Temporary errors in transport layers

[Matthieu Moy]

Robert Collins <robertc at robertcollins.net> writes:

> I agree that internal transient errors like this should not in general
> be exposed to the clients. However if there is risk of continual
> looping

There must clearly be a counter and a fatal error after one, or
perhaps two retries.

> or of not honouring any atomicity constraints, it should be raised to
> the client.
>
> For instance, reading a single file - if that is interrupted when the
> file is only partly downloaded, reconnecting and using REST in FTP *may
> not return the same file*. So that error must be raised to the client of
> the transport.

I *think* this is not a problem as soon as you restart the atomic
operation from start. In your case, you should probably throw away the
piece of file you've downloaded and get it again.

-- 
Matthieu