Thanks, Szehon. To clarify on compatibility, using the same format for the blobs makes it so that existing Delta readers can read and use the DVs written by Iceberg. I'd love for Delta to adopt Puffin, but if we adopt the extra fields they would not need to change how readers work. That's why I think there is a benefit to using the same format. We avoid fragmentation and make sure data and delete files are compatible. No unnecessary fragmentation.
Ryan On Tue, Oct 15, 2024 at 10:57 AM Szehon Ho <szehon.apa...@gmail.com> wrote: > This is awesome work by Anton and Ryan, it looks like a ton of effort has > gone into the V3 position vector proposal to make it clean and efficient, a > long time coming and Im truly excited to see the great improvement in > storage/perf. > > wrt to these fields, I think most of the concerns are already mentioned by > the other community members in the prs > https://github.com/apache/iceberg/pull/11238 and > https://github.com/apache/iceberg/pull/11238, so not much to add. The DV > itself is RoaringBitmap 64-bit format so that's great, the argument for CRC > seems reasonable, and I dont have enough data to be opinionated towards > endian/magic byte. > > But I do lean towards the many PR comments that the extra length field is > unnecessary, and would just add confusion. It seemed to me that the > Iceberg community has made so much effort to trim to spec to the bare > minimum for cleanliness and efficiency, so I do feel the field is not in > the normal direction of the project. Also Im not clear on the plan for old > Delta readers, they cant read Puffin anyway, if Delta adopts Puffin, then > new readers could adopt? Anyway great work again, thanks for raising the > issue on devlist! > > Thanks, > Szehon > > On Mon, Oct 14, 2024 at 5:14 PM rdb...@gmail.com <rdb...@gmail.com> wrote: > >> > I think it might be worth mentioning the current proposal makes some, >> mostly minor, design choices to try to be compatible with Delta Lake >> deletion vectors. >> >> Yes it does, and thanks for pointing this out, Micah. I think it's >> important to consider whether compatibility is important to this community. >> I just replied to Piotr on the PR, but I'll adapt some of that response >> here to reach the broader community. >> >> I think there is value in supporting compatibility with older Delta >> readers, but I acknowledge that this may be my perspective because my >> employer has a lot of Delta customers that we are going to support now and >> in the future. >> >> The main use case for maintaining compatibility with the Delta format is >> that it's hard to move old jobs to new code in a migration. We see a >> similar issue in Hive to Iceberg migrations, where unknown older readers >> prevent migration entirely because they are hard to track down and often >> read files directly from the backing object store. I'd like to avoid the >> same problem here, where all readers need to be identified and migrated at >> the same time. Compatibility with the format those readers expect makes it >> possible to maintain Delta metadata for them temporarily. That increases >> confidence that things won't randomly break and makes it easier to get >> people to move forward. >> >> The second reason for maintaining compatibility is that we want for the >> formats to become more similar. My hope is that we can work across both >> communities and come up with a common metadata format in a future version >> -- which explains my interest in smooth migrations. Maintaining >> compatibility in cases like this builds trust and keeps our options open: >> if we have compatible data layers, then it's easier to build a compatible >> metadata layer. I'm hoping that if we make the blob format compatible, we >> can get the Delta community to start using Puffin for better >> self-describing delete files. >> >> Other people may not share those goals, so I think it helps to consider >> what is being compromised for this compatibility. I don't think it is too >> much. There are 2 additional fields: >> * A 4-byte length field (that Iceberg doesn't need) >> * A 4-byte CRC to validate the contents of the bitmap >> >> There are also changes to how these would have been added if the Iceberg >> community were building this independently. >> * Our initial version didn't include a CRC at all, but now that we think >> it's useful compatibility means using a CRC-32 checksum rather than a newer >> one >> * The Delta format uses big endian for its fields (or mixed endian if you >> consider RoaringBitmap is LE) >> * The magic bytes (added to avoid reading the Puffin footer) would have >> been different >> >> Overall, I don't think that those changes to what we would have done are >> unreasonable. It's only 8 extra bytes and half of them are for a checksum >> that is a good idea. >> >> I'm looking forward to what the rest of the community thinks about this. >> Thanks for reviewing the PR! >> >> Ryan >> >> >> On Sun, Oct 13, 2024 at 10:45 PM Jean-Baptiste Onofré <j...@nanthrax.net> >> wrote: >> >>> Hi >>> >>> Thanks for the PRs ! I reviewed Anton's document, I will do a pass on >>> the PRs. >>> >>> Imho, it's important to get feedback from query engines, as, if delete >>> vectors is not a problem per se (it's what we are using as internal >>> representation), the use of Puffin files to store it is "impactful" >>> for the query engines (probably some query engines might need to >>> implement Puffin spec (read/write) using other language than Java, for >>> instance Apache Impala). >>> >>> I like the proposal, I just hope we won't "surprise" some query >>> engines with extra work :) >>> >>> Regards >>> JB >>> >>> On Thu, Oct 10, 2024 at 11:41 PM rdb...@gmail.com <rdb...@gmail.com> >>> wrote: >>> > >>> > Hi everyone, >>> > >>> > There seems to be broad agreement around Anton's proposal to use >>> deletion vectors in Iceberg v3, so I've opened two PRs that update the spec >>> with the proposed changes. The first, PR #11238, adds a new Puffin blob >>> type, delete-vector-v1, that stores a delete vector. The second, PR #11240, >>> updates the Iceberg table spec. >>> > >>> > Please take a look and comment! >>> > >>> > Ryan >>> >>
