build: Print dependency graph on failure

[jtojnar]

This is my attempt at finding out why I get get “Dependency cycle detected” errors.

• I am not sure how should I actually output the Graphviz data, currently just printing them to stdout.
  • Maybe make Logger output to stderr.
• It also increases dependency tree size: Nix package went from 4.5G to 4.8G. So we probably need to allow disabling it.
• Identifiers with the same path should be grouped together.
• Also print when the build succeeds.
• Avoid Orphan instance PrintDot Identifier
• Listed dependencies incomplete.
• Ensure graph nodes are escaped.
• Do not cluster single nodes.

  • Looks worse

    Patch

    --- a/lib/Hakyll/Core/Runtime.hs
    +++ b/lib/Hakyll/Core/Runtime.hs
    @@ -236,9 +237,12 @@ dependencyGraph fullDeps = graphElemsToDot graphParams (map labelNodes nodes) ed
         where
             nodes = S.toList (M.keysSet fullDeps `S.union` foldMap (S.map fst) fullDeps)
             edges = S.toList (M.foldMapWithKey (\ident -> S.map (\(dep, snap) -> (ident, dep, snap))) fullDeps)
    +        clusterElemCounts = foldr (\ident -> M.insertWith (+) (toFilePath ident) 1) M.empty nodes
             graphParams = defaultParams
                 { clusterBy = \(n, nl) -> C (toFilePath n) (N (n, nl))
                 , clusterID = Str . LT.pack
    +            -- Do not make paths without multiple versions into clusters.
    +            , isDotCluster = \cl -> M.findWithDefault 0 cl clusterElemCounts > 1
                 , fmtEdge = \(_, _, el) -> if null el then [] else [textLabel (LT.pack el)]
                 }
             labelNodes ident = (ident, show ident)

• Show snapshot names.
• Maybe move it to a separate command?

I recommend xdot for viewing the graphs. Edotor works in the web browser but does not support highlighting adjacent edges. Passing -f fdp to xdot might lay out the graph slightly more manageably.

[jaspervdj]

This is really cool work!

I wonder how to integrate this into the existing commands. Maybe one option would be to have something like this.

This is currently the most "low-level" way to run Hakyll:

hakyllWithExitCode :: Config.Configuration -> Rules a -> IO ExitCode

If we wanted to return more diagnostic information, we could do something like:

-- Other possible names: HakyllOutput, Diagnostics?
data HakyllResult = HakyllResult
  { hakyllDependencyGraph :: ...
  , hakyllExitCode :: ExitCode
  , -- Other stuff that people may care about in the future
  }

hakyllWithResult :: Config.Configuration -> Rules a -> IO HakyllResult

And then people could use it however they want in their main:

main = do
  info <- hakyllWithResult $ ...
  writeFile "graph.dot" $ hakyllDependencyGraph info

@Minoru What do you think?

[Minoru]

I see how this could be used to debug our dependency-checking code, but I wonder if end users actually need this. It seems to me that it'd be better to improve our error messages instead (and debug the deps checker, of course).