search

twitter / algebird

Abstract Algebra for Scala
https://twitter.github.io/algebird
Apache License 2.0
2.29k stars 346 forks source link

Count-Min-Sketch Heavy Hitters duplication. #139

Open wlue opened 11 years ago

wlue commented 11 years ago

When adding two CountMinSketch objects, is it intentional that the heavy hitters are duplicated when they're summed using the monoid? The heavyHitters method on CMS is correct because it filters out. Here's an example where I add two CMS objects, and inspect the heavy hitters internal representation.

scala> import com.twitter.algebird._
import com.twitter.algebird._

scala> val monoid = CMS.monoid(0.01, 0.01, 0, 0.01)
monoid: com.twitter.algebird.CountMinSketchMonoid = com.twitter.algebird.CountMinSketchMonoid@3219ab8d

scala> monoid.create(Seq(1L, 1L, 1L))
res0: com.twitter.algebird.CMS = CMSInstance(CMSCountsTable(Vector(Vector(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
scala> monoid.create(Seq(1L, 1L, 1L, 1L))
res1: com.twitter.algebird.CMS = CMSInstance(CMSCountsTable(Vector(Vector(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
scala> monoid.plus(res0, res1)
res2: com.twitter.algebird.CMS = CMSInstance(CMSCountsTable(Vector(Vector(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
scala> res2.asInstanceOf[CMSInstance].heavyHitters
res3: Set[Long] = TreeSet(1)

scala> res2.asInstanceOf[CMSInstance].hhs
res4: com.twitter.algebird.HeavyHitters = HeavyHitters(TreeSet(HeavyHitter(1,3), HeavyHitter(1,4)))

I would expect that the internal representation of hhs would remove the duplicate value. Since I serialize the data, it becomes a major waste of space storing a bunch of old values. For now, I just remove duplicate 'item' HeavyHitter keys when serialized, but I was wondering if this was intentional, since I can see that being the case for faster summing.

johnynek commented 11 years ago

https://github.com/twitter/algebird/blob/develop/algebird-core/src/main/scala/com/twitter/algebird/CountMinSketch.scala#L449

That's the bug line. The HeavyHitters object is considered different if it is the same value but different count.

johnynek commented 11 years ago

Actually this is busted:

https://github.com/twitter/algebird/blob/develop/algebird-core/src/main/scala/com/twitter/algebird/CountMinSketch.scala#L331

the problem is, it it assume the old count is known, but the appoximation may change it, and so we need to be more careful. I think what we need to do is just keep a set of keys, and not their counts at all. I don't think you need the sorted set.

johnynek commented 10 years ago

Need a failing unit test here and then a fix. I think this might still be broken.