Skip to content
Implementation of time series compression method from the Facebook's Gorilla paper
Branch: master
Clone or download


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

Build Status
Maven central


This is Java based implementation of the compression methods described in the paper "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.


The included tests are a good source for examples.



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


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.

long now =

LongArrayOutput output = new LongArrayOutput();
GorillaCompressor c = new GorillaCompressor(now, output);

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

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:


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


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.

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.

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().


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).

Table 1. Compression
GorillaCompressor (2.0.0) Compressor (1.1.0)

83.5M/s (~1.34GB/s)

31.2M/s (~499MB/s)

Table 2. 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.


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)


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.


File an issue and/or send a pull request.


   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

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   See the License for the specific language governing permissions and
   limitations under the License.
You can’t perform that action at this time.