Sorry for the quick double post here, but just wanted to clarify point #3 above. I understand JWTs are still private in the sense that you could beat the caller to the server with the same credentials it is trying to use (or rely on the fact that the service isn't using the JTI as a nonce). In any case, my point was that the attack vector is significantly reduced with short lived cryptographic tokens compared to, for example, bearer tokens. I believe the other points still remain strong.
Best, Colin On Monday, March 26, 2018 at 3:14:03 PM UTC-4, colin....@gmail.com wrote: > > Hey group, > > I've seen discussions before about CallCredentials and their ability to be > used on insecure channels. It seems that, at least today, they can't be > used for any C-based implementations of gRPC. I wanted to propose a change > to that, and suggest CallCredentials should be able to be used on insecure > channels (even if an option is required to enable this behavior). There are > a couple of reasons I think this should be changed: > > 1) At least gRPC-java does support this. At best, the inconsistency is > strange, at worst it could learn to painful realizations down the road if > starting on gRPC and assuming that similar patterns will "just work" on > other languages. This is what happened in my case, where our gRPC-Java > implementations worked fine, but attempting to do the same thing in Node > did not work and took a while before I realized this was the reason. > 2) While I understand gRPC's belief that it's insecure to exchange tokens > over plaintext channels, the reality is that the application-level > implementation really has no idea what channel the data will actually be > exchanged over. For example, in Istio deployments, the application may > think it's communicating insecurely (and thus not allow CallCredentials to > be sent), when in fact the traffic is going to hit an external container > that will perform mTLS auth with the destination service. From the client > and server perspective, this is an insecure channel, but in reality - it's > not (unless you're concerned about the ability to tcpdump the loopback > interface - at which point you're probably screwed anyway). > 3) There are plenty of cases where the CallCredentials themselves are not > necessarily private, and thus may be fine to exchange over plaintext (think > JWTs). This could be the case in scenarios where the services themselves > are not dealing with private information, but perhaps they perform an > action that should still be authenticated. Understandably, everything > should be TLS anyway, but see point #2 for cases in which the service might > be using TLS in ways that gRPC may not know about. > 4) Finally, from a developer experience perspective, it's still possible > to send this information anyway - but it results in more fragile > implementations of gRPC clients. For example, in Node, I've worked around > this limitation by simply pre-generating Metadata instances that can be > passed to calls (instead of using CallCredentials), but this requires me to > take care to ensure that, at all call-sites, I have valid metadata (i.e. it > hasn't expired since it was generated). CallCredentials provide a single > way for me to do this, but it's currently not possible because of the > restriction to use secure channels. > > Hopefully, these are some compelling reasons to consider it. But, if not, > at least this should hopefully start a conversation about the topic. > > Best, > Colin > -- You received this message because you are subscribed to the Google Groups "grpc.io" group. To unsubscribe from this group and stop receiving emails from it, send an email to grpc-io+unsubscr...@googlegroups.com. To post to this group, send email to grpc-io@googlegroups.com. Visit this group at https://groups.google.com/group/grpc-io. To view this discussion on the web visit https://groups.google.com/d/msgid/grpc-io/16e328d5-49bb-426d-aaa1-4c3a02cacad9%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
