 Spout Parallelism 25 100 200 500 Storm Heron Performance: Atmost Once
  165. 165. 165 Heron Performance % % Client Event Spout Distributor Bolt User Count Bolt % Aggregator Bolt Shuffle Grouping Fields Grouping Fields Grouping Performance: RTAC Topology
  166. 166. 166 Heron #coresused 0 100 200 300 400 Storm Heron latency(ms) 0 17.5 35 52.5 70 Storm Heron Latency CPU usage Performance: RTAC Atleast Once
  167. 167. 167 Heron #coresused 0 62.5 125 187.5 250 Storm Heron CPU usage Performance: RTAC Atmost Once
  168. 168. 168 Issues 3rd Generation Systems Bit early to tell Still no standardization and too many systems 6 % Slide  from  Mike  Franklin,  VLDB  2015  BIRTE  Talk  on  Real  Time  AnalyAcs
  169. 169. 169 Growing set Commercial Platforms 01 02 03 04 08 07 06 05 Infosphere Vibe Apama Event   Processor Data  Torrent Vitria  OI Blaze StreamBase
  170. 170. Prac`cal  Deployments "
  171. 171. 171 Combining batch and real time Lambda Architecture New  Data Client
  172. 172. 172 Lambda Architecture - The Good Message   Broker CollecAon  Pipeline Lambda  Architecture   AnalyAcs  Pipeline Results
  173. 173. 173 Lambda Architecture - The Bad Have to fix everything (may be twice)! How much Duct Tape required? Have to write everything twice! Subtle differences in semantics What about Graphs, ML, SQL, etc? $ *, 7#
  174. 174. 174 Summingbird Summingbird  Program Map  Reduce  Job HDFS Message  broker Storm/Heron  Topology Online  key  value  result   store Batch  key  value  result   store Client
  175. 175. 175 Near real-time processing SQL-on-Hadoop Com m ercial Commercial Apache Commercial Cloudera Hortonworks Pivotal MammothDB
  176. 176. Auto scaling the system in the presence of unpredictability 176 Technology Challenges The Road Ahead Auto tuning of real time analytics jobs/queries Exploiting faster networks for efficiently moving data Ä Ü J
  177. 177. Real-time personalization 177 Applications The Road Ahead Preferences,  Ame,  locaAon  and  social Wearable computing Screen  size  fragmentaAon Analytics: Image, Video, Touch PaPern  RecogniAon,  Anomaly  DetecAon +
  178. 178. 178 WHAT WHY WHERE WHEN WHO HOW Any Question ???
  179. 179. 179 @arun_kejariwal, @sanjeevrk, @karthikz Get in Touch
  180. 180. THANKS  FOR  ATTENDING  !!!

×