scalanb: Scala notebook

Status: PoC

Installation

Scalanb is not published yet.

// In build.sbt
// To use batch notebook, you need macro paradise plugin and additional compiler options.
addCompilerPlugin("org.scalamacros" % "paradise" % "2.1.0" cross CrossVersion.full)
scalacOptions += "-Yrangepos"

Batch Notebook

1. Setup dependencies, compiler plugin, scalac options in build.sbt
2. Create notebook class with @Notebook annotation
3. Run notebook (main method is automatically generated)
4. .ipynb is saved in ~/.scalanb/hist(default)

import com.todesking.{scalanb => nb}
@nb.Notebook
class MyNotebook {
  nb.markdown("# Example of scalanb")
  // add more code here
}

and

$ sbt 'runMain MyNotebook'

See Example1.scala and its output

To specify history location, use --out option.

$ sbt 'runMain MyNotebook --out=file:path=./hist/'

Spark Batch Notebook

Use spark.Notebook annotation

import com.todesking.{scalanb => nb}
@nb.spark.Notebook
class MyNotebook {
  // spark session available here
  val df = spark.read.csv("...")
  // Show dataframe as HTML tables via `nb.show` method
  df.nb.show(10)
}

$ sbt assembly # Make fatjar
$ spark-submit --class MyNotebook myapp.jar

Save history to HDFS

Requirement: scalanb-spark

$ sbt 'runMain MyNotebook --out=hdfs:path=/tmp/hist/'

Execution log

When --log option enabled, realtime log available.

$ sbt 'runMain MyNotebook --log'

# .scalanb/hist/{TIME}_{NOTE_NAME}.log
[2018-08-21 21:46:48] > nb.setShowTimeMillis(100)
[2018-08-21 21:46:48] > nb.markdown("# Scalanb Example")
[2018-08-21 21:46:48] > val a = 1
[2018-08-21 21:46:48] > val b = 2
[2018-08-21 21:46:48] > a
[2018-08-21 21:46:48] => 1
[2018-08-21 21:46:48] > println(s"a = $a")
[2018-08-21 21:46:48] stdout: a = 1

Caching

import com.todesking.{scalanb => nb}
@nb.Notebook
class BigData {
  val cp = nb.checkpoint
  val rawLog = cp.nocache { loadData("data/raw.csv") }
  val count = cp.cache(rawLog) { rawLog => rawLog.count() }
  cp.unwrap(count) { count =>
    println(s"count = $count")
  }
  val userId = 10
  val theUsersLogs = cp.cache((rawLog, userId)) { case (rawLog, userId) =>
    rawLog.where('user_id === userId)
  }
  cp.unwrap(theUsersLogs) { theUsersLogs =>
    theUsersLogs.count()
    theUsersLogs.show()
  }
}

Cache is based on value’s ID.
ID calculated from

• val name
• AST
• Dependent values
• Runtime value(if supported)

// ID: rawLog-{ loadData("data/raw.csv") }
val rawLog = cp.nocache { loadData("data/raw.csv") }
// ID: count-{ rawLog => rawLog.count() }(rawLog-{ loadData("data/raw.csv") })
val count = cp.cache(rawLog) { rawLog => rawLog.count() }
// Primitive values could be dependent value.
// ID: lit:10
val userId = 10
// ID: theUsersLogs-{ case (rawLog, userId) => rawLog.where('user_id === userId) }((rawLog-{ loadData("data/raw.csv") }, lit:10))
val theUsersLogs = cp.cache((rawLog, userId)) { case (rawLog, userId) =>
  rawLog.where('user_id === userId)
}

Cache location could specified by --cache option. Default is ~/.scalanb/cache/

--cache=file:path=/path/to/cache
--cache=hdfs:path=/path/to/cache # requires scalanb-spark

Cache file spec

• {root}/{namespace}/{name}
  • `{hex digest}
    • cache.json: metadata(TODO)
    • data: Serialized data(Format is type specific)

Plot using evilplot

To integrate EvilPlot, use this snippet:

import com.cibo.evilplot.plot
import plot.aesthetics.DefaultTheme._
implicit val plotFormat = nb.Format[plot.Plot] { plot =>
  val img = plot.render().asBufferedImage
  val buf = new java.io.ByteArrayOutputStream()
  val png = javax.imageio.ImageIO.write(img, "png", buf)
  buf.close()
  nb.Value.binary("image/png", buf.toByteArray)
}

And you can embed plot in notebook:

import com.cibo.evilplot.numeric.Point
val data = (0.0 until 1.0 by 0.02).map { v =>
  (v, v * scala.util.Random.nextDouble)
}.toSeq
plot.LinePlot(data.map { case (x, y) => Point(x, y) })
  .xAxis()
  .yAxis()
  .frame()
  .xLabel("x")
  .yLabel("y")