Distributed Streaming Quantiles
This module implements a count-min sketch-based streaming quantile computation, as described by Cormode and Muthukrishnan.
It is designed with PySpark in mind, but should be suitable with other distributed computing frameworks (e.g., MapReduce).
From the repo:
git clone https://github.com/laserson/dsq.git
cd dsq
python setup.py install
The count-min sketch is a probabilistic data structure, analogous to a Bloom filter that stores counts. The domain 63DB of possible values is split into dyadic intervals (powers of two) up to a specified number of levels. (This also means the domain must be specified in advance.) Each level has its own count-min sketch. When computing a quantile, we estimate the number of observations to the left of the specified quantile (along with the total count of observations). The estimate is built by taking the largest possible intervals from the highest levels of the tree; higher levels are more accurate.
The quantile accumulator requires the specification of a few parameters:
-
lower_bound,upper_bound-- the domain of possible values in the stream -
num_levels-- the depth of the binary tree that partitions the domain -
epsilon,delta-- precision parameters: "the error should be within a factor of epsilon with probability (1-delta)". Note that the higher the demanded precision, the more memory is needed to store the sketch.
Finally, the smallest interval in the dyadic partitioning represents the smallest precision that is attainable.
# Example PySpark usage
# create a SparkContext
# NOTE: you need to ship the dsq.py file around the cluster
sc = SparkContext("spark://your.spark.master", "YourSparkApp",
pyFiles=["/path/to/dsq.py"])
# create an RDD containing a bunch of random numbers
import random
values = sc.parallelize([random.uniform(0, 1) for i in xrange(10000)], 5)
# set parameters for QuantileAccumulator
lower_bound = 0
upper_bound = 1
num_levels = 12
epsilon = 0.001
delta = 0.01
seed = 1729 # optional
# create the accumulator function that will process a partition
from dsq import QuantileAccumulator
accum_fn = QuantileAccumulator(lower_bound, upper_bound, num_levels,
epsilon, delta, seed)
# stream the data through the accumulators
accumulators = values.mapPartitions(accum_fn)
# and merge them all together
quantile_accum = accumulators.reduce(lambda x, y: x.merge(y))
# compute the 95th percentile
quantile_accum.ppf(0.95)