Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/burmanm/gorilla-tsc

Implementation of time series compression method from the Facebook's Gorilla paper
https://github.com/burmanm/gorilla-tsc

compression java time-series timeseries

Last synced: 2 days ago
JSON representation

Implementation of time series compression method from the Facebook's Gorilla paper

Awesome Lists containing this project

README 

        
= Time series compression library, based on the Facebook's Gorilla paper

:source-language: java



ifdef::env-github[]

[link=https://travis-ci.org/burmanm/gorilla-tsc]

image::https://travis-ci.org/burmanm/gorilla-tsc.svg?branch=master[Build Status,70,18]

[link=https://maven-badges.herokuapp.com/maven-central/fi.iki.yak/compression-gorilla]

image::https://img.shields.io/maven-central/v/fi.iki.yak/compression-gorilla.svg[Maven central]

endif::[]



== Introduction



This is Java based implementation of the compression methods described in the paper link:http://www.vldb.org/pvldb/vol8/p1816-teller.pdf["Gorilla: A Fast, Scalable, In-Memory Time Series Database"]. For explanation on how the compression methods work, read the excellent paper.



In comparison to the original paper, this implementation allows using both integer values (`long`) as well as

floating point values (`double`), both 64 bit in length.



Versions 1.x and 2.x are not compatible with each other due to small differences to the stored array. Versions 2.x

will support reading and storing older format also, see usage for more details.



== Usage



The included tests are a good source for examples.



=== Maven



[source, xml]

----

    

        fi.iki.yak

        compression-gorilla

    

----



You can find latest version from the maven logo link above.



=== Compressing



To compress in the older 1.x format, use class ``Compressor``. For 2.x, use ``GorillaCompressor`` (recommended).

``LongArrayOutput`` is also recommended compared to ``ByteBufferBitOutput`` because of performance. One can supply

alternative predictor to the ``GorillaCompressor`` if required. One such implementation is included,

``DifferentialFCM`` that provides better compression ratio for some data patterns.



[source, java]

----

long now = LocalDateTime.now(ZoneOffset.UTC).truncatedTo(ChronoUnit.HOURS)

        .toInstant(ZoneOffset.UTC).toEpochMilli();



LongArrayOutput output = new LongArrayOutput();

GorillaCompressor c = new GorillaCompressor(now, output);

----



Compression class requires a block timestamp and an implementation of `BitOutput` interface.



[source, java]

----

c.addValue(long, double);

----



Adds a new floating-point value to the time series. If you wish to store only long values, use `c.addValue(long,

long)`, however do `not` mix these in the same series.



After the block is ready, remember to call:



[source, java]

----

c.close();

----



which flushes the remaining data to the stream and writes closing information.



=== Decompressing



To decompress from the older 1.x format, use class ``Decompressor``. For 2.x, use ``GorillaDecompressor`` (recommended).

``LongArrayInput`` is also recommended compared to ``ByteBufferBitInput`` because of performance if the 2.x

format was used to compress the time series. If the original compressor used different predictor than

``LastValuePredictor`` it must be defined in the constructor.



[source, java]

----

LongArrayInput input = new LongArrayInput(byteBuffer);

GorillaDecompressor d = new GorillaDecompressor(input);

----



To decompress a stream of bytes, supply `GorillaDecompressor` with a suitable implementation of `BitInput` interface.

 The LongArrayInput allows to decompress a long array or existing `ByteBuffer` presentation with 8 byte word

 length.



[source, java]

----

Pair pair = d.readPair();

----



Requesting next pair with `readPair()` returns the following series value or a `null` once the series is completely

read. The pair is a simple placeholder object with `getTimestamp()` and `getDoubleValue()` or `getLongValue()`.



== Performance



The following performance in reached in a Linux VM running on VMware Player in Windows 8.1 host. i7 2600K at 4GHz.

The benchmark used is the ``EncodingBenchmark``. These results should not be directly compared to other

implementations unless similar dataset is used.



Results are in millions of datapoints (timestamp + value) pairs per second. The values in this benchmark are

in doubles (performance with longs is slightly higher, around ~2-3M/s).



.Compression

|===

|GorillaCompressor (2.0.0) |Compressor (1.1.0)



|83.5M/s (~1.34GB/s)

|31.2M/s (~499MB/s)

|===



.Decompression

|===

|GorillaDecompressor (2.0.0) |Decompressor (1.1.0)



|77,9M/s (~1.25GB/s)

|51.4M/s (~822MB/s)

|===



Most of the differences in decompression / compression speed between versions come from implementation changes and

not from the small changes to the output format.



== Roadmap



There were few things I wanted to get to 2.0.0, but had to decide against due to lack of time. I will implement these

 later with potentially some breaking API changes:



 * Support timestamp only compressions (2.2.x)

 * Include ByteBufferLongOutput/ByteBufferLongInput in the package (2.2.x)

 * Move bit operations to inside the GorillaCompressor/GorillaDecompressor to allow easier usage with

 other allocators (2.2.x)



== Internals



=== Differences to the original paper



* Maximum number of leadingZeros is stored with 6 bits to allow up to 63 leading zeros, which are necessary when

storing long values. (>= 2.0.0)

* Timestamp delta-of-delta are stored by first turning them with ZigZag encoding to positive integers and then

reduced by one to fit in the necessary bits. In the decoding phase all the values are incremented by one to fetch the

 original value. (>= 2.0.0)

* The compressed blocks are created with a 27 bit delta header (unlike in the original paper, which uses a 14 bit delta

  header). This allows to use up to one day block size using millisecond precision. (>= 1.0.0)



=== Data structure



Values must be inserted in the increasing time order, out-of-order insertions are not supported.



The included ByteBufferBitInput and ByteBufferBitOutput classes use a big endian order for the data.



== Contributing



File an issue and/or send a pull request.



=== License



....

   Copyright 2016-2018 Michael Burman and/or other contributors.



   Licensed under the Apache License, Version 2.0 (the "License");

   you may not use this file except in compliance with the License.

   You may obtain a copy of the License at



       http://www.apache.org/licenses/LICENSE-2.0



   Unless required by applicable law or agreed to in writing, software

   distributed under the License is distributed on an "AS IS" BASIS,

   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

   See the License for the specific language governing permissions and

   limitations under the License.

....