Discussion: "Built-in" capability providers defined at runtime
Brooks: Some disambiguation before we get into it
The wasmCloud host has some builtins (ex. numbercore and logging), but it also has support for "external" providers
Teams are going to have a core of functionality that they can share and access, which is what providers are for
For example, if you're working on a project and everyone is using a HTTP client/server, then you'll want the HTTP client/server provider available
The key question is when you run a wasmcloud host, should there be a way to specify providers which are treated like built-ins (no need to configure link defs, etc)?
(audio issues)
Jordan: We started this discussion 2/3 talks ago, mentioning the cryptographic space and built-in numbers generators
Some providers are good for using locally, but we don't want people to remove builtin stuff (logging) & relink it
If there's a directory of providers, it would be nice to have them loaded and accessible as builtins
Don't want to rewrite CI to use different contracts and providers
Brooks: We've been talking about logging, but do you see this concept extending to other capabilities?
Jordan: Maybe there's a centralized ops team that maintains a set of shared plugins
Kevin: what we're doing right now is looking at all our contracts and seeing how they need to evolve with all the changes coming (ex. components)
Whether something is builtin or not should actually not be built into the contract
It's up to the host to decide if it satisfies the contract with a builtin or something else.
This is really a way to avoid creating link definitions
If you find yourself trying to make environment-specific changes to a contract, there's a bit of a problem -- you should be using the same contract across platforms
There is some impedance mismatch here
Jordan: Do you see the capabilities being pluggable or hardcoded into the host?
Kevin: depends on the host -- what the OTP host and Rust host (and other hosts) can do is different
For example a JS host might have more hard-coded builtins, and the OTP host might be able to accept builtin providers in Elixir/Erlang and have them loaded in
There are lots of possibilities here, but the idea is spot on -- we need to resolve discrepancies in the contracts
Jordan: Can we get more visibility into the builtins? They're the only ones we have no visibility into
Kevin: Turns out this is a bigger problem -- there's a general need for providers to emit their own events, and there's no good way to do that right now.
Also this rolls into not forcing the actors to know the difference between a builtin and not-builtin.
Vance: Assuming the difference between a builtin and an external provider is that there is no link definition...
Kevin: You can delete and add a new linkdef for providers, but you can't do that for builtins.
Vance: I'm assumign that today the builtins don't use the lattice?
Kevin: That's right, they don't
Vance: How does it work when you change from one provider to another? If you want to migrate from one http provider to another one, would that be destructive/cause downtime?
Kevin: Deleting the LinkDef could be potentially destructive
Vance: I thought maybe you'd be able to add the link first?
Kevin: It would be rejected since only one link for a given contract can exist at one time
Brooks: We've thought about this some...
Brooks: Linking an actor (creating a LinkDef) to a provider serves two main purposes:
Telling wasmcloud that the two things will talk
Provide optional configuration values
Here's the question: how much security validation are we getting at runtime by the link being explicit? If it was automatic, are we compromising on security somewhat?
Kevin: the dangerous case is that there's a signed contract, wasmCloud tries to find a suitable provider, there have some issues:
It doesn't know which providers need config and which ones don't
If there's an impostor (with that contract signed), then it's possible to choose it
Kevin: this is why the linkdefs up until now have been explicit -- you need to sign and have cluster issuers.
It'd be more unpredictable which would be an even bigger problem than security.
Link names can't change since the actor needs them
Brooks: It's interesting
Jochen: Is this a limitation of NATS or a chosen path? since it can only receive on single paht but listen on a *?
Brooks: It's more of a chosen path -- we use a different links (frotend and backend) to get around that, so we're flexible in routing.
Brooks: We're trying to find a way for linking to be more natural
Most of the time there's not even configuration required (ex. HTTP Client is just making outbound requests, versus HTTP Server which might need a port to listen on)
You could run two KVs but there's no way to know which one to talk to by default, which feels artificially restrictive
No concrete suggestions, but we've got some people with hands
Jordan: Concrete suggestion, rename the project to Wild Wild wasmCloud west.
More seriously: compiling link names to actors -- it's been a bit of a pain-point
Kevin: COmment on Brooks' usecase (implicit links)
In addition to the security problem, you could have many different providers all on the same link ID on the same lattice, and they could pick the wrong one (ex. looking for redis, and you find vault)
Brooks: definitely still things to think about here, but this settles it for me (for now)
Brooks: What should we take away from this discussion? What about Jordan's question from Slack about compiling link names
Jordan: Should washboard have a section that says "these builtins are up and running"
Jordan: Other question was answered -- Kevin says it's in the works
Bailey: we should probably write some RFCs
Jordan: If we do a bunch of hosts with different providers and builtins and lattice them all together, how do we...
Kevin: if the list of builtins is variable and we don't have contracts that specify that they're builtins, then only if a no link exists should a builtin be chosen
If an actor comes up on a host has a builtin for wasi:logging and there is no link definition, then the host should hook the actor up to the builtin wasi logging
Only if no link exists for the contract would the builtin be used
If you deploy a distributed logger, then a linkdef should be added to stay meaningful, explicit and durable
We do want a hierarchy of defaults here
Brooks: Love that, we should totally get an RFC
Kevin: one example of reasonable defaults -- right now you have to deploy an HTTP client provider to enable HTTP requests... But a host could easily provide this, and then you don't need a link definition
An issue is that if you have multiple different hosts -- one with a default and another one without -- then you could get a provider when you didn't expect it