A library for benign effects in Haskell
The current state is a proof of concept, it is not meant to be used in actual projects nor do I guarantee any stability yet
A core tenet of Haskell programming is “pure core, imperative shell”. We write most of our programs in pure style, and a small outer layer in imperative shell. This is great for correctness, but it comes at a great cost for observability. Because logging, tracing, or measuring time are all considered effects, we can't log, trace or measure time in pure code.
This means that most of our code is traditionally unobservable. There are some built-in facilities in GHC, such as cost-centre profiling. But besides putting pressure on the compiler to build these tools, these facilities are out of the language, hence not programmable in anyway way.
The Benign library is an attempt to address this gap. A difficulty is laziness. With laziness, things have a beginning (when the thunk is being forced), but not a well-defined end. So what do we measure the time of? The solution of the Benign library is to be less lazy. We keep laziness for algorithms, but use strict (or at least stricter) types to assemble bigger steps. It's fine to log or trace in pure code, since we don't consider that these observations are part of the semantics of the program.
The Benign library provides facilities to create benign effects, including to use strict types to assemble these large steps.
The premise underlying all this, as well as the implementation of the library, is that logging or tracing is not very fast. So we don't want to log or trace in places where performance is really essential. This is why at the most inner level, where tight loops and algorithms live, laziness is not a problem: we are not going to log there, this would cost too much performance. At a more outer level, we can use strictness to make steps with a beginning and an end. It's ok if there is a cost in terms of conversion, this is not where we need to optimise too much. The library can, and does, prevent optimisation through its functions (note that cost-centre profiling also prevents optimisation through cost centres; it isn't surprising that we are having a similar problem). With the optimisation not working at that level, strictness is much less of a problem.
Backends are shipped as separate package (with the exception of backends which require no additional dependencies).
I've got to confess that it's a major pain in my delicately dignified bottom. The language server works rather poorly across packages; duplicating the Cabal files is somewhat annoying, what with the common stanzas being lost and the version numbers being easier to get wrong, I'll have to publish several packages every time I want to make a release. I'm probably forgetting other issues. But until Hackage supports multiple-library packages, it's the only solution that I have to avoid pulling, say, Katip, when I want to use Opentelemetry.
You are welcome to write benign effects directly in your package. If you wish to upstream one of my backends: I'll deprecate mine.
Build as
$ stack buildThe formatter is Ormolu 0.5.0.1. Format with
$ just formatYou can use Nix (specifically nix-shell) to provide all the
necessary dependencies (including a compatible Haskell Language
Server).
For reasons unknown to me, Ormolu is not cached in the pinned version
of Nix. I haven't had time to investigate, instead I'm using
Cachix. For the nix-shell to load fast, install
cachix and run (once per
machine) prior to building the Nix shell.
$ cachix use aspiwackCredits go to
- Alexis King for helping me come up with the implementation strategy
based on
asyncand thread ids. - Thomas Bagrel for suggesting (in another context) the phrase “lexical state” to me. I don't know whether the phrase has been used before, a quick googling didn't turn up anything.