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.
