Re: [BuildStream] CAS server resource names & instance names

[Jürg Billeter <j bitron ch>]

On Mon, 2018-12-10 at 13:41 +0000, Jim MacArthur wrote:
> On 10/12/2018 12:46, Jürg Billeter wrote:
>
> > > Remote execution storage will take an optional instance name and not
> > > provide it or the initial slash if it's not provided, although I will
> > > allow specification of "" as a valid instance name, distinct from not
> > > specifying one.
> > I don't understand this. In the protocol there is no difference between
> > an empty instance name and not specifying an instance name (as in
> > protobuf it's the same). Why (and how) would we treat these two cases
> > differently in BuildStream, possibly violating the spec?
> I don't see how you've got to the conclusion that empty instance and no 
> instance are the same thing. The Bazel protocol does say "For servers 
> which do not support multiple instances, then the `instance_name` is the 
> empty path" but that doesn't necessary imply behaviour for servers which 
> do support it. I can also see mappings to http URLs which look like 
> they'll always include an instance name or empty string, but I wasn't 
> considering them authoritative specifications.

For example, for Execution.Execute() instance_name is a protobuf string
field in the ExecuteRequest message. URL mapping is handled internally
by gRPC, clients can't normally influence it. Protobuf messages don't
allow distinction between unset string fields and empty string fields
(the serialization is identical).

I.e., at least for Execution.Execute() empty instance_name and no
instance_name are identical. It's impossible for the server to treat
them differently. I don't think it makes sense to distinguish the two
cases in other places where it would theoretically be possible. That
would be inconsistent and confusing, in my opinion.

Cheers,
Jürg