Row-type polymorphism in the context of Cardano scripts

[klntsky]

Regarding row type polymorphism (RTP for short from now on), which is the only non-trivial language feature that we have to implement, we have to consider two questions:

I. What value does it bring to the PS ecosystem and how is it relevant for the on-chain scripts domain
II. What costs are we willing to pay for RTP?

[klntsky]

I. What value does it bring to the PS ecosystem and how is it relevant for the on-chain scripts domain

As far as I'm aware of the ecosystem, row type polymorphism is mainly used to achieve the following goals:

1. Operations with records bypassing built-in syntactic constructs: e.g. updating record fields using Proxy - this approach is of course more composable than record update syntax. Also, lenses.
2. Defining data-generic encoders/decoders for record types (most notably, argonaut for JSON)
3. FFI stuff for simpler JS interop, e.g. untagged-union
4. Custom data machinery, like type-safe variants - this is a clever hack that is arguably not widely used (no core libraries depend on it)

[klarkc]

I believe RTP is highly valuable when decoding/encoding JSON, which is very common in web development - a context where PureScript is likely to be used. I suggest not giving up on RTP, or at least making it a low priority / future effort.

[klntsky]

II. What costs are we willing to pay for RTP?

I think that if we fail to find a way to implement RTP as a zero-runtime-cost abstraction, we should give up on it, because, as we know from Haskell experience, using a FP language with no RTP is just fine. And we are not even building a general purpose language - just a glorified DSL.

[klntsky]

Let's examine the runtime costs of the approaches/libs involing RTP I listed here with the existing JS backend. My argument is: if the PS folks didn't figure out a zero-runtime-cost approach for JS, then it is unproductive for us to try doing that for UPLC, given the limited budget. A few years ago even Proxy was not zero-cost, but we need to figure out the current state of affairs.

[klarkc]

AFAIK, records in PureScript translate to plain runtime JS objects, enabling unique optimizations at runtime, beyond just compile time.

[paluh]

In practice having polymorphic types is also useful to share type definitions like:

type Base r = { a :: A | r }
type R1 = Base (b :: B)
type R2 = Base (c :: C)

I wonder if it is feasible to keep row RTP in the type system but require full manual specialization of value types.

I think that if we fail to find a way to implement RTP as a zero-runtime-cost abstraction, we should give up on it (..)

I fully agree with this if I can share my user opinion. Maybe even smaller language (better error messages?) would be easier to use in this context.