Commit graph

3 commits

Author SHA1 Message Date
sterni
d485ebf01a fix(bubblegum): import lib from the new correct place
Since cl/2910 depot has no lib attribute anymore. Import it from the
depot fix point via depot.third_party.nixpkgs.lib to avoid passing
another argument and enlargening the shebang further.

Change-Id: I3c719eba38a5ceb36689ebf0409bd19d4f46a609
Reviewed-on: https://cl.tvl.fyi/c/depot/+/3050
Tested-by: BuildkiteCI
Reviewed-by: sterni <sternenseemann@systemli.org>
2021-04-20 08:42:58 +00:00
sterni
1c0f89f4ca feat(web/bubblegum): report some errors to the user via HTTP
We can actually catch some errors that may be generated in bubblegum
applications where we can report them to the user in a way that doesn't
require curl -vv:

* Type errors in the status argument: By removing yants completely we
  not only (presumably) gain some performance, but also the ability to
  return an internal server error on an unexpected type instead of
  throwing.

* User generated evaluation errors: by using builtins.tryEval we can
  catch throws and asserts the user inserted when generating the body
  and report to the user that something went wrong. To do: also support
  for the headers.

Change-Id: I8363b9825c6c730e624eb8016a5482d63cbc1890
Reviewed-on: https://cl.tvl.fyi/c/depot/+/2849
Tested-by: BuildkiteCI
Reviewed-by: sterni <sternenseemann@systemli.org>
2021-04-05 10:54:32 +00:00
sterni
93a746aaaa feat(web/bubblegum): nix CGI programming framework
So here is what has been keeping me up at night: At some point I
realized that nix actually made a somewhat passable language for CGI
programming:

* That `builtins.getEnv` exists as one of the impurities of Nix is
  perfect as environment variables are the main way of communication
  from the web server to the CGI application.

* We can actually read from the filesystem via builtins.readDir and
  builtins.readFile with bearable overhead if we avoid importing the
  used paths into the nix store.

* Templating and routing are convenient to implement via indented strings
  and attribute sets respectively.

Of course there are obvious limitation:

* The overhead of derivations is probably much to great for them to be
  useful via IfD.

* Even without derivations, nix evaluation is very slow to the point
  were a trivial application takes between 100ms and 400ms to produce a
  response.

* We can't really cause effects other than producing a response which
  makes it not viable for a lot of applications. There are some ways
  around this:

  * With a custom interpreter we could have streaming and multiplexed
    I/O (using lazy lists emulated via attrsets) to cause such effects,
    but it would probably perform terribly.

  * We can use builtins.fetchurl to call other HTTP-based microservices,
    but only in very limited constraints, i. e. only GET, no headers,
    and only if the tarball ttl is set to 0 in the global nix.conf.

* Terrible error handling capabilities because builtins.tryEval actually
  doesn't catch a lot of errors.

To prove that it actually works, there are some demo applications,
which I invite you to run and potentially break horribly:

    nix-build -A web.bubblegum.examples && ./result
    # navigate to http://localhost:9000

The setup uses thttpd and executes the nix CGI scripts using
users.sterni.nint which automatically passed `depot`, so they can
import the cgi library.

Change-Id: I3a22a749612211627e5f8301c31ec2e7a872812c
Reviewed-on: https://cl.tvl.fyi/c/depot/+/2746
Tested-by: BuildkiteCI
Reviewed-by: tazjin <mail@tazj.in>
2021-04-01 18:50:36 +00:00