Hi Martin, I agree that bs_zstd would be a good place to start. Regarding the choice of language, Java, C++ and Python are your options. As far as I know, the Java implementation of Parquet has more users from the business sector, where decimal is preferred over floating point data types. It is also much more tightly integrated with the Hadoop ecosystem (it is even called parquet-mr, as in MapReduce), making for a steeper learning curve.
The Python and C++ language bindings have more scientific users, so users of these may be more interested in the new encodings. Python is a good language for rapid prototyping as well, but the Python binding of Parquet may use the C++ library under the hood, I'm not sure (I'm more familiar with the Java implementation). In any case, there are at least two Python bindings: pyarrow and fastparquet. I think we can extend the format before the actual implementations are ready, provided that the specification is clear and nobody objects to adding it to the format. For this, I would wait for the opinion of a few more Parquet developers first, since changes to the format that are only supported by a single committer usually have a hard time getting into the spec. Additionally, could you please clarify which language bindings you plan to implement yourself? This will help the developers of the different language bindings assess how much work they will have to do to add support. Thanks, Zoltan On Fri, Jul 5, 2019 at 4:34 PM Radev, Martin <martin.ra...@tum.de> wrote: > Hello Zoltan and Parquet devs, > > > do you think it would be appropriate to start with a Parquet prototype > from my side? > > I suspect that integrating 'bs_zstd' would be the simplest to > integrate and from the report we can see an improvement in both ratio and > speed. > > > Do you think that Apache Arrow is an appropriate place to prototype the > extension of the format? > > Do you agree that the enum field 'Encodings' is a suitable place to add > the 'Byte stream-splitting transformation'? In that way it could be used > with any of the other supported compressors. > > It might be best to also add a Java implementation of the transformation. > Would the project 'parquet-mr' be a good place? > > > Would the workflow be such that I write my patches, we verify for > correctness, get reviews, merge them AND just then we make adjustments to > the Apache Parquet spec? > > > Any piece of advice is welcome! > > > Regards, > > Martin > > > ------------------------------ > *From:* Zoltan Ivanfi <z...@cloudera.com> > *Sent:* Friday, July 5, 2019 4:21:39 PM > *To:* Radev, Martin > *Cc:* Parquet Dev; Raoofy, Amir; Karlstetter, Roman > *Subject:* Re: Floating point data compression for Apache Parquet > > > Hi Martin, > > Thanks for the explanations, makes sense. Nice work! > > Br, > > Zoltan > > On Thu, Jul 4, 2019 at 12:22 AM Radev, Martin <martin.ra...@tum.de> wrote: > >> Hello Zoltan, >> >> >> *> **Is data pre-loaded to RAM before making the measurements?* >> Yes, the file is read into physical memory. >> >> For mmap-ed files, read from external storage, I would expect, but not >> 100% sure, that the IO-overhead would be big enough that all algorithms >> compress quite close at the same speed. >> >> >> >> *>In "Figure 3: Decompression speed in MB/s", is data size measured >> before or after uncompression? * >> *> In "Figure 4: Compression speed in MB/s", is data size measured before >> or after compression?* >> For both the reported result is "size of the original file / time to >> compress or decompress". >> >> *> **According to "Figure 3: Decompression speed in MB/s", decompression >> of bs_zstd is almost twice as fast as plain zstd. Do you know what causes >> this massive speed improvement?* >> >> I do not know all of the details. As you mentioned, the written out data >> is less, this could potentially lead to improvement in speed as less data >> has to be written out to memory during compression or read from memory >> during decompression. >> >> Another thing to consider is that ZSTD uses different techniques to >> compress a block of data - "raw", "RLE", "Huffman coding", "Treeless >> coding". >> >> I expect that "Huffman coding" is more costly than "RLE" and I also >> expect that "RLE" to be applicable for the majority of the sign bits thus >> leading to a performance win for when the transformation is applied. >> >> I also expect that zstd has to do some form of "optimal parsing" to >> decide how to process the input in order to compress it well. This is >> something every wanna-be-good LZ-like compressor has to do ( >> https://martinradev.github.io/jekyll/update/2019/05/29/writing-a-pe32-x86-exe-packer.html >> , >> http://cbloomrants.blogspot.com/2011/10/10-24-11-lz-optimal-parse-with-star.html >> ). >> It might be so that the transformed input is somehow easy which leads to >> faster compression rates and also easier to decompress data which leads to >> faster decompression rates. >> >> I used this as a reference: >> https://www.rfc-editor.org/rfc/pdfrfc/rfc8478.txt.pdf. I am not familiar >> with ZSTD in particular. >> >> I also checked that the majority of the time is spent in zstd. >> >> Example run for msg_sweep3d.dp using zstd at level 1. >> - Transformation during compression: 0.086s, ZSTD compress on transformed >> data: 0.08s >> >> - regular ZSTD: 0.34s >> - ZSTD decompress from compressed transformed data: 0.067s, Transformation >> during decompression: 0.021s >> - regular ZSTD decompress: 0.24s >> >> >> Example run for msg_sweep3d.dp using zstd at level 20. >> >> - Transformation during compression: 0.083s, ZSTD compress on transformed >> data: 14.35s >> >> - regular ZSTD: 183s >> - ZSTD decompress from compressed transformed >> data: 0.075s, Transformation during decompression: 0.022s >> - regular ZSTD decompress: 0.31s >> Here it's clear that the transformed input is easier to parse (compress). >> Maybe also the blocks are of type which takes less time to decompress. >> >> *> **If considering using existing libraries to provide any of the >> compression algorithms, license compatibility is also an important factor >> and therefore would be worth mentioning in Section 5.* >> This is something I forgot to list. I will back to you and the other devs >> with information. >> >> The filter I proposed for lossless compression can be integrated without >> any concerns for a license. >> >> >> *> **Are any of the investigated strategies applicable to DECIMAL >> values?* >> The lossy compressors SZ and ZFP do not support that outside of the box. >> I could communicate with the SZ developers to come to a decision how this >> can be added to SZ. An option is to losslessly compress the pre-decimal >> number and lossyly compress the post-decimal number. >> >> For lossless compression, we can apply a similar stream splitting >> technique for decimal types though it might be somewhat more complex and I >> have not really though about this case. >> >> >> Regards, >> >> Martin >> ------------------------------ >> *From:* Zoltan Ivanfi <z...@cloudera.com> >> *Sent:* Wednesday, July 3, 2019 6:07:50 PM >> *To:* Parquet Dev; Radev, Martin >> *Cc:* Raoofy, Amir; Karlstetter, Roman >> *Subject:* Re: Floating point data compression for Apache Parquet >> >> Hi Martin, >> >> Thanks for the thorough investigation, very nice report. I would have a >> few questions: >> >> - Is data pre-loaded to RAM before making the measurements? >> >> - In "Figure 3: Decompression speed in MB/s", is data size measured >> before or after uncompression? >> >> - In "Figure 4: Compression speed in MB/s", is data size measured before >> or after compression? >> >> - According to "Figure 3: Decompression speed in MB/s", decompression of >> bs_zstd is almost twice as fast as plain zstd. Do you know what causes this >> massive speed improvement? Based on the description provided in section >> 3.2, bs_zstd uses the same zstd compression with an extra step of >> splitting/combining streams. Since this is extra work, I would have >> expected bs_zstd to be slower than pure zstd, unless the compressed data >> becomes so much smaller that it radically improves data access times. >> However, according to "Figure 2: Compression ratio", bs_zstd achieves >> "only" 23% better compression than plain zstd, which can not be the reason >> for the 2x speed-up in itself. >> >> - If considering using existing libraries to provide any of the >> compression algorithms, license compatibility is also an important factor >> and therefore would be worth mentioning in Section 5. >> >> - Are any of the investigated strategies applicable to DECIMAL values? >> Since floating point values and calculations have an inherent inaccuracy, >> the DECIMAL type is much more important for storing financial data, which >> is one of the main use cases of Parquet. >> >> Thanks, >> >> Zoltan >> >> On Mon, Jul 1, 2019 at 10:57 PM Radev, Martin <martin.ra...@tum.de> >> wrote: >> >>> 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 <z...@cloudera.com.INVALID> >>> 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 <martin.ra...@tum.de> >>> 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 >>> > >>> >>
