cube-js / datasketches-rs

Rusty wrapper for Apache DataSketches

DataSketches in Rust

A Rust binding for the Apache DataSketches library and command-line tool.

On the command-line, we provide

• dsrs [--key] [--raw] [--merge] for approximate distinct line-counting, and
• dsrs --hh k for heavy hitters (approximate most frequent lines).

For instance, the following experiment checks how many unique lines exist when you print all numbers up to 100M twice.

m100=$((100 * 1000 * 1000))
(seq $m100 && seq $m100) | \
  /usr/bin/time -f "%e sec %M KB" dsrs
102055590
5.22 sec 4288 KB

(seq $m100 && seq $m100) | \
  /usr/bin/time -f "%e sec %M KB" sort -u | wc -l
438.66 sec 12880 KB
100000000

(seq $m100 && seq $m100) | \
  /usr/bin/time -f "%e sec %M KB" awk '{a[$0]=1}END{print length(a)}'
100000000
39.28 sec 898240 KB

Next, we can ask for the most popular lines from a stream (there is a topfew Rust package, but it does not support streams).

m10=$((10 * 1000 * 1000))
seq $m10 | sed 's/$/\n1\n2\n3/' | \
  /usr/bin/time -f "%e sec %M KB" sort | \
  uniq -c | sort -rn | head -3
54.88 sec 8968 KB
10000001 3
10000001 2
10000001 1
  
# exact hashmap solution, requires go
pushd /tmp && \
  (test -d topfew || git clone [email protected]:timbray/topfew.git topfew) && \
  pushd topfew && make && popd && popd
seq $m10 | sed 's/$/\n1\n2\n3/' | \
  /usr/bin/time -f "%e sec %M KB" /tmp/topfew/bin/tf -f 1 -n 3
10000001 2
10000001 3
10000001 1
10.67 sec 1060332 KB
  
seq $m10 | sed 's/$/\n1\n2\n3/' | \
  /usr/bin/time -f "%e sec %M KB" target/release/dsrs --hh 3
10000001 2
10000001 1
10000001 3
4.48 sec 4560 KB

Here's a sophisticated example of the tool in action, used to compute rolling average active reviewers for Amazon over a couple decades. The equivalent non-sketch based solution OOMs. Similarly, we can use dsrs --hh to extract the most popular SciHub downloads using multiple orders of magnitude less memory than an exact solution.

Installation

Assumes a modern Rust cargo is installed. The command line tool dsrs can be installed with:

cargo install dsrs

The library may be used as a regular Rust dependency by adding it to your Cargo.toml file.

Embedded C++ Library

This Rust library contains manually-copied header files from the header-only datasketches-cpp library at commit 043b947f.

This was done by extracting all headers. Assuming you're in the datasketches-rs directory, which has a sibling datasketches-cpp:

# make all required directories
find ../datasketches-cpp/ -name "*.h" -or -name "*.hpp" | \
  xargs dirname | \
  sort -u |
  cut -d/ -f2- | \
  xargs mkdir -p
# copy over the actual headers
find ../datasketches-cpp/ -name "*.h" -or -name "*.hpp" | \
  cut -d/ -f2- | \
  xargs -I {} cp ../{} {}
# and the license info too
cp ../datasketches-cpp/{NOTICE,LICENSE} datasketches-cpp/

# some manual interventions were required for the heavy hitters
# implementation, which requires the C++ side to temporarily own
# keys from Rust, so additional management code needs to be injected.
git apply fi.patch
git grep -l "uint16_t DRIFT_LIMIT = [0-9]*;" | xargs sed -i 's/uint16_t DRIFT_LIMIT = [0-9]*;/uint32_t DRIFT_LIMIT = 1024 * 1024 * 1024;/'

This is all only possible thanks to the excellent dtolnay/cxx library!

Why DataSketches in Rust?

There are quite a few crates containing HyperLogLog sketches. However, when I attempted to use them (as of 2021-06-20), I found that their APIs panicked on certain inputs (e.g., try amadeus_streaming::HyperLogLog::<u64>::new(0.0001);), or did not have merge operations. A very rudimentary cargo criterion on 1M unique keys finds that CPC has better accuracy anyway (for all of the below, the same nominal accuracy configuration was set, so these should be using roughly the same amount of memory):

repeat-ten/dsrs::CpcSketch/1000000
                        time:   [144.95 ms 149.27 ms 155.42 ms]
repeat-ten/amadeus_streaming::HyperLogLog/1000000
                        time:   [132.89 ms 134.01 ms 135.49 ms]
repeat-ten/probabilistic_collections::HyperLogLog/1000000
                        time:   [159.99 ms 165.94 ms 172.32 ms]
repeat-ten/probably::HyperLogLog/1000000
                        time:   [119.47 ms 123.95 ms 127.84 ms]
repeat-ten/hyperloglogplus::HyperLogLogPlus/1000000
                        time:   [120.74 ms 121.32 ms 122.10 ms]

relative errors
size: 1000000
  relerr:   1.1% name: dsrs::CpcSketch
  relerr:   3.3% name: amadeus_streaming::HyperLogLog
  relerr:   4.3% name: hyperloglogplus::HyperLogLogPlus
  relerr:  50.7% name: probably::HyperLogLog
  relerr:   inf% name: probabilistic_collections::HyperLogLog

while overall update speed doesn't change too much between implementations.