| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
I wanted to even further simplify the build tooling overhead. My general
goal is to not have to think about declaring packages, or dependencies,
or really anything that you might find in a cabal file. Not all of these
goals are possible, but we can get pretty close. With this commit all I
need for the 'buildGhc/buildGhcjs' functions is the path to the
entrypoint file; everything else is either inferred by the Nix code or
declared in the Haskell code comments.
The strategy is to map a Haskell module to an executable artifact, and
pass just that module to 'ghc --make'. Then we can rely on ghc to handle
walking the local filesystem for imports. The only thing ghc really
needs to know is a name for the output executable; it is hard to
automatically infer this, so we have a simple comment syntax to declare
this in the file. The comment syntax is inspired by existing Haskell
'LANGUAGE' pragmas; having this in the same file keeps the configuration
as close to the real code as possible. The Nix code then extracts this
info from the code comments, and sets the required ghc flags.
Second, we need to declare the set of 3rd-party packages that our
program relies on. For this we can re-use the same comment syntax and
just list the dependencies, then extract them in Nix and construct a
package set as we were before.
This reduces the amount of "package declaration" code we have to write
in default.nix, and reduces the amount of time we have to spend
switching between the Haskell code and the Nix code (I find such context
switching super annoying). I also think having the configuration in with
the Haskell code encourages us to write smaller, simpler modules and
only write code that we need.
Additionally, I refactored the bild and ghci (now called 'repl') scripts
to work in any directory. The .envrc uses direnv to set the path so that
you can run these scripts anywhere. That means the following works:
$ cd Run/Que
$ bild Website
$ repl Server
λ> :l Run.Que.Server
I find this to be a rather nice workflow.
|
|
|
|
I also block before taking, instead of doing a mult and tap. This is a
simple way to fix the memory issue, and makes them conceptually simpler
to work with I think.
The channels are still mutli-consumer and multi-producer, which is
fine. I'm not sure now I will implement the regular pubsub, but I'm not
sure there is a great usecase for that anyway.
|
|
|
|
|
|
I'm using serval.simatime.com as a catch-all production app server for
now. The 'que.run' domain is pointed at that instance, and the service
is just installed as a regular NixOS systemd service.
I had to do some troubleshooting because I wasn't getting any DNS names
to resolve. I think changing the nameservers fixed it. Don't know why
the 127 number was in there.
Another issue concerns how to add our packages to the set of nixpkgs in
the generated NixOS. I played around with this for a while and landed on
using an overlay to put our set of packages under 'pkgs.biz.<name>', and
then passing that in to the 'buildOS' function. This isn't really the
best solution because it is confusing and rather disconnected. I'm
starting to realize that it might be good to separate nix artifacts into
"machines" and "programs", but I don't want to do that just yet. I'd
like to finish designing my bild program before making any large design
decisions or re-organizations.
|
|
Now that I have the domain name que.run! Aw yeah.
|
|
|
|
|
|
The performance is reportedly better. The API is simpler. Also with STM
channels, I couldn't get multiconsumer to work. I was able to get it to
work with unagi. Also I could write 'mult' and 'tap' which bring me back
to my Clojure days.
|
|
|
|
More generally, we could extend this to other settings, like 'main-is'
and target architecture to compile for and so on. It would be best to
define the parameters in Nix first, then later inline them to the
code comments after we've worked out the interface.
|
|
|
|
As much as I like these operators, I have to remove them because they
don't work as expected. Haskell doesn't allow you to have unary prefix
operators. I can't find a way around this, and it's not that important
anyway.
|
|
It's easier to remember what operators do, and thus easier to write and
read condens code, if they follow some symbolic pattern or visually
represent the concept to which they map. This is in part inspired by
hoon, in part by OCaml's operators.
I'm not married to these operators specifically, but I think they are
good so far.
|
|
|
|
|
|
|
|
Two functions makes it simpler to reason about what is being built and
when, even if it is a bit more explicit. I also removed the dumb
Apex/Aero naming thing because Server/Client is just easier to remember.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
rouns are no longer in use anyway.
|
|
|
|
|
|
|
|
|
|
|
|
- make it actually compile
- add fmt
- export stuff I'm using elsewhere
- comment stuff that doesn't work yet
|
|
|
|
I don't have runserval.com yet, so I'm not gonna put the cart before the horse
|
|
|
|
|
|
|
|
This was causing the system to revert once per day.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
