Hello folks,
thank you for your input. I am finished with my investigation regarding introducing special support for FP compression in Apache Parquet. My report also includes an investigation of lossy compressors though there are still some things to be cleared out. Report: https://drive.google.com/open?id=1wfLQyO2G5nofYFkS7pVbUW0-oJkQqBvv Sections 3 4 5 6 are the most important to go over. Let me know if you have any questions or concerns. Regards, Martin ________________________________ From: Zoltan Ivanfi <[email protected]> Sent: Thursday, June 13, 2019 2:16:56 PM To: Parquet Dev Cc: Raoofy, Amir; Karlstetter, Roman Subject: Re: Floating point data compression for Apache Parquet Hi Martin, Thanks for your interest in improving Parquet. Efficient encodings are really important in a big data file format, so this topic is definitely worth researching and personally I am looking forward to your report. Whether to add any new encodings to Parquet, however, can not be answered until we see the results of your findings. You mention two paths. One has very small computational overhead but does not provide significant space savings. The other provides significant space savings but at the price of a significant computational overhead. While purely based on these properties both of them seem "balanced" (one is small effort, small gain; the other is large effort, large gain) and therefore sound reasonable options, I would argue that one should also consider development costs, code complexity and compatibility implications when deciding about whether a new feature is worth implementing. Adding a new encoding or compression to Parquet complicates the specification of the file format and requires implementing it in every language binding of the format, which is not only a considerable effort, but is also error-prone (see LZ4 for an example, which was added to both the Java and the C++ implementation of Parquet, yet they are incompatible with each other). And lack of support is not only a minor annoyance in this case: if one is forced to use an older reader that does not support the new encoding yet (or a language binding that does not support it at all), the data simply can not be read. In my opinion, no matter how low the computational overhead of a new encoding is, if it does not provide significant gains, then the specification clutter, implementation costs and the potential of compatibility problems greatly outweigh its advantages. For this reason, I would say that only encodings that provide significant gains are worth adding. As far as I am concerned, such a new encoding would be a welcome addition to Parquet. Thanks, Zoltan On Wed, Jun 12, 2019 at 11:10 PM Radev, Martin <[email protected]> wrote: > > Dear all, > > thank you for your work on the Apache Parquet format. > > We are a group of students at the Technical University of Munich who would > like to extend the available compression and encoding options for 32-bit and > 64-bit floating point data in Apache Parquet. > The current encodings and compression algorithms offered in Apache Parquet > are heavily specialized towards integer and text data. > Thus there is an opportunity in reducing both io throughput requirements and > space requirements for handling floating point data by selecting a > specialized compression algorithm. > > Currently, I am doing an investigation on the available literature and > publicly available fp compressors. In my investigation I am writing a report > on my findings - the available algorithms, their strengths and weaknesses, > compression rates, compression speeds and decompression speeds, and licenses. > Once finished I will share the report with you and make a proposal which ones > IMO are good candidates for Apache Parquet. > > The goal is to add a solution for both 32-bit and 64-bit fp types. I think > that it would be beneficial to offer at the very least two distinct paths. > The first one should offer fast compression and decompression speed with some > but not significant saving in space. The second one should offer slower > compression and decompression speed but with a decent compression rate. Both > lossless. A lossy path will be investigated further and discussed with the > community. > > If I get an approval from you – the developers – I can continue with adding > support for the new encoding/compression options in the C++ implementation of > Apache Parquet in Apache Arrow. > > Please let me know what you think of this idea and whether you have any > concerns with the plan. > > Best regards, > Martin Radev >
