My vote would be that things in there should "generally work" in that we'll try to address obvious bugs, but not really provide any real support. No assurances of binary serialization stability, for instance. So clone and use at your own risk.
But along with that, I think we should consider a "promote or remove" policy, so it doesn't become a graveyard of abandoned ideas or a way to avoid the last mile challenges of real releases. Anything that's been there for N months with no activity either gets moved to the library or killed off. But maybe that causes too many of its own issues? jon On Thu, Feb 6, 2020, 7:11 PM Kenneth Knowles <[email protected]> wrote: > Thanks for the summary! > > Smart move to have an experimental repository. Do you release the > works-in-progress there? Rephrased: how do people try them out? Do they > have to clone & build? > > Kenn > > On Thu, Feb 6, 2020 at 10:52 AM leerho <[email protected]> wrote: > >> Folks, >> >> This is a brief summary of our Science meeting of February 4th. >> >> Most of the discussion revolved around theoretical approaches for a >> relative-error quantiles sketch. There is certainly user interest in such >> a sketch if it could be practically implemented. Apparently, there are >> several research groups also interested in solving this problem, however, >> to date, there is not consensus on the best algorithmic approach nor is >> there consensus on what the theoretical error bounds would be. >> >> We did briefly discuss the possibility of re-creating a >> datasketches-experimental repository for algorithms like this that are >> still experimental and not ready for production. This would allow early >> adopters to experiment with the algorithms and provide feedback. We had >> such a repository in the github.com/datasketches organization, but as of >> yet have not seen the need to move it to Apache. If and when this new >> quantiles algorithm becomes available in the future it will be a good >> candidate to place in an experimental repository. >> >> Because this work is still preliminary and very early in development, >> there are no action items for the DataSketches engineering. >> >> Lee. >> >> >>
