The detail of whether an `InputStream` or `OutputStream` is aligned to 8- or 
64-byte boundary had contaminated a number of IPC functions, causing generated 
metadata (plus padding, if any) to be non-deterministic. This introduces a new 
`ipc::AlignStream` function and the desired use is to align the stream at the 
highest level possible. I've implemented the changes in the Python tensor 
serialization code paths. I removed various cruft where alignment issues leaked 
into the API

[ Full content available at: https://github.com/apache/arrow/pull/2615 ]
This message was relayed via gitbox.apache.org for [email protected]

Reply via email to