Mining Big Data Streams Better Algorithms or Faster Systems? 
 
 Gianmarco De Francisci Morales
 [email protected]
 @gdfm7 

Vision Algorithms & Systems Distributed stream mining platform Development and collaboration framework
 for researchers Library of state-of-the-art algorithms
 for practitioners 2 

Agenda SAMOA
 (Scalable Advanced Massive Online Analysis) VHT
 (Vertical Hoeffding Tree) PKG
 (Partial Key Grouping) 3 System Algorithm API 

Visiting Scientist 
 @Aalto DMG Scientist @Yahoo Labs PPMC @ Apache SAMOA Committer @ Apache Pig Contributor for Hadoop, 
 Giraph, Storm, S4, Grafos.ml 4 

What do I work on? Systems Distributed Mining News Streaming Grid Admin —2008 —2009 —2010 —2011 —2012 —2013 -—2014 -—2015 • IMT Lucca • M.Eng • Y!R Barcelona • PhD 5 PhD Student Postdoc Scientist 

“Panta rhei”
 (everything flows) -Heraclitus 6 

Importance$of$O •  As$spam$trends$change retrain$the$model$with Importance Example: spam detection in comments on Yahoo News Trends change in time Need to retrain model with new data 7 

Stream Batch data is a snapshot of streaming data 8 

Challenges Operational Need to rerun the pipeline and redeploy the model when new data arrives Paradigmatic New data lies in storage without generating new value until new model is retrained 9 Gather Clean Model Deploy 

Present of big data Too big to handle 10 

Future of big data Drinking from a firehose 11 

Evolution of SPEs 12 —2003 —2004 —2005 —2006 —2008 —2010 —2011 —2013 Aurora STREAM Borealis SPC SPADE Storm S4 1st generation 2nd generation 3rd generation Abadi et al., “Aurora: a new model and architecture for data stream management,” VLDB Journal, 2003 Arasu et al., “STREAM: The Stanford Data Stream Management System,” Stanford InfoLab, 2004. Abadi et al., “The Design of the Borealis Stream Processing Engine,” in CIDR ’05 Amini et al., “SPC: A Distributed, Scalable Platform for Data Mining,” in DMSSP ’06 Gedik et al., “SPADE: The System S Declarative Stream Processing Engine,” in SIGMOD ’08 Neumeyer et al., “S4: Distributed Stream Computing Platform,” in ICDMW ’10 http://storm-project.net Samza http://samza.incubator.apache.org 

Actor Model 13 PE PE Input Stream PEI PEI PEI PEI PEI Output Stream Event routing 

Paradigm Shift 14 Gather Clean Model Deploy + = 

Apache SAMOA Scalable Advanced Massive Online Analysis
 G. De Francisci Morales, A. Bifet
 JMLR 2015 15 

Taxonomy 16 Data Mining Distributed Batch Hadoop Mahout Stream Storm, S4, Samza SAMOA Non Distributed Batch R, WEKA, … Stream MOA 

What about Mahout? SAMOA = Mahout for streaming But… More than JBoA (just a bunch of algorithms) Provides a common platform Easy to port to new computing engines 17 

Architecture 18 SA SAMOA% 

Status Status 19 https://samoa.incubator.apache.org 

Status Status Parallel algorithms 19 https://samoa.incubator.apache.org 

Status Status Parallel algorithms Classification (Vertical Hoeffding Tree) 19 https://samoa.incubator.apache.org 

Status Status Parallel algorithms Classification (Vertical Hoeffding Tree) Clustering (CluStream) 19 https://samoa.incubator.apache.org 

Status Status Parallel algorithms Classification (Vertical Hoeffding Tree) Clustering (CluStream) Regression (Adaptive Model Rules) 19 https://samoa.incubator.apache.org 

Status Status Parallel algorithms Classification (Vertical Hoeffding Tree) Clustering (CluStream) Regression (Adaptive Model Rules) Execution engines 19 https://samoa.incubator.apache.org 

Is SAMOA useful for you? Only if you need to deal with: Large fast data Evolving process (model updates) What is happening now? Use feedback in real-time Adapt to changes faster 20 

Advantages (operational) Avoid deploy cycle No need to choose update frequency No system downtime No complex backup/update procedures Program once, run everywhere Reuse existing computational infrastructure 21 

Advantages (paradigmatic) Model freshness Immediate data value No stream/batch impedance mismatch 22 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 23 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 24 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 24 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 24 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 25 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 25 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 25 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 26 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 26 

PE PE PEI PEI PEI PEI Groupings Key Grouping 
 (hashing) Shuffle Grouping
 (round-robin) All Grouping
 (broadcast) 26 

VHT Vertical Hoeffding Tree
 A. Murdopo, A. Bifet, G. De Francisci Morales, N. Kourtellis
 (under submission) 27 

Decision Tree Nodes are tests on attributes Branches are possible outcomes Leafs are class assignments
 
 28 Class Instance Attributes Road Tested? Mileage? Age? No Yes High ✅ ❌ Low Old Recent ✅ ❌ Car deal? 

Hoeffding Tree Sample of stream enough for near optimal decision Estimate merit of alternatives from prefix of stream Choose sample size based on statistical principles When to expand a leaf? Let x1 be the most informative attribute,
 x2 the second most informative one Hoeffding bound: split if 29 G ( x1, x2) > ✏ = r R 2 ln(1 / ) 2 n P. Domingos and G. Hulten, “Mining High-Speed Data Streams,” KDD ’00 

Parallel Decision Trees 30 

Parallel Decision Trees Which kind of parallelism? 30 

Parallel Decision Trees Which kind of parallelism? Task 30 

Parallel Decision Trees Which kind of parallelism? Task Data 30 Data Attributes Instances 

Parallel Decision Trees Which kind of parallelism? Task Data Horizontal 30 Data Attributes Instances 

Parallel Decision Trees Which kind of parallelism? Task Data Horizontal Vertical 30 Data Attributes Instances 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates Single attribute tracked in multiple node 31 

Horizontal Parallelism Y. Ben-Haim and E. Tom-Tov, “A Streaming Parallel Decision Tree Algorithm,” JMLR, vol. 11, pp. 849–872, 2010 31 Stats Stats Stats Stream Histograms Model Instances Model Updates Aggregation to compute splits 31 

Hoeffding Tree Profiling 32 Other 6% Split 24% Learn 70% CPU time for training
 100 nominal and 100 numeric attributes 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Stats Stats Stats Stream Model Attributes Splits 

Vertical Parallelism 33 Single attribute tracked in single node Stats Stats Stats Stream Model Attributes Splits 

Advantages of Vertical High number of attributes => high level of parallelism
 (e.g., documents) Vs task parallelism Parallelism observed immediately Vs horizontal parallelism Reduced memory usage (no model replication) Parallelized split computation 34 

Vertical Hoeffding Tree 35 Control Split Result Source (n) Model (n) Stats (n) Evaluator (1) Instance Stream Shuffle Grouping Key Grouping All Grouping 

Accuracy 36 No. Leaf Nodes VHT2 – tree-100 30 Very close and very high accuracy 

Performance 37 35 0 50 100 150 200 250 MHT VHT2-par-3 Execution Time (seconds) Classifier Profiling Results for text-10000 with 100000 instances t_calc t_comm t_serial Throughput VHT2-par-3: 2631 inst/sec MHT : 507 inst/sec 

PKG Partial Key Grouping
 M. A. Uddin Nasir, G. De Francisci Morales, D. Garcia-Soriano, N. Kourtellis, 
 M. Serafini, “The Power of Both Choices: Practical Load Balancing for Distributed Stream Processing Engines”, ICDE 2015 38 

10-14 10-12 10-10 10-8 10-6 10-4 10-2 100 100 101 102 103 104 105 106 107 108 CCDF key frequency words in tweets wikipedia links Systems Challenges Skewed key distribution 39 

Key Grouping and Skew 40 Source Source Worker Worker Worker Stream 

Problem Statement Input stream of messages Load of worker Imbalance of the system Goal: partitioning function that minimizes imbalance 41 m = ht, k, vi Li(t) = |{h⌧, k, vi : P⌧ (k) = i ^ ⌧  t}| Pt : K ! N i 2 W I ( t ) = max i ( Li( t )) avg i ( Li( t )) , for i 2 W 

Shuffle Grouping 42 Source Source Worker Worker Stream Aggr. Worker 

Existing Stream Partitioning Key Grouping Memory and communication efficient :) Load imbalance :( Shuffle Grouping Load balance :) Additional memory and aggregation phase :( 43 

Solution 1: Rebalancing At regular intervals move keys around workers Issues How often? Which keys to move? Key migration not supported with Storm/Samza API Many large routing tables (consistency and state) Hard to implement 44 

Solution 2: PoTC Balls and bins problem For each ball, pick two bins uniformly at random Put the ball in the least loaded one Issues Consensus and state to remember choice Load information in distributed system 45 

Solution 3: PKG Fully distributed adaptation of PoTC, handles skew Consensus and state to remember choice Key splitting:
 assign each key independently with PoTC Load information in distributed system Local load estimation:
 estimate worker load locally at each source 46 

Power of Both Choices 47 Source Source Worker Worker Stream Aggr. Worker 

Throw m balls with k colors in n bins with d choices Ball = msg, bin = worker, color = key, choice = hash Necessary condition: Imbalance is Chromatic Balls and Bins 48 p1  d n I(m) = ( O m n · ln n ln ln n , if d = 1 O m n , if d 2 

Comparison 49 Stream Grouping Pros Cons Key Grouping Memory efficient Load imbalance Shuffle Grouping Load balance Memory overhead Aggregation O(W) Partial Key Grouping Memory efficient Load balance Aggregation O(1) 

Experimental Design What is the effect of key splitting? How does local estimation compare to a global oracle? How does PKG perform in a real system? Measures: imbalance, throughput, latency, memory Datasets: Twitter, Wikipedia, graphs, synthetic 50 

Effect of Key Splitting 51 Average Imbalance 0% 1% 2% 3% 4% Number of workers 5 10 50 100 PKG Off-Greedy PoTC KG 

Local vs Global 52 10-10 10-9 10-8 10-7 10-6 10-5 10-4 10-3 10-2 10-1 5 10 50 100 Fraction of Imbalance workers TW G L5 5 10 50 100 workers WP L10 5 10 50 100 workers CT L15 5 LN1 Fig. 2: Fraction of average imbalance with respect to total number of messages workers and number of sources. 5 10 50 100 workers W L5 5 10 50 100 workers WP L10 5 10 50 100 workers CT L15 5 LN1 L2 ction of average imbalance with respect to total number of messages fo d number of sources. 100 5 10 50 100 workers WP L10 5 10 50 100 workers CT L15 5 10 50 100 workers LN1 L20 rage imbalance with respect to total number of messages for each datas sources. 5 10 50 100 workers P L10 5 10 50 100 workers CT L15 5 10 50 100 workers LN1 L20 5 LN2 ance with respect to total number of messages for each dataset, for diffe 0 5 10 50 100 workers CT L15 5 10 50 100 workers LN1 L20 5 10 50 100 workers LN2 H ect to total number of messages for each dataset, for different number 10-10 10-9 10-8 10-7 10-6 10-5 10-4 10-3 10-2 10-1 5 10 50 100 workers TW G L5 5 10 50 100 workers WP L10 5 10 50 100 workers CT L15 g. 2: Fraction of average imbalance with respect to total number of mes rkers and number of sources. 10-10 10-9 10-8 10-7 10-6 10-5 10-4 10-3 10-2 10-1 5 10 50 100 Fraction of Imbalance workers TW G L5 5 10 50 100 workers WP L10 5 10 50 100 workers CT L15 Fig. 2: Fraction of average imbalance with respect to total number of m workers and number of sources. 

Throughput vs Memory 53 0 200 400 600 800 1000 1200 1400 1600 0 0.2 0.4 0.6 0.8 1 Throughput (keys/s) (a) CPU delay (ms) PKG SG KG 1000 1100 1200 0 1.105 2.105 3.105 4.105 (b) Memory (counters) 10s 30s 60s 300s 300s 600s 600s PKG SG KG Fig. 5: (a) Throughput for PKG, SG and KG for different CPU delays. (b) Throughput for PKG and SG vs. average memory for different aggregation periods. 

Latency 54 In the second experiment, we fix the CPU delay to 0.4ms per key, as it seems to be the limit of saturation for kg Table 4: Complete latency per message (ms) for di↵erent techniques, CPU delays and aggregation periods. CPU delay D (ms) Aggregation period T (s) D=0.1 D=0.5 D=1 T=10 T=30 T=60 pkg 3.81 6.24 11.01 6.93 6.79 6.47 sg 3.66 6.11 10.82 7.01 6.75 6.58 kg 3.65 9.82 19.35 

Impact Open source https://github.com/gdfm/partial-key-grouping Integrated in Apache Storm 0.10 (STORM-632, STORM-637) Plan to integrate it in Samza 55 

Conclusions Mining big data streams is an open field Needs collaboration between 
 algorithms and systems communities SAMOA: a platform for mining big data streams And for collaboration on distributed stream mining Algorithm-system co-design Promising future direction 56 System Algorithm API 

Future Work Algorithms Lift assumptions of ideal systems Systems New primitives targeted to mining algorithms Impact Open-source involvement with ASF 57 

Thanks! 58 https://samoa.incubator.apache.org @ApacheSAMOA @gdfm7 [email protected]