burmanm/gorilla-tsc

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

gorilla-tsc's Issues

Incorrect Maven dependency

Dependency group id is incorrect in the README file, here's the correct one.

<dependency>
	<groupId>fi.iki.yak</groupId>
	<artifactId>compression-gorilla</artifactId>
	<version>1.1.0</version>
</dependency>

Move benchmarks to test part

The Benchmark parts are still residing in the main/ section of the source code, while they should be in the test. This causes the jmh-libraries and all its subdependencies to flow to implementing programs.

Fix travis issues

Travis should not try to use all the maven build modules (such as gpg).

LongArrayOutput#flipByteWithoutExpandCheck runs into ArrayIndexOutOfBoundsException

private void checkAndFlipByte() { // Wish I could avoid this check in most cases... if(bitsLeft == 0) { flipByte(); } }

We only call flipByte() (and in turn expand allocation) when bitsLeft is exactly 0. This might not always be the case. Also when bits > bitsLeft, we call flipByteWithoutExpandCheck and if position has already reached the long array size, then we run into ArrayIndexOutOfBoundsException.

Support splitIterator in Decompressor

Support the Java 8 Stream functionality in Decompressor, by implementing the SplitIterator interface (without the parallelism support)

Allow compressing and retrieving long values

The compression methods themselves work just fine with long values also, thus there the API should be extended to support this use-case also.

Create copy constructors for LongArrayOutput and ByteBufferBitOutput

Provide a mechanism to get an exact copy of each BitOutput implementation.

If a user keeps a reference to the BitOutput implementation passed to a Compressor/GorillaCompressor, this will allow an exact copy to be made, finalized (as if Compressor.close() was called), and passed to a Decompressor/GorillaDecompressor without affecting the original Compressor/GorillaCompressor.

(pull request to follow)

Don't write end stream

There should be support for avoiding to write the end stream as this consumes several bytes per chunk. For chunks with small amount of datapoints, this is large overhead and completely avoidable. Perhaps, make it configurable and of course backwards compatible with older chunks.

Reconsider ZigZag for storing longs

When storing long values, ZigZag encoding should be considered (like with timestamps). While this can be done manually as well outside the library, it could as well be supported in the library.

This improves compression ratio when on negative/positive fluctuating serie (such as -1, 1, -1, 1)

Enforce positive timestamps and maximal gaps

The timestamp of value 0 is somewhat specially treated by the library and used as initialiser for the field storedTimestamp in both compression & decompression.
This leads to a number of strange behaviours if in the middle of a series some timestamp is 0.

It might be a good idea to disallow all non-positive timestamps to protect users from those behaviours.

Also, the same could be applied for timestamps that are too far apart in time.

Support use of only timestamp or value compression

In some cases, the use of only timestamp or only the value compression is preferred. One of these cases could be the use of high precision in timestamps, where this compression method is not at its best.

Make nextTimestamp() and nextValue() public and allow the use of them directly.

Still being maintained? And a few PRs if so

Notice this library is in use in a few OSS projects but the issues are quite old.

I had a few PRs to make but seeing as there isn't activity here in quite a while I was wondering @burmanm if you were interesting in passing the reins for maintenance elsewhere.

PR's I wanted to open:

• Support for 5-bit leading zeros instead of 6 (consistent with original paper and used in some other libraries
• Explicit Float32 support with 1 less bit

Thanks for the effort to create this library it's been quite useful for us!

Why must values be inserted in the increasing time order

In the data structure section of the readme, it says "Values must be inserted in the increasing time order, out-of-order insertions are not supported". If I understand it correctly, it requires the library user to insert data points with strict time order (no out of order data points). Can you elaborate on the reasoning of this? There seems no such limitation of the original paper and I did not see such login in your code either.

Thanks!

Ability to use DFCM predictor

While last value predictor as used in the Gorilla library is often good for monitoring data, it's not very good predictor for many patterns. It should be configurable which predictor is used and implement the DFCM predictor as one opportunity (and turning current predictor to LastValuePredictor).

Also, by allowing to share the predictor instance between the blocks it's possible to use larger prediction tables and learn from previously stored values (instead of relearning always after a new block is used). This of course requires to rebuild the predictor from multiple blocks (instead of making a single block independent), but in theory could allow better compression ratio. This is of course up to the user to decide, this library should not do such decisions.

Gorilla style compression