On Thu, Jun 25, 2020 at 5:31 AM Antoine Pitrou <anto...@python.org> wrote: > > > Le 25/06/2020 à 12:18, Antoine Pitrou a écrit : > > > > Le 25/06/2020 à 00:40, Wes McKinney a écrit : > >> hi folks, > >> > >> This has come up in some other contexts, but I believe it would be a > >> good idea to increment the version number in Schema.fbs starting with > >> 1.0.0 to separate the pre-1.0 and post-1.0 worlds > >> > >> https://github.com/apache/arrow/blob/master/format/Schema.fbs#L22 > >> > >> Given that we are contemplating a number of changes to assist with > >> forward compatibility and a breaking serialization change for unions, > >> this would seem prudent so that we do not risk breaking compatibility > >> with 0.17.1 and prior. > >> > >> Given that there are no major backwards incompatibilities, there > >> should be no problem with 1.0.0 readers reading data generated by > >> libraries <= 0.17.1. > > > > Actually, it seems that a dense array with top-level null values > > (represented in 0.17.1 fashion) would need non-trivial rewriting of its > > offsets and child arrays (at least one child array) to represent the > > nulls at the child level. > > > > This is unless we keep the top-level union null bitmap in C++ and only > > avoid emitting it on the IPC side. Which would be a slightly weird > > arrangement, but would limit incompatibilites on the C++ API side. > > Actually, if we do this, the same problem will appear on the IPC write > side (C++-created dense union arrays with a top-level null bitmap will > need regenerating some of the child buffers).
I see. Well I think we can shut down this issue by giving up on Union forward compatibility V4 / pre-1.0 libraries. > Regards > > Antoine.
