by· September 12, 2013
I’d like to acknowledge Xavier Léauté for his extensive contributions (in particular, for suggesting several algorithmic improvements and work on implementation), helpful comments, and fruitful discussions. Featured image courtesy of CERN.
Many businesses care about accurately computing quantiles over their key metrics, which can pose several interesting challenges at scale. For example, many service level agreements hinge on these metrics, such as guaranteeing that 95% of queries return in < 500ms. Internet service providers routinely use burstable billing, a fact that Google famously exploited to transfer terabytes of data across the US for free. Quantile calculations just involve sorting the data, which can be easily parallelized. However, this requires storing the raw values, which is at odds with a pre-aggregation step that helps Druid achieve such dizzying speed. Instead, we store smaller, adaptive approximations of these values as the building blocks of our “approximate histograms.” In this post, we explore the related problems of accurate estimation of quantiles and building histogram visualizations that enable the live exploration of distributions of values. Our solution is capable of scaling out to aggregate billions of values in seconds.
When we first met Druid, we considered the following example of a raw impression event log:
timestamp publisher advertiser gender country dimensions click price 2011-01-01T01:01:35Z bieberfever.com google.com Male USA 0 0.65 2011-01-01T01:03:63Z bieberfever.com google.com Male USA 0 0.62 2011-01-01T01:04:51Z bieberfever.com google.com Male USA 1 0.45 ... 2011-01-01T01:00:00Z ultratrimfast.com google.com Female UK 0 0.87 2011-01-01T02:00:00Z ultratrimfast.com google.com Female UK 0 0.99 2011-01-01T02:00:00Z ultratrimfast.com google.com Female UK 1 1.53 ...
By giving up some resolution in the timestamp column (e.g., by truncating the timestamps to the hour), we can produce a summarized dataset by grouping by the dimensions and aggregating the metrics. We also introduce the “impressions” column, which counts the rows from the raw data with that combination of dimensions:
timestamp publisher advertiser gender country impressions clicks revenue 2011-01-01T01:00:00Z ultratrimfast.com google.com Male USA 1800 25 15.70 2011-01-01T01:00:00Z bieberfever.com google.com Male USA 2912 42 29.18 2011-01-01T02:00:00Z ultratrimfast.com google.com Male UK 1953 17 17.31 2011-01-01T02:00:00Z bieberfever.com google.com Male UK 3194 170 34.01
All is well and good if we content ourselves with computations that can be distributed efficiently such as summing hourly revenue to produce daily revenue, or calculating click-through rates. In the language of Gray et al., the former calculation is distributive: we can sum the raw event prices to produce hourly revenue over each combination of dimensions and in turn sum this intermediary for further coarsening into daily and quarterly totals. The latter is algebraic: it is a combination of a fixed number of distributive statistics, in particular, clicks / impressions.
However, sums and averages are of very little use when one wants to ask certain questions of bid-level data. Exchanges may wish to visualize the bid landscape so as to provide guidance to publishers on how to set floor prices. Because of our data-summarization process, we have lost the individual bid prices–and knowing that the 20 total bids sum to $5 won’t tell us how many exceed $1 or $2. Quantiles, by contrast, are holistic: there is no constant bound on the size of the storage needed to exactly describe a sub-aggregate.
Although the raw data contain the unadulterated prices–with which we can answer these bid landscape questions exactly–let’s recall why we much prefer the summarized dataset. In the above example, each raw row corresponds to an impression, and the summarized data represent an average compression ratio of ~2500:1 (in practice, we see ratios in the 1 to 3 digit range). Less data is both cheaper to store in memory and faster to scan through. In effect, we are trading off increased ETL effort against less storage and faster queries with this pre-aggregation.
One solution to support quantile queries is to store the entire array of ~2500 prices in each row:
timestamp publisher advertiser gender country impressions clicks prices 2011-01-01T01:00:00Z ultratrimfast.com google.com Male USA 1800 25 [0.64, 1.93, 0.93, ...] 2011-01-01T01:00:00Z bieberfever.com google.com Male USA 2912 42 [0.65, 0.62, 0.45, ...] 2011-01-01T02:00:00Z ultratrimfast.com google.com Male UK 1953 17 [0.07, 0.34, 1.23, ...] 2011-01-01T02:00:00Z bieberfever.com google.com Male UK 3194 170 [0.53, 0.92, 0.12, ...]
But the storage requirements for this approach are prohibitive. If we can accept approximate quantiles, then we can replace the complete array of prices with a data structure that is sublinear in storage–similar to our sketch-based approach to cardinality estimation.
Ben-Haim and Tom-Tov suggest summarizing the unbounded-length arrays with a fixed number of (count, centroid) pairs. Suppose we attempt to summarize a set of numbers with a single pair. The mean (centroid) has the nice property of minimizing the sum of the squared differences between it and each value, but it is sensitive to outliers because of the squaring. The median is the minimizer of the sum of the absolute differences and for an odd number of observations, corresponds to an actual bid price. Bid prices tend to be skewed due to the mechanics of second price auctions–some bidders have no problem bidding $100, knowing that they will likely only have to pay $2. So a median of $1 is more representative of the “average” bid price than a mean of $20. However, with the (count, median) representation, there is no way to merge medians: knowing that 8 prices have a median of $.43 and 10 prices have a median of $.59 doesn’t tell you that the median of all 18 prices is $.44. Merging centroids is simple–just use the weighted mean. Given some approximate histogram representation of (count, centroid) pairs, we can make online updates as we scan through data.
Of course, there is no way to accurately summarize an arbitrary number of prices with a single pair, so we are confronted with a classical accuracy/storage/speed tradeoff. We can fix the number of pairs that we store like so:
timestamp publisher advertiser gender country impressions clicks prices 2011-01-01T01:00:00Z ultratrimfast.com google.com Male USA 1800 25 [(1, .16), (48, .62), (83, .71), ...] 2011-01-01T01:00:00Z bieberfever.com google.com Male USA 2912 42 [(1, .12), (3, .15), (30, 1.41), ...] 2011-01-01T02:00:00Z ultratrimfast.com google.com Male UK 1953 17 [(2, .03), (1, .62), (20, .93), ...] 2011-01-01T02:00:00Z bieberfever.com google.com Male UK 3194 170 [(1, .05), (94, .84), (1, 1.14), ...]
In the first row, there is one bid at $.16, 48 bids with an average price of $.62, and so on. But given a set of prices, how do we summarize them as (count, centroid) pairs? This is a special case of the k-means clustering problem, which in general is NP-hard, even in the plane. Fortunately, however, the one-dimensional case is tractable and admits a solution via dynamic programming. The B-H/T-T approach is to iteratively combine the closest two pairs together by taking weighted means until we reach our desired size.
Here we illustrate the B-H/T-T summarization process for the integers 1 through 10, 15 and 20, and 12 and 25 each repeated 3 times, for 3 different choices of the number of (count, centroid) pairs.
There 4 salient operations on these approximate histogram objects:
Adding new values to the histogram: add a new pair, (1, value), and merge the closest pair if we exceed the size parameter
Merging two histograms together: repeatedly add all pairs of values from one histogram to another
Estimating the count of values below some reference value: build trapezoids between the pairs and look at the various areas
Estimating the quantiles of the values represented in a histogram: walk along the trapezoids until you reach the desired quantile We apply operation 1 during our ETL phase, as we group by the dimensions and build a histogram on the resulting prices, serializing this object into a Druid data segment. The compute nodes repeat operation 2 in parallel, each emitting an intermediate histogram to the query broker for combination (another application of operation 2). Finally, we can apply operation 3 repeatedly to estimate counts in between various breakpoints, producing a histogram plot. Or we can estimate quantiles of interest with operation 4.
Here we review the trapezoidal estimation of Ben-Haim and Tom-Tov with an example. Suppose we wanted to estimate the number of values less than or equal to 10 (the exact answer is 10) knowing that there are 10 points with mean 5.5, 4 with mean 12.8, and 4 with mean 23.8. We assume that half of the values lie to the left and half lie to the right (we shall improve upon this assumption in the next section) of the centroid. So we mark off that 5 values are smaller than the first centroid (this turns out to be correct). We then draw a trapezoid connecting the next two centroids and assume that the number of values between 5.5 and 10 is proportional to the area that this sub-trapezoid occupies (the latter half of which is marked in blue). We assume that half of the 10 values near 5.5 lie to its right, and half of the 4 values near 12.8 lie to its left and multiply the sum of 7 by the ratio of areas to come up with our estimate of 5.05 in this region (the exact answer is 5). Therefore, we estimate that there are 10.05 values less than or equal to 10.
Here we describe some improvements and efficiencies specific to our implementation of the B-H/T-T approximate histogram.
Computational efficiency at query time (operation 2) is more dear to us than at ETL time (operation 1). That is, we can spend a few more cycles in building the histograms if it allows for a very efficient means of combination. Our Java-based implementation of operation 2 using a heap to keep track of the differences between pairs can combine roughly 200K (size 50) histograms per second per core (on an i7-3615QM). This compares unfavorably with core scan rates an order of magnitude or two higher for count, sum, and group by queries. Although, to be fair, a histogram contains 1-2 orders of magnitude more information than a single count or sum. Still, we sought a faster solution. If we know ahead of time what the proper threshold below which to merge pairs is, then we can do a linear scan through the sorted pairs (which we can do at ETL time), choosing to merge or not based on the threshold. The exact determination of this threshold is difficult to do efficiently, but eschewing the heap-based solution for this approximation results in core aggregation rates of ~1.3M (size 50) histograms per second.
We have 3 different serialization formats when indexing depending on the nature of the data, for which we use the most efficent encoding:
a dense format, storing all counts and centroids up to the configurable size parameter
a sparse format, storing some number of pairs below the limit
a compact format, storing the individual values themselves
It is important to emphasize that we can specify different levels of accuracy hierarchically. The above formats come into play when we index the data, turning the arrays of raw values into (count, centroid) pairs. Because indexing is slow and expensive and Druid segments are immutable, it’s better at this level to err on the side of accuracy. So we do something like specify a maximum of 100 (count, centroid) pairs in indexing, which will allow for greater flexibility at query time, when we aggregate these together into some possibly different number of (count, centroid) pairs.
We use the superfluous sign bit of the count to determine whether a (count, centroid) pair with count > 1 is exact or not. Does a value of (2, 1.51) indicate 2 bid prices of $1.51, or 2 unequal bid prices that average to $1.51? The trapezoid method of count estimation makes no such distinction and will “spread out” its uncertainty equally. This can be problematic for the discrete, multimodal distributions characteristic of bid data. But given knowledge of which (count, centroid) pairs are exact, we can make more accurate estimates.
Recall that our data typically exhibit high skewness. Because the closest histogram pairs are continuously merged until the number of pairs is small enough, the remaining pairs are necessarily (relatively) far apart. It is logical to summarize 12 prices around $.10 and 6 prices around $.12 as 18 prices around $.11, but we wouldn’t want to merge all prices under $2 because of the influence of 49 wildly-high prices–unless we are particularly interested in those outliers, that is. At the very least, we would like to be able to control our “area of interest”–do we care about the majority of the data or about those few outliers? For when we aggregate millions or billions of values, even with the tiniest skew, we’ll end up summarizing the bulk of the distribution with a single (count, centroid) pair. Our solution is to define special limits, inside of which we maintain the accuracy of our estimates. This typically jives well with setting x-axis limits for our histogram visualization.
Here, we plot a histogram over ~18M prices, using default settings for the x-axis limits and bin widths. Due to the high degree of skew, the inferred limits are suboptimal, as they include prices ~$100. In addition, there are even negative bid prices (which could be erroneous or a way of expressing uninterest in the auction)!
Below, we set our resolution limits to $0 and $1 and vary the number of (count, centroid) pairs in our approximate histogram datastructure. The accuracy using only 5 pairs is abysmal and doesn’t even capture the second mode in the $.20 to $.25 bucket. 50 pairs fare much better, and 200 are very accurate.
Let’s take a look at some benchmarks on our modest demo cluster (4 m2.2xlarge compute nodes) with some wikipedia data. We’ll look at the performance of the following aggregators:
a count aggregator, which simply counts the number of rows
a uniques aggregator, which implements a version of the HyperLogLog algorithm
approximate histogram aggregators, varying the resolution from 10 pairs to 50 pairs to 200 pairs We get about 1-3M summarized rows of data per week from Wikipedia, and the benchmarks over the full 32 week period cover 84M rows. There appears to be a roughly linear relationship between the query time and the quantity of data:
Indeed, the cluster scan rates tend to flatten out once we hit enough data:
We previously obtained cluster scan rates of 26B rows per second on a beefier cluster. Very roughly speaking, the approximate histogram aggregator is 1/10 the speed of the count aggregator, so we might expect speeds of 2-3B rows per second on such a cluster. Recall that our summarization step compacts 10-100 rows of data into 1, for typical datasets. This means that it is possible to construct histograms representing tens to hundreds of billions of prices in seconds.
Finally, my colleague Fangjin Yang and I will continue the discussion in October in New York at the Strata Conference where we will present, “Not Exactly! Fast Queries via Approximation Algorithms.”