Hi, I've been looking into Adrien's suggestion to migrate (Byte/Float)VectorValues to an unabashedly random-access API. We can easily enough support iteration on top of that (which we use extensively during indexing). I think this would represent a great simplification; preliminary implementation shows a big reduction in boilerplate code and awkward casting, extra interfaces, and so on. And it would make binary-partitioning of HNSW graphs more straightforward. But I doubt this would be ready in time for 10.0 so I wouldn't wait for it. Just wanted to let you all know since it would be a breaking change.
On Mon, Sep 9, 2024 at 3:18 PM Luca Cavanna <java...@apache.org> wrote: > > The intra-segment concurrency PR has been pretty close for quite a few days > already. I ran benchmarks last week, made adjustments, and just finished > addressing comments from Mike's review. My plan would be to merge it tomorrow > unless there are objections. > > Regarding the usage of the deprecated search method, I agree that we should > not delay the release for that, yet I hope that we get that done. I will try > to tackle the remaining issues, and Greg has been helping there too, thanks a > lot! For anyone else interested, the description of the issue lists a number > of remaining items that need fixing: > https://github.com/apache/lucene/issues/12892 . > > Cheers > Luca > > On Mon, Sep 9, 2024 at 7:46 PM Adrien Grand <jpou...@gmail.com> wrote: >> >> Thanks to all who replied to this thread and worked on getting these >> blockers addressed. In particular I see that support for JDK 23, backporting >> the Arena work, and the removal of CollectorOwner are merged. >> >> I just reviewed the humongous PR that migrates more classes to records, it >> looks pretty good to me. If someone can look at my comments (hopefully much >> quicker than reviewing the whole PR!), I would appreciate it. >> >> To be transparent, the more usage of the deprecated >> IndexSearcher#search(Query, Collector) we can remove, the better, but I >> don't plan on delaying the release if it's not finished. Likewise, I don't >> plan on delaying 10.0 if support for native dot product is not merged, it >> could make it to a minor release later on if needed. >> >> I haven't taken a deep look at RaBitQ. If it's ready in time for 9.12 and >> 10.0, I'm fine with it getting merged, but as a new feature that doesn't >> look like it requires breaking changes to our public API, it could be >> introduced in a minor release later on, so I don't plan to treat it as a >> blocker. >> >> I would like to get support for intra-segment search concurrency in, as it >> is breaking enough that we could not easily introduce it in a minor later >> on. It seems to be almost ready, so hopefully it will get merged before >> feature freeze next week? >> >> I'm not clear if someone is actively looking into the recall issues with >> 8-bit quantization? >> >> >> On Thu, Sep 5, 2024 at 10:14 PM Shubham Chaudhary <shubhmas...@gmail.com> >> wrote: >>> >>> > Maybe mark it blocker so we don't lose track? >>> >>> Hi Mike, it's already linked to the 10.0 milestone. Is there some way to >>> mark or track it as a blocker for 10.0? It would be great if I could get >>> some reviews on it. The PR has been accumulating merge conflicts over time. >>> I'm happy to address the comments and iterate on it to get this done for >>> the 10.0 release. >>> >>> - Shubham >>> >>> On Wed, Sep 4, 2024 at 7:20 PM Michael McCandless >>> <luc...@mikemccandless.com> wrote: >>>> >>>> On Sat, Aug 31, 2024 at 2:00 PM Shubham Chaudhary <shubhmas...@gmail.com> >>>> wrote: >>>>> >>>>> Hi, regarding the 10.0 release, should we also consider >>>>> https://github.com/apache/lucene/pull/13328. It was planned for 10.0 >>>>> (https://github.com/apache/lucene/issues/13207) and is waiting on review, >>>>> so I think it'll be good if we could consider it. Looking forward to >>>>> views and seeing if there are any concerns with the change I'm unaware of. >>>> >>>> >>>> +1 >>>> >>>> It looks like this one is super close? A couple of rounds of feedback >>>> from Uwe, folded into the PR. Maybe mark it blocker so we don't lose >>>> track? >>>> >>>> Thanks Shubham. >>>> >>>> Mike McCandless >>>> >>>> http://blog.mikemccandless.com >>>> >>>>> >>>>> - Shubham >>>>> >>>>> On Thu, Aug 8, 2024 at 10:20 PM Adrien Grand <jpou...@gmail.com> wrote: >>>>>> >>>>>> Hello everyone, >>>>>> >>>>>> As previously discussed, I plan on releasing 9.last and 10.0 under the >>>>>> following timeline: >>>>>> - ~September 15th: 10.0 feature freeze - main becomes 11.0 >>>>>> - ~September 22nd: 9.last release, >>>>>> - ~October 1st: 10.0 release. >>>>>> >>>>>> Unless someone shortly volunteers to do a 9.x release, this 9.last >>>>>> release will likely be 9.12. >>>>>> >>>>>> As these dates are coming shortly, I would like to start tracking >>>>>> blockers. Please reply to this thread with issues that you know about >>>>>> that should delay the 9.last or 10.0 releases. >>>>>> >>>>>> Chris, Uwe: I also wanted to check with you if this timeline works well >>>>>> with regards to supporting Java 23 in 9.last and 10.0? >>>>>> >>>>>> -- >>>>>> Adrien >> >> >> >> -- >> Adrien --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
