## Consistent Hashing in Clojure

I wrote this post to teach myself consistent hashing - a simple hash family that Akamai’s founders came up with. This was originally done to prepare for a talk in my grad algorithms class (I made a horlicks of the talk but whatever). I am going to provide intuition, analysis and a clojure implementation.

## The Problem Setting

A server contains viral resources (frequently hit webpages) and is swamped with requests. The obvious solution is to replicate this page using a set of caches (a set because a single cache can be overwhelmed by the same traffic surge) and after a cache miss, we send the request to the server.

We employ consistent hashing to assign and retrieve resources from these caches.

The set of caches is not fixed and even the set of caches a client can observe is not fixed. So one would like to be at the sweet spot of redundancy and uniformity (i.e. an even distribution of items among caches). Add to this, the requirement that we cannot predict a traffic surge.

The natural solution is thus to use hashing to determine which cache to use to store a resource (or which cache to retrieve a resource from). A traditional hash function of the type $(ax + b) \% p$ is a poor choice since anytime $p$ changes (which is quite often), the entire assignment of resources changes and we are stuck with a massive reassignment overhead.

Consistent hashing achieves all these properties and has the added advantage of being quite simple.

Anywho the operations are:

• Get a circle of circumference 1.

• Take your items and map them to some point on this circle (using something like MD5).

• Then take the available caches (i.e. whatever you can see) and map them to the circle as well.

• To assign an item to a cache, start at the item, keep going clockwise and the first cache you hit is the one you assign the item to.

And that is it! (this blog post has a few good graphics illustrating this algorithm).

Now, I wanted to perform a simulation of this algorithm in clojure to see empirically how good the performance is. I basically simulated 100 assignments. This experiment essentially can be thought of as: 100 caches, 10000 items and the routine that assigns an item to a cache updates its assignment ten times - each time observing a (random) subset of the available caches. This is a distribution of the load:

Essentially the the size of the circle is an indication of the number of items assigned. A very large circle indicates that a large number of items were mapped to that cache. As you can see, only a few caches have been overloaded with resources (possibly because one of the workers picked up very few caches and mapped a lot of resources to them).

Clearly the empirical evidence shows that the distribution of resources in these caches is really good. Only a few caches out of 100 have an extraordinarily high number of resources assigned to them. That experiment and the displayed gif were generated using code in this repo.

In particular, the code for consistent hashing looks like:

  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 (ns consistent-hashing.core (:require [digest]) (:import [java.nio ByteBuffer])) (def items (range 10000)) (def caches (range 100)) (defn map-to-unit-circle [x] (let [arr (ByteBuffer/allocate 4) val (do (.putInt arr x) (->> arr (.array) digest/md5 (.getBytes) bigint))] (double (/ (rem val 100000) 100000)))) (defn map-cache-to-unit-circle [x] (let [val (->> x str digest/md5 (.getBytes) bigint)] (double (/ (rem val 100000) 100000)))) (defn assign-item [[item pt] caches-points] (let [dsts (sort-by second (map (fn [[c p]] [c (if (< p pt) (+ p (- 1 pt)) (- p pt))]) caches-points))] (first (first dsts)))) (defn assign "Items: a set of items Caches: a set of caches" [items caches] (let [mapped-items (into {} (map vector items (map map-to-unit-circle items))) mapped-caches (sort-by second (map vector caches (map map-cache-to-unit-circle caches)))] (map (fn [x] (assign-item x mapped-caches)) mapped-items))) (defn random-take [coll n] (let [coll-set (set coll) item (rand-nth (into [] coll-set))] (if (zero? n) [] (cons item (random-take (clojure.set/difference coll-set (set [item])) (dec n)))))) (defn simulation [] (let [num-assgns 100] (map (fn [i] (let [to-take (rand-nth (range 1 (count caches))) seen-caches (random-take caches to-take)] (assign items seen-caches))) (range num-assgns)))) 

And the gif was generated using:

  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 (ns consistent-hashing.animate (:require [consistent-hashing.core :as core]) (:import [javax.swing JFrame JLabel] [java.awt.image BufferedImage] [java.awt Graphics Dimension Color] [gifAnimation GifEncoder])) (defn paint-canvas [size graphics caches assignment] ;; draw the unit circle (.setColor graphics (Color. 255 255 255)) (.fillRect graphics 0 0 (+ 10 size) (+ 10 size)) (doseq [[c x] caches] (.setColor graphics (Color. 255 0 0)) (when (assignment c) (let [diam (+ 10 (/ (assignment c) 250))] (.fillOval graphics (+ (/ size 2) (- (/ diam 2)) (int (* (/ size 2) (Math/cos (* x 2 Math/PI))))) (+ (/ size 2) (- (/ diam 2)) (int (* (/ size 2) (Math/sin (* x 2 Math/PI))))) diam diam)))) (.drawOval graphics 0 0 size size)) (defn draw [size caches assignment] (let [image (BufferedImage. (+ 10 size) (+ 10 size) BufferedImage/TYPE_INT_RGB) ;; canvas (proxy [JLabel] [] ;; (paint [g] (.drawImage g image 0 0 this))) ] (paint-canvas size (.createGraphics image) caches assignment) ;; (doto (JFrame.) ;; (.add canvas) ;; (.setSize (Dimension. (+ 10 size) (+ 10 size))) ;; (.show)) image)) (defn animate-load [] (let [mapped-items (map vector core/items (map core/map-to-unit-circle core/items)) mapped-caches (map vector core/caches (map core/map-cache-to-unit-circle core/caches)) simulated (core/simulation) load-pics (map (fn [xs] (reduce (fn [acc [c is]] (merge-with +' acc {c (count is)})) {} xs)) (reductions (fn [acc x] (let [cache-items (reduce (fn [acc [c is]] (merge-with clojure.set/union acc {c (set is)})) {} (map vector x (map vector core/items)))] (merge-with clojure.set/union acc cache-items))) {} simulated)) encoder (new GifEncoder)] ;;(draw 800 mapped-caches load-pics) (.start encoder "load.gif") (.setRepeat encoder 0) (doseq [assignments (rest load-pics)] (println assignments) (let [img (draw 500 mapped-caches assignments)] (.addFrame encoder img))) (.finish encoder))) 

Next, I will provide a proof of the claims made about consistent hashing.

The paper suggests that there are 3 dimensions to a good distribution of items in caches:

• Monotonicity

• Balance

Let us see each of these properties in detail:

## Monotonicity

Essentially, when new caches are added, a resource can get mapped only to one of the new caches but can’t be mapped to an existing cache (this clearly limits the amount of remapping you will engage in).

This property ensures that a single item is not mapped to too many caches (thus ensuring that items are evenly spread across the set of caches).

We could empirically see the evidence for this in the gif above. The proof of this metric has the following flavor:

• Assume that items are uniformly distributed on the circle.

• An arc length of $O(\log{n})$ on this circle contains a cache with very high probability (demonstrated using a simple Chernoff bound application that is available in the original paper).

• Stemming from this, we can be sure that, you don’t have to travel far from an item to hit a cache.

• From that very lemma, you can also conclude that not too many caches lie in a small interval around the item.

• Use the union bound and you get a nice probability expression for the spread. i.e. with a high probability, you get a good spread.

Load guarantees that a single cache doesn’t contain too many items (again the evidence for this is available in the gif above).

The proof is identical to the proof for the spread case (just switch the items with the caches).

## Balance

Balance ensures that overall, the probability that an item is mapped to a cache is uniform (ish) across all caches.

Essentially a chernoff bound gives you a uniform distribution.

Consistent hashing is pretty much everywhere (memcached, Akamai, etc. etc.) and is one of the all time greats of computer science.