by· February 18, 2014
At Metamarkets, we specialize in converting mountains of programmatic ad data into real-time, explorable views. Because these datasets are so large and complex, we’re always looking for ways to maximize the speed and efficiency of how we deliver them to our clients. In this post, we’re going to continue our discussion of some of the techniques we use to calculate critical metrics such as unique users and device IDs with maximum performance and accuracy.
Approximation algorithms are rapidly gaining traction as the preferred way to determine the unique number of elements in high cardinality sets. In the space of cardinality estimation algorithms, HyperLogLog has quickly emerged as the de-facto standard. Widely discussed by technology companies and popular blogs, HyperLogLog trades accuracy in data and query results for massive reductions in data storage and vastly improved system performance.
In our previous investigation of HyperLogLog, we briefly discussed our motivations for using approximate algorithms and how we leveraged HyperLogLog in Druid, Metamarkets’ open source, distributed data store. Since implementing and deploying HyperLogLog last year, we’ve made several optimizations to further improve performance and reduce storage cost. This blog post will share some of those optimizations. This blog post assumes that you are already familiar with how HyperLogLog works. If you are not familiar with the algorithm, there are plenty of resources online.
In our initial implementation of HLL, we allocated 8 bits of memory for each register. Recall that each value stored in a register indicates the position of the first ‘1’ bit of a hashed input. Given that 2^255 ~== 10^76, a single 8 bit register could approximate (not well, though) a cardinality close to the number of atoms in the entire observable universe. Martin Traverso, et. al. of Facebook’s Presto , realized that this was a bit wasteful and proposed an optimization, exploiting the fact that the registers increment in near lockstep.
Given that each register is initially initialized with value 0, with 0 uniques, there is no change in any of the registers. Let’s say we have 8 registers. Then with 8 * 2^10 uniques, each register will have values ~ 10. Of course, there will be some variance, which can be calculated exactly if one were so inclined, given that the distribution in each register is an independent maximum of Negative Binomial (1, .5) draws.
With 4 bit registers, each register can only approximate up to 2^15 = 32,768 uniques. In fact, the reality is worse because the higher numbers cannot be represented and are lost, impacting accuracy. Even with 2,048 registers, we can’t do much better than ~60M, which is one or two orders of magnitude lower than what we need.
Since the register values tend to increase together, the FB folks decided to introduce an offset counter and only store positive differences from it in the registers. That is, if we have register values of 8, 7, and 9, this corresponds to having an offset of 7 and using register difference values of 1, 0, and 2. Given the smallish spread that we expect to see, we typically won’t observe a difference of more than 15 among register values. So we feel comfortable using 2,048 4 bit registers with an 8 bit offset, for 1025 bytes of storage < 2048 bytes (no offset and 8 bit registers).
In fact, others have commented on the concentrated distribution of the register values as well. In her thesis, Marianne Durand suggested using a variable bit prefix encoding. Researchers at Google have had success with difference encodings and variable length encodings.
This optimization has served us well, with no appreciable loss in accuracy when streaming many uniques into a single HLL object, because the offset increments when all the registers get hit. Similarly, we can combine many HLL objects of moderate size together and watch the offsets increase. However, a curious phenomenon occurs when we try to combine many “small” HLL objects together.
Suppose each HLL object stores a single unique value. Then its offset will be 0, one register will have a value between 1 and 15, and the remaining registers will be 0. No matter how many of these we combine together, our aggregate HLL object will never be able to exceed a value of 15 in each register with a 0 offset, which is equivalent to an offset of 15 with 0’s in each register. Using 2,048 registers, this means we won’t be able to produce estimates greater than ~ .7 * 2048^2 * 1 / (2048 / 2^15) ~ 47M. (Flajolet, et al. 2007)
Not good, because this means our estimates are capped at 10^7 instead of 10^80, irrespective of the number of true uniques. And this isn’t just some pathological edge case. Its untimely appearance in production a while ago was no fun trying to fix.
The root problem in the above scenario is that the high values (> 15) are being clipped, with no hope of making it into a “small” HLL object, since the offset is 0. Although they are rare, many cumulative misses can have a noticeably large effect. Our solution involves storing one additional pair, a “floating max” bucket with higher resolution. Previously, a value of 20 in bucket 94 would be clipped to 15. Now, we store (20, 94) as the floating max, requiring at most an additional 2 bytes, bringing our total up to 1027 bytes. With enough small HLL objects so that each position is covered by a floating max, the combined HLL object can exceed the previous limit of 15 in each position. It also turns out that just one floating max is sufficient to largely fix the problem.
Let’s take a look at one measure of the accuracy of our approximations. We simulate 1,000 runs of streaming 1B uniques into an HLL object and look at the proportion of cases in which we observed clipping with the offset approximation (black) and the addition of the floating max (red). So for 1e9 uniques, the max reduced clipping from 95%+ to ~15%. That is, in 85% of cases, the much smaller HLL objects with the floating max agreed with HLL versus less than 5% without the floating max.
For the cost of only 2 bytes, the floating max register allowed us to union millions of HLL objects with minimal measurable loss in accuracy.
We first discussed the concept of representing HLL buckets in either a sparse or dense format in our first blog post. Since that time, Google has also written a great paper on the matter. Data undergoes a summarization process when it is ingested in Druid. It is unnecessarily expensive to store raw event data and instead, Druid rolls ingested data up to some time granularity.
In practice, we see tremendous reductions in data volume by summarizing our data. For a given summarized row, we can maintain HLL objects where each object represents the estimated number of unique elements for a column of that row.
When the summarization granularity is sufficiently small, only a limited number of unique elements may be seen for a dimension. In this case, a given HLL object may have registers that contain no values. The HLL registers are thus ‘sparsely’ populated.
Our normal storage representation of HLL stores 2 register values per byte. In the sparse representation, we instead store the explicit indexes of buckets that have valid values in them as (index, value) pairs. When the sparse representation exceeds the size of the normal or ‘dense’ representation (1027 bytes), we can switch to using only the dense representation. Our actual implementation uses a heuristic to determine when this switch occurs, but the idea is the same. In practice, many dimensions in real world data sets are of low cardinality, and this optimization can greatly reduce storage versus only storing the dense representation.
One of the simpler optimizations that we implemented for faster cardinality calculations was to use lookups for register values. Instead of computing the actual register value by summing the register offset with the stored register value, we instead perform a lookup into a precalculated map. Similarly, to determine the number of zeros in a register value, we created a secondary lookup table. Given the number of registers we have, the cost of storing these lookup tables is near trivial. This problem is often known as the Hamming Weight problem.
Many of our optimizations came out of necessity, both to provide the interactive query latencies that Druid users have come to expect, and to keep our storage costs reasonable. If you have any further improvements to our optimizations, please share them with us! We strongly believe that as data sets get increasingly larger, estimation algorithms are key to keeping query times acceptable. The approximate algorithm space remains relatively new, but it is something we can build together.
For more information on Druid, please visit druid.io and follow @druidio. We’d also like to thank Eric Tschetter and Xavier Léauté for their contributions to this work. Featured image courtesy of Donna L Martin.