> Instead of using an existing columnar format like parquet (one file for one type of stats) to store indexes, any reason why we have developed our own format and any benchmarks taken against Puffin vs other formats?
The format needs to store large blobs, which can easily be multiple megabytes for indexes like bloom filters. Formats like Parquet aren't designed for large values like this. Some Parquet writers use page sizes that are on the order of 64kb and the default for the Java implementation is 1 MB. When single values are this large and you want to be able to quickly get to any one value, it makes no sense to batch them up into row groups. Instead, a simple container format makes sense. > How these Puffin files are linked to Iceberg's metadata files is still a missing link for me. Most likely, the initial use of Puffin files will be storing NDV sketches. The Puffin file will be tracked by Snapshot. > do we need to read the previous snapshot's Puffin file and write back a new file with updated stats? No, we will carry stats files forward without modification. Stats can be a little out of date. > do we plan to asynchronously support collecting the stats like "ANALYZE table" and modify the table metadata with the stats file names? (might need an Iceberg commit to write new table metadata) Stats can be handled however engines choose to implement this. ANALYZE is a good option to begin with. > we are looking forward to storing [partition level stats] in the Puffin format. But I'm not sure about storing it as a single file with millions of rows. These aren't designed for millions of rows. Instead, each sketch or index would cover a whole partition. This format is mainly for storing a few large payloads, not tabular data with many rows and columns. Ryan On Mon, Jun 20, 2022 at 11:52 PM Ajantha Bhat <ajanthab...@gmail.com> wrote: > Thank you Piotr for all of the work you’ve put into this. > > I just checked the spec. I have a few newbie questions. > > a. Instead of using an existing columnar format like parquet (one file for > one type of stats) to store indexes, any reason why we have developed our > own format and any benchmarks taken against Puffin vs other formats? > > b. How these Puffin files are linked to Iceberg's metadata files is still > a missing link for me. As the Puffin spec says, these stats are table level > (updated per snapshots). So, do we need an Iceberg spec change to store the > file names of these Puffin files so that remove_orphan_files will not > clean it up accidentally? (also needed for expire_snapshots) > > c. NDV's are column level stats. So, I expect the latest puffin file of > that snapshot will have one row of stats representing stats for each > column. But if we are to implement secondary index or table level partition > stats, there can be many rows (millions) in puffin based on the dataset. > So, for every commit, do we need to read the previous snapshot's Puffin > file and write back a new file with updated stats? (the file might be very > huge when data grows?). I think it will affect the commit time. Any > thoughts on this? > > d. Slightly related to the above point, do we plan to asynchronously > support collecting the stats like "ANALYZE table" and modify the table > metadata with the stats file names? (might need an Iceberg commit to write > new table metadata) > > e. Even though table level partition stats are available from _parition > metadata table (along with filter push down support), computing metadata > table per query will be expensive. > Hence, we are looking forward to storing them in the Puffin format. But > I'm not sure about storing it as a single file with millions of rows. > I Would like to collaborate and discuss more on this. > > Thanks, > Ajantha > > On Mon, Jun 13, 2022 at 2:45 AM Miao Wang <miw...@adobe.com.invalid> > wrote: > >> +1 on the format! It looks great! >> >> >> >> Thanks for materializing the initial design idea. >> >> >> >> Miao >> >> *From: *Kyle Bendickson <kjbendick...@gmail.com> >> *Date: *Sunday, June 12, 2022 at 1:55 PM >> *To: *dev@iceberg.apache.org <dev@iceberg.apache.org> >> *Subject: *Re: [VOTE] Adopt Puffin format as a file format for >> statistics and indexes >> >> *EXTERNAL: Use caution when clicking on links or opening attachments.* >> >> >> >> +1 [non-binding] >> >> >> >> Thank you Piotr for all of the work you’ve put into this. >> >> >> >> This should greatly benefit not only Iceberg on Trino, but hopefully can >> be used in many novel ways due to its well thought out generic design and >> incorporation of the ability to extend with new sketches. >> >> >> >> Looking forward to the improvements this will bring. >> >> >> >> - Kyle >> >> >> >> On Fri, Jun 10, 2022 at 1:47 PM Alexander Jo <alex...@starburstdata.com> >> wrote: >> >> +1, let's do it! >> >> >> >> On Fri, Jun 10, 2022 at 2:47 PM John Zhuge <jzh...@apache.org> wrote: >> >> +1 Looking forward to the features it enables. >> >> >> >> On Fri, Jun 10, 2022 at 10:11 AM Yufei Gu <flyrain...@gmail.com> wrote: >> >> +1. Looking forward to the partition stats. >> >> Best, >> >> >> >> Yufei >> >> >> >> >> >> On Thu, Jun 9, 2022 at 6:32 PM Daniel Weeks <dwe...@apache.org> wrote: >> >> +1 as well. Excited about the progress here. >> >> >> >> -Dan >> >> On Thu, Jun 9, 2022, 6:25 PM Junjie Chen <chenjunjied...@gmail.com> >> wrote: >> >> +1, really nice! Indexes are coming! >> >> >> >> On Fri, Jun 10, 2022 at 8:04 AM Szehon Ho <szehon.apa...@gmail.com> >> wrote: >> >> +1, it's an exciting step for Iceberg, look forward to all the new >> statistics and secondary indices it will allow. >> >> >> >> Had a few questions of what the reference to Puffin file(s) will be in >> the Iceberg spec, but it's orthogonal to Puffin file format itself. >> >> >> >> Thanks, >> >> Szehon >> >> >> >> On Thu, Jun 9, 2022 at 3:32 PM Ryan Blue <b...@tabular.io> wrote: >> >> +1 from me! >> >> >> >> There may also be people that haven't followed the design discussions and >> we can start a DISCUSS thread if needed. But if everyone is comfortable >> with the design and implementation, I think it's ready for a vote as well. >> >> >> >> Huge thanks to Piotr for getting this ready! I think the format is going >> to be really useful for both stats and indexes in Iceberg. >> >> >> >> On Thu, Jun 9, 2022 at 3:35 AM Piotr Findeisen <pi...@starburstdata.com> >> wrote: >> >> Hi Everyone, >> >> I propose that we adopt Puffin file format as a file format for >> statistics and indexes in Iceberg tables. >> >> >> >> Puffin file format specification: >> >> https://github.com/apache/iceberg/blob/master/format/puffin-spec.md >> <https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Ficeberg%2Fblob%2Fmaster%2Fformat%2Fpuffin-spec.md&data=05%7C01%7Cmiwang%40adobe.com%7Cba30cde28d1b4e3abe5108da4cb5ef83%7Cfa7b1b5a7b34438794aed2c178decee1%7C0%7C0%7C637906641543835876%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=3Y04jqMw6ZIc%2BojDmWlpOeLL5zQ3YvLcdAgoHJTwL8c%3D&reserved=0> >> >> (previous discussions: https://github.com/apache/iceberg/pull/4944 >> <https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Ficeberg%2Fpull%2F4944&data=05%7C01%7Cmiwang%40adobe.com%7Cba30cde28d1b4e3abe5108da4cb5ef83%7Cfa7b1b5a7b34438794aed2c178decee1%7C0%7C0%7C637906641543835876%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=tWuoyTfEaIWmOFivROQRt0fD1KRYc%2FqwRO2KoZhIoi8%3D&reserved=0> >> , https://github.com/apache/iceberg-docs/pull/69 >> <https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Ficeberg-docs%2Fpull%2F69&data=05%7C01%7Cmiwang%40adobe.com%7Cba30cde28d1b4e3abe5108da4cb5ef83%7Cfa7b1b5a7b34438794aed2c178decee1%7C0%7C0%7C637906641543835876%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=Uf8XiuLSLEO8YtCMkk%2BSXWS6lefw95O22K844P5Iovc%3D&reserved=0> >> ) >> >> >> >> Intend use: >> >> * statistics in Iceberg tables (see >> https://github.com/apache/iceberg/pull/4945 >> <https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Ficeberg%2Fpull%2F4945&data=05%7C01%7Cmiwang%40adobe.com%7Cba30cde28d1b4e3abe5108da4cb5ef83%7Cfa7b1b5a7b34438794aed2c178decee1%7C0%7C0%7C637906641543835876%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=swByVgNPD6lbSlJjHIZZX4jgeVzC%2BT%2BWUvxrrg0Wpx8%3D&reserved=0> >> and associated proposed implementation >> https://github.com/apache/iceberg/pull/4741 >> <https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Ficeberg%2Fpull%2F4741&data=05%7C01%7Cmiwang%40adobe.com%7Cba30cde28d1b4e3abe5108da4cb5ef83%7Cfa7b1b5a7b34438794aed2c178decee1%7C0%7C0%7C637906641543835876%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=dYckyv1f36iQqs9%2FaRQRsumtB2xEmwcFJAQihYZRYlw%3D&reserved=0> >> ) >> >> * in the future: storage for secondary indexes >> >> >> >> Puffin file reader and writer implementation: >> >> https://github.com/apache/iceberg/pull/4537 >> <https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fapache%2Ficeberg%2Fpull%2F4537&data=05%7C01%7Cmiwang%40adobe.com%7Cba30cde28d1b4e3abe5108da4cb5ef83%7Cfa7b1b5a7b34438794aed2c178decee1%7C0%7C0%7C637906641543835876%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=YY%2B52Eq%2FcnnseM5Nd4E0D3Xw8IWMsD4QaI98LXFMu9c%3D&reserved=0> >> >> >> >> Thanks, >> >> PF >> >> >> >> >> >> >> -- >> >> Ryan Blue >> >> Tabular >> >> >> >> >> -- >> >> Best Regards >> >> >> >> >> -- >> >> John Zhuge >> >> -- Ryan Blue Tabular
