Discuss the usage of Span vs an owned bag of bytes type

[FranzBusch]

#4 (comment)

[Lukasa]

So the principal problem with Span is that it essentially forces users to copy bytes in order to do anything non-trivial with their data. Put another way, Span is always less useful to users than a data type that can give you a Span, because any data type that can give you a Span can immediately do whatever Span can do, but it can also do other things.

As for what the owned bag of bytes type will be, I think that's a substantially more challenging question. My preference would be to sink a very simple type into the standard library: essentially a ~Copyable struct wrapping a buffer pointer. This type would then be able to vend Spans. It would also be able to be the backing object for ByteBuffer, or at least be able to be extracted from a ByteBuffer very cheaply, and it would be able to back a Data very cheaply as well.

In any case, we should be strongly avoiding the use of Span as the de-facto transfer mechanism for bytes. Span is for processing bytes, not for transferring them.

As a sidebar, Span<UInt8> would also be wrong, it should be RawSpan.

[tfpauly]

I agree that a non-copyable storage would be much preferable here. Using a UniqueArray<UInt8> is something that works well for moving around ownership over bytes uniquely, as one option.