How to add "provided" dependencies back to run/test tasks' classpath?

Here's an example build.sbt:

import AssemblyKeys._

assemblySettings

buildInfoSettings

net.virtualvoid.sbt.graph.Plugin.graphSettings

name := "scala-app-template"

version := "0.1"

scalaVersion := "2.9.3"

val FunnyRuntime = config("funnyruntime") extend(Compile)

libraryDependencies += "org.spark-project" %% "spark-core" % "0.7.3" % "provided"

sourceGenerators in Compile <+= buildInfo

buildInfoPackage := "com.psnively"

buildInfoKeys := Seq[BuildInfoKey](name, version, scalaVersion, target)

assembleArtifact in packageScala := false

val root = project.in(file(".")).
  configs(FunnyRuntime).
  settings(inConfig(FunnyRuntime)(Classpaths.configSettings ++ baseAssemblySettings ++ Seq(
    libraryDependencies += "org.spark-project" %% "spark-core" % "0.7.3" % "funnyruntime"
  )): _*)

The goal is to have spark-core "provided" so it and its dependencies are not included in the assembly artifact, but to reinclude them on the runtime classpath for the run- and test-related tasks.

It seems that using a custom scope will ultimately be helpful, but I'm stymied on how to actually cause the default/global run/test tasks to use the custom libraryDependencies and hopefully override the default. I've tried things including:

(run in Global) := (run in FunnyRuntime)

and the like to no avail.

To summarize: this feels essentially a generalization of the web case, where the servlet-api is in "provided" scope, and run/test tasks generally fork a servlet container that really does provide the servlet-api to the running code. The only difference here is that I'm not forking off a separate JVM/environment; I just want to manually augment those tasks' classpaths, effectively "undoing" the "provided" scope, but in a way that continues to exclude the dependency from the assembly artifact.

Solution 1:

For a similar case I used in assembly.sbt:

run in Compile <<= Defaults.runTask(fullClasspath in Compile, mainClass in (Compile, run), runner in (Compile, run)) 

and now the 'run' task uses all the libraries, including the ones marked with "provided". No further change was necessary.

Update:

@rob solution seems to be the only one working on latest SBT version, just add to settings in build.sbt:

run in Compile := Defaults.runTask(fullClasspath in Compile, mainClass in (Compile, run), runner in (Compile, run)).evaluated,
runMain in Compile := Defaults.runMainTask(fullClasspath in Compile, runner in(Compile, run)).evaluated

Solution 2:

Adding to @douglaz' answer,

runMain in Compile <<= Defaults.runMainTask(fullClasspath in Compile, runner in (Compile, run))

is the corresponding fix for the runMain task.

Solution 3:

Another option is to create separate sbt projects for assembly vs run/test. This allows you to run sbt assemblyProj/assembly to build a fat jar for deploying with spark-submit, as well as sbt runTestProj/run for running directly via sbt with Spark embedded. As added benefits, runTestProj will work without modification in IntelliJ, and a separate main class can be defined for each project in order to e.g. specify the spark master in code when running with sbt.

val sparkDep = "org.apache.spark" %% "spark-core" % sparkVersion

val commonSettings = Seq(
  name := "Project",
  libraryDependencies ++= Seq(...) // Common deps
)

// Project for running via spark-submit
lazy val assemblyProj = (project in file("proj-dir"))
  .settings(
    commonSettings,
    assembly / mainClass := Some("com.example.Main"),
    libraryDependencies += sparkDep % "provided"
  )

// Project for running via sbt with embedded spark
lazy val runTestProj = (project in file("proj-dir"))
  .settings(
    // Projects' target dirs can't overlap
    target := target.value.toPath.resolveSibling("target-runtest").toFile,
    commonSettings,
    // If separate main file needed, e.g. for specifying spark master in code
    Compile / run / mainClass := Some("com.example.RunMain"),
    libraryDependencies += sparkDep
  )

Solution 4:

If you use sbt-revolver plugin, here is a solution for its "reStart" task:

fullClasspath in Revolver.reStart <<= fullClasspath in Compile

UPD: for sbt-1.0 you may use the new assignment form:

fullClasspath in Revolver.reStart := (fullClasspath in Compile).value