So I think the issue is that we are serializing record batches in a
distributed fashion, and then concatenating them in the streaming
format. However, the message serialization only aligns the start of the
buffers, which requires it to know the current absolute offset of the
output stream. Would there be any problem with padding the end of the
message, so any single serialized record batch would always be a
multiple of 8 bytes?
I've put together a branch that does this, and the existing java tests
all pass. I'm having some trouble running the integration tests though.
Thanks,
Emilio
On 08/08/2017 09:18 AM, Emilio Lahr-Vivaz wrote:
Hi Wes,
You're right, I just realized that. I think the alignment issue might
be in some unrelated code, actually. From what I can tell the the
arrow writers are aligning buffers correctly; if not I'll open a bug.
Thanks,
Emilio
On 08/08/2017 09:15 AM, Wes McKinney wrote:
hi Emilio,
From your description, it isn't clear why 8-byte alignment is causing
a problem (as compare with 64-byte alignment). My understanding is
that JavaScript's TypedArray classes range in size from 1 to 8 bytes.
The starting offset for all buffers should be 8-byte aligned, if not
that is a bug. Could you clarify?
- Wes
On Tue, Aug 8, 2017 at 8:52 AM, Emilio Lahr-Vivaz
<elahrvi...@ccri.com> wrote:
After looking at it further, I think only the buffers themselves
need to be
aligned, not the metadata and/or schema. Would there be any problem
with
changing the alignment to 64 bytes then?
Thanks,
Emilio
On 08/08/2017 08:08 AM, Emilio Lahr-Vivaz wrote:
I'm looking into buffer alignment in the java writer classes.
Currently
some files written with the java streaming writer can't be read due
to the
javascript TypedArray's restriction that the start offset of the
array must
be a multiple of the data size of the array type (i.e. Int32Vectors
must
start on a multiple of 4, Float64Vectors must start on a multiple
of 8,
etc). From a cursory look at the java writer, I believe that the
schema that
is written first is not aligned at all, and then each record batch
pads out
its size to a multiple of 8. So:
1. should the schema block pad itself so that the first record
batch is
aligned, and is there any problem with doing so?
2. is there any problem with changing the alignment to 64 bytes, as
recommended (but not required) by the spec?
Thanks,
Emilio
