Join us for RedisConf and Hackathon, April 20-21
This post is about what are, in my opinion, two of the most exciting things in the world: probabilistic data structures and Redis modules. If you’ve heard about one or the other then you can surely relate to my enthusiasm, but in case you want to catch up on the coolest stuff on earth just continue reading.
I’ll start with this statement: large-scale low-latency data processing is challenging. The volumes and velocities of data involved can make real-time analysis extremely demanding. Due to its high performance and versatility, Redis is commonly used to solve such challenges. Its ability to store, manipulate and serve multiple forms of data in sub-millisecond latency makes it the ideal data container in many cases where online computation is needed.
But everything is relative, and there are scales so extreme that they make accurate real-time analysis a practical impossibility. Complex problems only become more difficult as they get bigger, but we tend to forget that simple problems follow the same rule. Even something as basic as summing numbers can become a monumental task once data is too big, too fast or when we don’t have enough resources to process it. And while resources are always finite and expensive, data is constantly growing like nobody’s business.
Software engineering is all about making trade-offs, and a popular approach to addressing such challenges in a cost-effective way is to forgo accuracy in favor of efficiency. This approach is exemplified by Redis’ implementation of the HyperLogLog, a data structure that’s designed to efficiently provide answers to queries about set cardinality. The HyperLogLog is a member of a family of data structures called “sketches” that, just like their real-world artistic counterparts, convey information via an approximation of their subjects.
In broad strokes, sketches are data structures (and their accompanying algorithms) that capture the nature of your data—the answers to your questions about the data, without actually storing the data itself. Formally stated, sketches are useful because they have sublinear asymptotic computational complexity, thus requiring less computing power and/or storage. But there are no free lunches and the gain in efficiency is offset by the accuracy of the answers. In many cases, however, errors are acceptable as long as they can be kept under a threshold. A good data sketch is one that readily admits its errors, and in fact many sketches parameterize their errors (or the probability of said errors) so that they can be defined by the user.
Good sketches are efficient and have bound probability of error, but excellent sketches are those that can be computed in parallel. A parallelizable sketch is one that can be prepared independently on parts of the data and that allows combining its parts into a meaningful and consistent aggregate. Because each piece of an excellent sketch can be computed at a different location and/or time, parallelism makes it possible to apply a straightforward divide-and-conquer approach to solving scaling challenges.
The aforementioned HyperLogLog is an excellent sketch but it is only suited for answering a specific type of question. Another such invaluable tool is the Count Min Sketch (CMS), as described in the paper “An Improved Data Stream Summary: The Count-Min Sketch and its Applications” by G. Cormode and S. Muthukrishnan. This ingenious contraption was contrived to provide answers about the frequency of samples, a common building block in a large percentage of analytical processes.
Given enough time and resources, calculating samples’ frequency is a simple process – just keep a count of observations (times seen) for each sample (thing seen) and then divide that by the total number of observations to obtain that sample’s frequency. However, given the context of high-scale low-latency data processing, there’s never enough time or resources. Answers are to be supplied instantly as the data streams in, regardless of its scale, and the sheer size of the sampling space makes it unfeasible to keep a counter for each sample.
So instead of accurately keeping track of each sample, we can try estimating the frequency. One way of going about that is to randomly sample the observations and hope that the sample generally reflects the properties of the whole. The problem with this approach is that ensuring true randomness is a difficult task, so the success of random sampling may be limited by our selection process and/or the properties of the data itself. That is where CMS comes in with an approach so radically different, that at first it may seem like the opposite of an excellent sketch: not only does CMS record each and every observation, each one is recorded in multiple counters!
Of course there is a twist, and it is as clever as it is simple. The original paper (and its lighter version called “Approximating Data with the Count-Min Data Structure”) does a great job of explaining it, but I’ll try to summarize it anyway. The cleverness of CMS is in the way that it stores samples: instead of tracking each unique sample independently, it uses its hash value. The hash value of a sample is used as the index to a constant-sized (parameterized as d in the paper) array of counters. By employing several (the parameter w) different hash functions and their respective arrays, the sketch handles hash collisions found while querying the structure by picking the minimum value out of all relevant counters for the sample.
An example is called for, so let’s make a simple sketch to illustrate the data structure’s inner workings. To keep the sketch simple, we’ll use small parameter values. We’ll set w to 3, meaning we’ll use three hash functions – denoted h1, h2 and h3 respectively, and d to 4. To store the sketch’s counters we’ll use a 3×4 array with a total of 12 elements initialized to 0.
Now we can examine what happens when samples are added to the sketch. Let’s assume samples arrive one by one and that the hashes for the first sample, denoted as s1, are: h1(s1) = 1, h2(s1) = 2 and h3(s1) = 3. To record s1 in the sketch we’ll increment each hash function’s counter at the relevant index by 1. The following table shows the array’s initial and current states:
Although there’s only one sample in the sketch, we can already query it effectively. Remember that the number of observations for a sample is the minimum of all its counters, so for s1 it is obtained by:
min(array[h1(s1)], array[h2(s1)], array[h3(s1)]) =The sketch also answers queries about the samples not yet added. Assuming that h1(s2) = 4, h2(s2) = 4, h3(s2) = 4, note that querying for s2 will return the result 0. Let’s continue to add s2 and s3 (h1(s3) = 1, h2(s3) = 1, h3(s3) = 1) to the sketch, yielding the following:
min(array, array, array) =
min(1,1,1) = 1
In our contrived example, almost all of the samples’ hashes map to unique counters, with the one exception being the collision of h1(s1) and h1(s3). Because both hashes are the same, h1‘s 1st counter now holds the value 2. Since the sketch picks the minimum of all counters, the queries for s1 and s3 still return the correct result of 1. Eventually, however, once enough collisions have occurred, the queries’ results will become less accurate.
CMS’ two parameters – w and d – determine its space/time requirements as well as the probability and maximal value of its error. A more intuitive way to initialize the sketch is to provide the error’s probability and cap, allowing it to then derive the required values for w and d. Parallelization is possible because any number of sub-sketches can be trivially merged as the sum of arrays, as long as the same parameters and hash functions are used in constructing them.
Count Min Sketch’s efficiency can be demonstrated by reviewing its requirements. The space complexity of CMS is the product of w, d and the width of the counters that it uses. For example, a sketch that has 0.01% error rate at probability of 0.01% is created using 10 hash functions and 2000-counter arrays. Assuming the use of 16-bit counters, the overall memory requirement of the resulting sketch’s data structure clocks in at 40KB (a couple of additional bytes are needed for storing the total number of observations and a few pointers). The sketch’s other computational aspect is similarly pleasing—because hash functions are cheap to produce and compute, accessing the data structure, whether for reading or writing, is also performed in constant time.
There’s more to CMS; the sketch’s authors have also shown how it can be used for answering other similar questions. These include estimating percentiles and identifying heavy hitters (frequent items), but are out of scope for this post.
CMS is certainly an excellent sketch, but there are at least two things that prevent it from achieving perfection. My first reservation about CMS is that it can be biased, and thus overestimate the frequencies of samples with a small number of observations. CMS’ bias is well known and several improvements have been suggested. The most recent is the Count Min Log Sketch (“Approximately counting with approximate counters” from G. Pitel and G. Fouquier), which essentially replaces CMS’ linear registers with logarithmic ones to reduce the relative error and allow higher counts without increasing the width of counter registers.
While the above reservation is shared by everyone (albeit only those who grok data structures), my second reservation is exclusive to the Redis community. In order to explain, I’ll have to introduce Redis Modules.
Redis Modules were unveiled earlier this year at RedisConf by antirez and have literally turned our world upside-down. Nothing more or less than server-loadable dynamic libraries, Modules allow Redis users to move faster than Redis itself and go places never dreamt of before. And while this post isn’t an introduction to what Modules are or how to make them, this one is (as well as this post, and this webinar).
There are several reasons I wanted to write the Count Min Sketch Redis module, aside from its extreme usefulness. Part of it was a learning experience and part of it was an evaluation of the modules API, but mostly it was just a whole lot of fun to model a new data structure into Redis. The module provides a Redis interface for adding observations to the sketch, querying it and merging multiple sketches into one.
The module stores the sketch’s data in a Redis String and uses direct memory access (DMA) for mapping the contents of the key to its internal data structure. I’ve yet to conduct exhaustive performance benchmarks on it, but my initial impression from testing it locally is that it is as performant as any of the core Redis commands. Like our other modules, countminsketch is open source and I encourage you to try it out and hack on it.
Before signing off, I’d like to keep my promise and share my Redis-specific reservation about CMS. The issue, which also applies to other sketches and data structures, is that CMS requires you to set it up/initialize it/create it before using it. Requiring a mandatory initialization stage, such as CMS’ parameters setup, breaks one of Redis’ fundamental patterns—data structures do not need to be explicitly declared before use as they can be created on demand. To make the module seem more Redis-ish and work around that anti-pattern, the module uses default parameter values when a new sketch is implicitly implied (i.e. using the CMS.ADD command on a non-existing key) but also allows creating new empty sketches with given parameters.
Probabilistic data structures, or sketches, are amazing tools that let us keep up with the growth of big data and the shrinkage of latency budgets in an efficient and sufficiently accurate way. The two sketches mentioned in this post, and others such as the Bloom Filter and the T-digest, are quickly becoming indispensable tools in the modern data monger’s arsenal. Modules allow you to extend Redis with custom data types and commands that operate at native speed and have local access to the data. The possibilities are endless and nothing is impossible.
Want to learn more about Redis modules and how to develop them? Is there a data structure, whether probabilistic or not, that you want to discuss or add to Redis? Feel free to reach out to me with anything at my Twitter account or via email – I’m highly-available 🙂