Le 02/07/2019 à 00:20, Wes McKinney a écrit : > Thanks for the references. > > If we decided to make a change around this, we could call the first 4 > bytes a stream continuation marker to make it slightly less ugly > > * 0xFFFFFFFF: continue > * 0x00000000: stop
Do you mean it would be a separate IPC message? > > On Mon, Jul 1, 2019 at 4:35 PM Micah Kornfield <emkornfi...@gmail.com> wrote: >> >> Hi Wes, >> I'm not an expert on this either, my inclination mostly comes from some >> research I've done. I think it is important to distinguish two cases: >> 1. unaligned access at the processor instruction level >> 2. undefined behavior >> >> From my reading unaligned access is fine on most modern architectures and it >> seems the performance penalty has mostly been eliminated. >> >> Undefined behavior is a compiler/language concept. The problem is the >> compiler can choose to do anything in UB scenarios, not just the "obvious" >> translation. Specifically, the compiler is under no obligation to generate >> the unaligned access instructions, and if it doesn't SEGVs ensue. Two >> examples, both of which relate to SIMD optimizations are linked below. >> >> I tend to be on the conservative side with this type of thing but if we have >> experts on the the ML that can offer a more informed opinion, I would love >> to hear it. >> >> [1] http://pzemtsov.github.io/2016/11/06/bug-story-alignment-on-x86.html >> [2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65709 >> >> On Mon, Jul 1, 2019 at 1:41 PM Wes McKinney <wesmck...@gmail.com> wrote: >>> >>> The <0xffffffff><int32_t size> solution is downright ugly but I think >>> it's one of the only ways that achieves >>> >>> * backward compatibility (new clients can read old data) >>> * opt-in forward compatibility (if we want to go to the labor of doing >>> so, sort of dangerous) >>> * old clients receiving new data do not blow up (they will see a >>> metadata length of -1) >>> >>> NB 0xFFFFFFFF <length> would look like: >>> >>> In [13]: np.array([(2 << 32) - 1, 128], dtype=np.uint32) >>> Out[13]: array([4294967295, 128], dtype=uint32) >>> >>> In [14]: np.array([(2 << 32) - 1, 128], >>> dtype=np.uint32).view(np.int32) >>> Out[14]: array([ -1, 128], dtype=int32) >>> >>> In [15]: np.array([(2 << 32) - 1, 128], dtype=np.uint32).view(np.uint8) >>> Out[15]: array([255, 255, 255, 255, 128, 0, 0, 0], dtype=uint8) >>> >>> Flatbuffers are 32-bit limited so we don't need all 64 bits. >>> >>> Do you know in what circumstances unaligned reads from Flatbuffers >>> might cause an issue? I do not know enough about UB but my >>> understanding is that it causes issues on some specialized platforms >>> where for most modern x86-64 processors and compilers it is not really >>> an issue (though perhaps a performance issue) >>> >>> On Sun, Jun 30, 2019 at 6:36 PM Micah Kornfield <emkornfi...@gmail.com> >>> wrote: >>>> >>>> At least on the read-side we can make this detectable by using something >>>> like <0xffffffff><int32_t size> instead of int64_t. On the write side we >>>> would need some sort of default mode that we could flip on/off if we >>>> wanted to maintain compatibility. >>>> >>>> I should say I think we should fix it. Undefined behavior is unpaid debt >>>> that might never be collected or might cause things to fail in difficult >>>> to diagnose ways. And pre-1.0.0 is definitely the time. >>>> >>>> -Micah >>>> >>>> On Sun, Jun 30, 2019 at 3:17 PM Wes McKinney <wesmck...@gmail.com> wrote: >>>>> >>>>> On Sun, Jun 30, 2019 at 5:14 PM Wes McKinney <wesmck...@gmail.com> wrote: >>>>>> >>>>>> hi Micah, >>>>>> >>>>>> This is definitely unfortunate, I wish we had realized the potential >>>>>> implications of having the Flatbuffer message start on a 4-byte >>>>>> (rather than 8-byte) boundary. The cost of making such a change now >>>>>> would be pretty high since all readers and writers in all languages >>>>>> would have to be changed. That being said, the 0.14.0 -> 1.0.0 version >>>>>> bump is the last opportunity we have to make a change like this, so we >>>>>> might as well discuss it now. Note that particular implementations >>>>>> could implement compatibility functions to handle the 4 to 8 byte >>>>>> change so that old clients can still be understood. We'd probably want >>>>>> to do this in C++, for example, since users would pretty quickly >>>>>> acquire a new pyarrow version in Spark applications while they are >>>>>> stuck on an old version of the Java libraries. >>>>> >>>>> NB such a backwards compatibility fix would not be forward-compatible, >>>>> so the PySpark users would need to use a pinned version of pyarrow >>>>> until Spark upgraded to Arrow 1.0.0. Maybe that's OK >>>>> >>>>>> >>>>>> - Wes >>>>>> >>>>>> On Sun, Jun 30, 2019 at 3:01 AM Micah Kornfield <emkornfi...@gmail.com> >>>>>> wrote: >>>>>>> >>>>>>> While working on trying to fix undefined behavior for unaligned memory >>>>>>> accesses [1], I ran into an issue with the IPC specification [2] which >>>>>>> prevents us from ever achieving zero-copy memory mapping and having >>>>>>> aligned >>>>>>> accesses (i.e. clean UBSan runs). >>>>>>> >>>>>>> Flatbuffer metadata needs 8-byte alignment to guarantee aligned >>>>>>> accesses. >>>>>>> >>>>>>> In the IPC format we align each message to 8-byte boundaries. We then >>>>>>> write a int32_t integer to to denote the size of flat buffer metadata, >>>>>>> followed immediately by the flatbuffer metadata. This means the >>>>>>> flatbuffer metadata will never be 8 byte aligned. >>>>>>> >>>>>>> Do people care? A simple fix would be to use int64_t instead of >>>>>>> int32_t >>>>>>> for length. However, any fix essentially breaks all previous client >>>>>>> library versions or incurs a memory copy. >>>>>>> >>>>>>> [1] https://github.com/apache/arrow/pull/4757 >>>>>>> [2] https://arrow.apache.org/docs/ipc.html
