Algebird : Abstract Algebra for Big Data Analytics.

Devoxx 2014. Antwerp Belgium.
Tools in Action Room 4
Antwerp, Tue 11th Nov. 2014

Transcript

1. Algebird Abstract Algebra for Analytics Sam BESSALAH @samklr
5. Abstract Algebra
6. From WikiPedia
7. Algebraic Structure “ Set of values, coupled with one or more finite operations,and a set of laws those operations must obey. “
8. Algebraic Structure “ Set of values, coupled with one or more finite operations, and a set of laws those operations must obey. “ e.g Sum, Magma, Semigroup, Groups, Monoid, Abelian Group, Semi Lattices, Rings, Monads, etc.
9. Semigroup Semigroup Law : (x <> y) <> z = x <> (y <> z) (associativity)
10. Semigroup Semigroup Law : (x <> y) <> z = x <> (y <> z) (associativity) trait Semigroup[T] { def aggregate(x : T, y : T) : T }
11. Monoids Monoid Laws : (x <> y) <> z = x <> (y <> z) (associativity) identity <> x = x x <> identity = x (identity)
12. Monoids Monoid Laws : (x <> y) <> z = x <> (y <> z) (associativity) identity <> x = x x <> identity = x (identiy / zero) trait Monoid[T] { def identity : T def aggregate (x, y) : T }
13. Monoids Monoid Laws : (x <> y) <> z = x <> (y <> z) (associativity) identity <> x = x x <> identity = x trait Monoid[T] extends Semigroup[T]{ def identity : T }
14. Groups Group Laws: (x <> y) <> z = x <> (y <> z) (associativity) identity <> x = x x <> identity = x (identity) x <> inverse x = identity inverse x <> x = identity (invertibility)
15. Groups Group Laws (x <> y) <> z = x <> (y <> z) identity <> x = x x <> identity = x x <> inverse x = identity inverse x <> x = identity trait Group[T] extends Monoid[T]{ def inverse (v : T) :T }
16. Many More - Abelian groups (Commutative Sets) - Rings - Semi Lattices - Ordered Semigroups - Fields .. Many of those are in Algebird ….
17. Examples - (a min b) min c = a (b min c) with Int. - a max ( b max c) = (a max b) max c ** - a or (b or c) = (a or b) or c - a and (b and c) = (a and b) and c - int addition - set union - harmonic sum - Integer mean - Priority queue
19. Why do we need those algebraic structures ?
20. We want to : - Build scalable analytics systems - Leverage distributed computing to perform aggregation on really large data sets. - A lot of operations in analytics are just sorting and counting at the end of the day
21. Distributed Computing → Parallellism
22. Distributed Computing → Parallellism Associativity → enables parallelism
24. Distributed Computing → Parallellism Associativity enables parallelism Identity means we can ignore some data Commutativity helps us ignore order
25. Typical Map Reduce ...
26. Finding Top-K Elements in Scalding ... class TopKJob(args : Args) extends Job (args) { Tsv ( args(‘input’), visitScheme) .filter (. ..) .leftJoinWithTiny ( … ) .filter ( … ) .groupBy( ‘fieldOne) { _.sortWithTake (visitScheme -> top } (biggerSale) .write(Tsv(...) ) }
27. .sortWithTake( … ) Looking into .sortWithTake in Scalding, there’s one nice thing : class PiorityQueueMonoid[T] (max : Int) (implicit order : Ordering[T] ) extends Monoid[Priorityqueue[T] ]
28. class PiorityQueueMonoid[T] (max : Int) (implicit order : Ordering[T] ) extends Monoid[Priorityqueue[T] ] Let’s take a look : PQ1 : 55, 45, 21, 3 PQ2: 100, 80, 40, 3 top-4 (PQ1 U PQ2 ): 100, 80, 55, 45 Priority Queue : Can be empty Two Priority Queues can be “added” in any order Associative + Commutative
29. class PiorityQueueMonoid[T] (max : Int) (implicit order : Ordering[T] ) extends Monoid[Priorityqueue[T] ] Let’s take a look : PQ1 : 55, 45, 21, 3 PQ2: 100, 80, 40, 3 top-4 (PQ1 U PQ2 ): 100, 80, 55, 45 Priority Queue : Can be empty Two Priority Queues can be “added” in any order Associative + Commutative Makes Scalding go fast, by doing sorting, filtering and extracting in one single “map” step.
30. Stream Mining Challenges - Update predictions after each observation - Single pass : can’t read old data or replay the stream - Full size of the stream often unknown - Limited time for computation per observation - O(1) memory size
31. Stream Mining Challenges http://radar.oreilly.com/2013/10/stream-mining-essentials.html
32. Tradeoff : Space and speed over accuracy.
33. Tradeoff : Space and speed over accuracy. use sketches.
34. Sketches Probabilistic data structures that store a summary (hashed mostly)of a data set that would be costly to store in its entirety, thus providing most of the time, sublinear algorithmic properties. E.g Bloom Filters, Counter Sketch, KMV counters, Count Min Sketch, HyperLogLog, Min Hashes
35. Bloom filters Approximate data structure for set membership Behaves like an approximate set BloomFilter.contains(x) => NO | Maybe P(False Positive) > 0 P(False Negative) = 0
36. Internally : Bit Array of fixed size add(x) : for all element i, b[h(x,i)]=1 contains(x) : TRUE if b[h(x,i)] = = 1 for all i. (Boolean AND => associative) Both are associative => BF can be designed as a Monoid
37. Bloom filters import com.twitter.algebird._ import com.twitter.algebird.Operators._ // generate 2 lists val A = (1 to 300).toList // Generate a Bloomfilter val NUM_HASHES = 6 val WIDTH = 6000 // bits val SEED = 1 implicit val bfm = new BloomFilterMonoid(NUM_HASHES, WIDTH, SEED) // approximate set with bloomfilter val A_bf = A.map{i => bfm.create(i.toString)}.reduce(_ + _) val approxBool = A_bf.contains(“150”) ---> ApproximateBoolean(true, 0.9995…)
38. Count Min Sketch Gives an approximation of the number of occurrences of an element in a set.
39. Count Min Sketch Count min sketch Adding an element is a numerical addition Querying uses a MIN function. Both are associative. useful for detecting heavy hitters, topK, LSH We have in Algebird :
40. HyperLogLog Popular sketch for cardinality estimtion. Gives within a probilistic distribution of an error the number of distinct values in a data set. HLL.size = Approx[Number] Intuition Long runs of trailings 0 in a random bits chain are rare But the more bit chains you look at, the more likely you are to find a long one The longest run of trailing 0-bits seen can be an estimator of the number of unique bit chains observed.
41. Adding an element uses a Max and Sum function. Both are associative and Monoids. (Max is an ordered semigroup in Algebird really) Querying for an element uses an harmonic mean which is a Monoid. In Algebird :
42. Many More juicy sketches ... - MinHashes to compute Jaccard similarity - QTree for quantiles estimation. Neat for anomaly detection. - SpaceSaverMonoid, Awesome to find the approximate most frequent and top K elements. - TopKMonoid - SGD, PriorityQueues, Histograms, etc.
43. SummingBird : Lamba in a box
44. Heard of Lambda Architecture ?
45. SummingBird Same code for both batch and real time processing.
46. SummingBird Same code, for both batch and real time processing. But works only on Monoids. Uses Storehaus, as a mergeable store layer.
47. http://github.com/twitter/algebird
48. http://github.com/twitter/algebird
49. These slides : http://bit.ly/1szncAZ http://slidesha.re/1zhhXKU
51. -Algebra for analytics by Oscar Boykin (Creator of Algebird) http://speakerdeck.com/johnynek/algebra-for-analytics - Take a look into HLearn https://github.com/mikeizbicki/HLearn - Great intro into Algebird by Michael Noll http://www.michael-noll.com/blog/2013/12/02/twitter-algebird-monoid- monad-for-large-scala-data-analytics/ -Aggregate Knowledge http://research.neustar.biz/2012/10/25/sketch-of- the-day-hyperloglog-cornerstone-of-a-big-data-infrastructure - Probabilistic data structures for web analytics. http://highlyscalable.wordpress.com/2012/05/01/probabilistic- structures-web-analytics-data-mining/ - http://debasishg.blogspot.fr/2014/01/count-min-sketch-data- structure-for.html - http://infolab.stanford.edu/~ullman/mmds/ch3.pdf Links