I updated the PR to fix some issues with my edits that Antoine pointed out. I can start working on a C++ patch to implement the C++ changes in the next few days if that helps. Given the time urgency of deciding what to do on this if anyone else could express opinions it would be helpful.
I see one of the major benefits of this that when forming a sparse union (the one without the offsets buffer) the child bitmaps do not have to be spliced together to create the top-level bitmap with the assistance of the type ids buffer, which when you think about it is a rather complicated operation. On Wed, Jun 24, 2020 at 5:34 PM Wes McKinney <wesmck...@gmail.com> wrote: > > I drafted the specification changes that would be associated with the > union changes > > https://github.com/apache/arrow/pull/7535 > > I'll start a separate discussion about incrementing the > MetadataVersion since that must be discussed independently. > > Please take a look > > On Wed, Jun 24, 2020 at 3:50 PM Wes McKinney <wesmck...@gmail.com> wrote: > > > > I should also add that we could (with some effort) use the > > MetadataVersion V4/V5 indicator to offer backward compatibility for > > old serialized union data > > > > In any case, if there is consensus about this, we would need to have a > > vote and get busy with implementing and testing the changes. I could > > assist with the C++ changes > > > > On Wed, Jun 24, 2020 at 1:14 PM Wes McKinney <wesmck...@gmail.com> wrote: > > > > > > On Wed, Jun 24, 2020 at 1:07 PM Francois Saint-Jacques > > > <fsaintjacq...@gmail.com> wrote: > > > > > > > > OTOH, > > > > > > > > how do we handle NullType -> UnionType<T...> cast conversion? Do we > > > > require some convention like the first children ArrayData null bitmap > > > > to be set and all tags set to 0? > > > > > > Sure, that sounds like a reasonable implementation should this > > > operation actually be required. > > > > > > > François > > > > > > > > On Wed, Jun 24, 2020 at 1:09 PM Antoine Pitrou <anto...@python.org> > > > > wrote: > > > > > > > > > > > > > > > Le 24/06/2020 à 18:34, Wes McKinney a écrit : > > > > > > On Wed, Jun 24, 2020 at 11:08 AM Antoine Pitrou > > > > > > <anto...@python.org> wrote: > > > > > >> > > > > > >> > > > > > >> Le 24/06/2020 à 16:57, Wes McKinney a écrit : > > > > > >>> hi folks, > > > > > >>> > > > > > >>> As discussed on the recent GitHub PR [1], as a means of > > > > > >>> reconciling > > > > > >>> the long-standing cross-implementation incompatibilities with > > > > > >>> Union > > > > > >>> types, it's been proposed to remove the top-level validity bitmap > > > > > >>> from > > > > > >>> the Union data layout and let validity be determined exclusively > > > > > >>> by > > > > > >>> the child arrays of the union. So the only additional data needed > > > > > >>> to > > > > > >>> form a union are the type ids (and for the dense union, the > > > > > >>> offsets). > > > > > >>> > > > > > >>> I do not think this change meaningfully alters the semantics of > > > > > >>> Union > > > > > >>> types and I think it also simplifies their construction, so I > > > > > >>> would be > > > > > >>> in favor of making it for 1.0.0. > > > > > >> > > > > > >> So it sounds like this may break compatibility with existing only > > > > > >> uses > > > > > >> of Arrow C++ (and the relevant bindings: PyArrow, Arrow C/GLib, Red > > > > > >> Arrow); not only on the API side, but on the data side. > > > > > > > > > > > > Right. However, I don't think these changes will be very disruptive, > > > > > > and we always knew that this disruption was possible because of the > > > > > > hitherto unreconciled issues with Unions. The applications that I'm > > > > > > aware of that use Union serialization (e.g. Ray) use it only for > > > > > > ephemeral serialization. > > > > > > > > > > Ok, that's a convincing argument. > > > > > > > > > > > In general, I think that we should be bumping the metadata version > > > > > > [1] > > > > > > for 1.0.0 to create a forcing function for upgrade to the > > > > > > format-stable line of libraries. The C++/Python libraries could > > > > > > have a > > > > > > "compatibility mode" (like the "write_legacy_ipc_format" options) > > > > > > that > > > > > > writes MetadataVersion::V4 (v0.8.0 -> v0.17.1) with certain features > > > > > > (like unions -- which are not needed for Spark for example) > > > > > > disabled. > > > > > > > > > > Hmm, I hope we can keep the negotiation minimal. We should take from > > > > > the Jon Postel principle - be liberal in what you accept, strict in > > > > > what > > > > > you emit. > > > > > > > > > > So the IPC reader can have a simple detection that goes this way: > > > > > > > > > > * if we receive 1 buffer for sparse union or 2 buffers for dense > > > > > union > > > > > => it's the new-style format, there's nothing to do > > > > > > > > > > * if we receive 2 (non-null) buffers for sparse union or 3 > > > > > (non-null) > > > > > buffers for dense union > > > > > => it's the old format, we should AND the parent bitmap into each of > > > > > the > > > > > child bitmaps > > > > > > > > > > We can also add a flag to IpcOptions to enable/disable compatibility > > > > > tricks. > > > > > > > > > > Regards > > > > > > > > > > Antoine.
