Spark and the Berkeley Data Analytics Stack (BDAS) represent a unified, distributed, and parallel high-performance big data processing and analytics platform. Written in Scala, Spark supports multiple languages including Python, Java, Scala, and even R. Commonly seen as the successor to Hadoop, Spark is fully compatible with Hadoop including UDFs, SerDe’s, file formats, and compression algorithms. The high-level Spark libraries include stream processing, machine learning, graph processing, approximating, sampling - and every combination therein. The most active big data open source project in existence, Spark boasts ~500 of contributors and 10,000 commits to date. Spark recently broke the Daytona GraySort 100 TB record with almost 3 times the throughput, 1/3rd less time, and 1/10th of the resources!

1. After Dark
Generating High-Quality Recommendations using
Real-time Advanced Analytics and Machine Learning with
Chris Fregly
Data Solutions Engineer @ Databricks

2. Who am I?
2
Data Platform Engineer
playboy.com
Streaming Platform Engineer
NetflixOSS Committer
netflix.com, github.com/Netflix
Data Solutions Engineer
Apache Spark Contributor
databricks.com, github.com/apache/spark

3. Why After Dark?
Playboy After Dark
Late 1960’s TV Show
Progressive Show For Its Time
And it rhymes!!
3

4. What is ?
4
Spark Core
Spark
Streaming
real-timeSpark SQL
structured data
MLlib
machine
learning
GraphX
graph
analytics
…
BlinkDB
approx queries

5. in Production
5

6. What is ?
6
Founded by the creators of
as a Service
Powerful Visualizations
Collaborative Notebooks
Scala/Java, Python, SQL, R
Flexible Cluster Management
Job Scheduling and Monitoring

7. in Production
7

8. 8
① Generate high-quality recommendations
② Demonstrate Spark high-level libraries:
③ Spark Streaming -> Kafka, Approximates
④ Spark SQL -> DataFrames, Cassandra
① GraphX -> PageRank, Shortest Path
① MLlib -> Matrix Factor, Word2Vec
Goals of After Dark?
Images courtesy of tinder.com. Not affiliated with Tinder in any way.

9. Popular Dating Sites
9

10. Themes of this Talk
10
① Performance
② Parallelism
③ Columnar Storage
④ Approximations
⑤ Similarity
⑥ Minimize Shuffle

11. Performance
11

12. Daytona Gray Sort Contest
12
On-disk only
250,000 partitions
No in-memory caching
(2014)(2013) (2014)

13. Improved Shuffle and Network Layer
13
① Introduced sort-based shuffle
Mapper maintains large buffer grouped by keys
Reducer seeks directly to group and scans
② Minimizes OS resources
Less mapper-reducer open files,connections
③ Netty: Async keeps CPU hot, reuse ByteBuffer
④ epoll: disk-network comm in kernel space only

14. Project Tungsten: CPU and Memory
14
① Largest change to Spark exec engine to date
② Cache-aware data structs and sorting
->
③ Expand JVM bytecode gen, JIT optimizations
④ Custom mem manage, serializers, HashMap

15. DataFrames and Catalyst
15
15
https://ogirardot.wordpress.com/2015/05/29/rdds-are-the-new-bytecode-of-apache-spark/
Tip: Use DataFrames! -->
JVM bytecode
generation

16. Parallelism
16

17. Brady Bunch circa 1980
17
Season 5, Episode 18: “Two Petes in a Pod”

18. Parallel Algorithm : O(log n)
18
O(log n)

19. Non-parallel Algorithm : O(n)
19
O(n)

20. Columnar Storage
20

21. Columnar Storage Format
21
*Skip whole chunks with min-max heuristics
stored in each chunk (sorted data only)

22. Parquet File Format
22
① Based on Google Dremel Paper
② Implemented by Twitter and Cloudera
③ Columnar storage format
④ Optimized for fast columnar aggregations
⑤ Tight compression
⑥ Supports pushdowns
⑦ Nested, self-describing, evolving schema

23. Types of Compression
23
① Run Length Encoding
Repeated data
② Dictionary Encoding
Fixed set of values
③ Delta, Prefix Encoding
Sorted dataset

24. Types of Pushdowns
24
① Column, Partition Pruning
② Row, Predicate Filtering

25. Approximations
25

26. Sketch Algorithm: Count Min Sketch
26
① Approximate counters
② Better than HashMap
③ Fixed, low memory
④ Known error bounds
⑤ Large num of counters
⑥ Available in Twitter’s Algebird
⑦ Streaming example in Spark

27. Probabilistic Data Structure: HyperLogLog
27
① Fixed memory
② Known error distribution
③ Measures set cardinality
④ Approx count distinct
⑤ Number of unique users
⑥ From Twitter’s Algebird
⑦ Streaming example in Spark
⑧ RDD: countApproxDistinctByKey()

28. Similarity
28

29. Types of Similarity
29
① Euclidean: linear measure
Magnitude bias
② Cosine: angle measure
Adjusts for magnitude bias
③ Jaccard: set intersection divided by union
Popularity bias
④ Log Likelihood
Adjusts for bias -->
Ali
Matei
Reynold
Patrick
Andy
Kimberly
1
1
1
1
Paula
1
Lisa
1
Cindy
1
1
Holden
1
1
1
1
1
z

30. All-pairs Similarity
30
① Compare everything to everything
② aka. “pair-wise similarity” or “similarity join”
③ Naïve shuffle: O(m*n^2); m=rows, n=cols
④ Minimize shuffle: reduce data size & approx
Reduce m (rows)
Sampling and bucketing
Reduce n (cols)
Remove most frequent value (0?)

31. Minimize Shuffle
31

32. Sampling Algo: DIMSUM
32
① "Dimension Independent Matrix Square
Using MR”
② Remove rows with low similarity probability
③ MLlib: RowMatrix.columnSimilarities(…)
④ Twitter: 40% efficiency gain over Cosine

33. Bucket Algo: Locality Sensitive Hashing
33
① Split into b buckets using similarity hash algo
Requires pre-processing of data
② Compare bucket contents in parallel
③ Converts O(m*n^2) -> O(m*n/b*b^2);
m=rows, n=cols, b=buckets
④ Example: 500k x 500k matrix
O(1.25E17) -> O(1.25E13); b=50
⑤ github.com/mrsqueeze/spark-hash

34. MLlib: SparseVector vs. DenseVector
34
① Remove columns using sparse vectors
② Converts O(m*n^2) -> O(m*nnz^2);
nnz=num nonzeros, nnz << n
Tip: Choose most frequent value … may not be 0

35. Interactive Demo!
35

36. Audience Participation Needed!
36
① Navigate to sparkafterdark.com
② Click 3 actors and 3 actresses
->
You are here
->

37. Recommendation Terminology
37
① User
User seeking likeable recommendations
② Item
User who has been liked
*Also a user seeking likeable recommendations!
③ Types of Feedback
Explicit: Ratings, Like/Dislike
Implicit: Search, Click, Hover, View, Scroll

38. Types of Recommendations
38
① Non-personalized
Cold Start
No preference or behavior data for user, yet
② Personalized
Items that others with similar prefs have liked
User-Item Similarity
Items similar to your previously-liked items
Item-Item Similarity

39. Non-personalized
Recommendations
39

40. Summary Statistics and Aggregations
40
① Top Users by Like Count
“I might like users with the highest sum aggregation
of likes overall.”
SparkSQL + DataFrame: Aggregations

41. Like Graph Analysis
41
② Top Influencers by Like Graph
“I might like users who have the highest probability of
me liking them randomly while walking the like graph.”
GraphX: PageRank

42. Demo!
Spark SQL + DataFrames + GraphX
42

43. Personalized
Recommendations
43

44. Collaborative Filtering Personalized Recs
44
③ Like behavior of similar users
“I like the same people that you like.
What other people did you like that I haven’t seen?”
MLlib: Matrix Factorization, User-Item Similarity

45. Text-based Personalized Recs
45
④ Similar profiles to each other
“Our profiles have similar, unique k-skip n-grams.
We might like each other.”
MLlib: Word2Vec, TF/IDF, Doc Similarity

46. More Text-based Personalized Recs
46
⑤ Similar profiles from my past likes
“Your profile shares a similar feature vector space to
others that I’ve liked. I might like you.”
MLlib: Word2Vec, TF/IDF, Doc Similarity

47. More Text-based Personalized Recs
47
⑥ Relevant, High-Value Emails
“Your initial email has similar named entities to my profile.
I might like you just for making the effort.”
MLlib: Word2Vec, TF/IDF, Entity Recognition
^
Her Email
< My Profile

48. Demo!
MLlib + ALS + Word2Vec + TF/IDF
48

49. Bonus!
The Future of Recommendations
49

50. Facial Recognition
50
⑦ Eigenfaces
“Your face looks similar to others that I’ve liked.
I might like you.”
MLlib: RowMatrix, PCA, Item-Item Similarity
Image courtesy of http://crockpotveggies.com/2015/02/09/automating-tinder-with-eigenfaces.html

51. Conversation Starter Bot
51
⑧ NLP and DecisionTrees
“If your responses to my trite opening lines are positive,
I might actually read your profile.”
MLlib: TF/IDF, DecisionTree,
Sentiment Analysis
Positive
responses ->
Negative
<- responses
Image courtesty of http://crockpotveggies.com/2015/02/09/automating-tinder-with-eigenfaces.html

52. Double Bonus!
52
Maintaining the

53. Compromise Recommendations (Couples)
53
⑨ Similarity Pathways
“I want Mad Max. You want Message In a Bottle.
Let’s find something in between to watch tonight.”
MLlib: RowMatrix, Item-Item Similarity
GraphX: Nearest Neighbors, Shortest Path
similar similar
plots -> <- actors
… …

54. And the Final,
54
⑩ Personalized Recommendation

55. My Personalized Recommendation
55
⑩ Get Off Your Computer and Be Social!!
Thank you!
cfregly@databricks.com
@cfregly
Image courtesy of http://www.duchess-france.org/