scoozie

Scala DSL on top of Oozie XML

Why?

Scoozie is designed to solve problems concerning developer productivity when creating and running oozie workflow jobs.

Problems that Scoozie solves:

Job hierarchy and re-use: No more copy/pasting XML to create a new job
Type and syntax checking: No more errors from typos in XML
Safety / Verification: Scala provides a type-safe environment. Various other verificaitons and sanity checks are made. For example, it is impossible to create a cyclic graph or provide a fork/join structure that is not supported by oozie.
Developer Overhead: The Scoozie DSL code makes it easy to read and write workflows - no more messy XML.
Multiple sources of truth: The Scoozie specification of a job will be the source of truth for any workflow

Version:

As of version 0.5.6, scoozie has been upgraded to cdh5.3.3. Please depend on an earlier scoozie version to support lower versions of CDH libraries.

How it works:

There are three main layers to Scoozie:

Specification of job in DSL. This encapsulates the syntactic sugar that makes it easy to read and write Scoozie workflows. The general format for a Scoozie workflow is as follows:
```
val job1 = JobType(params) dependsOn Start
val job2 = JobType(params) dependsOn job2
val job3 = JobType(params) dependsOn job3, etc
```
This is the code that a worflow developer will actually be writing. The intent is to make it easy to read and to see the dependency path between jobs. Scoozie will automatically figure out forks/joins, etc. Specific samples of this can be found in src/main/scala/samples.scala
Logic / conversion from DSL spec to workflow graph. This is the real workhorse of Scoozie, as it converts the programmer-specified workflow into an intermediate graph of workflow nodes. Forks, joins, and decisions are figured out and made ready for the final step - conversion to XML. This step also includes verification on the result graph.
Conversion from intermediate graph to XML. This step makes use of scalaxb, an XML data-binding tool for scala. In this step each node in the intermediate graph is converted to an oozie-specific type. Finally, these nodes are converted to XML. This XML can be run in oozie by a scala command, examples of which can be found in src/main/scala/runSamples.scala

DSL Class Structure:

Node: Contains work and a list of dependencies. A node is also a dependency.

Dependency: The parent of everything that a workflow job could depend on. (nodes or decisions)

Work: Any work that a workflow node performs. This could be End, Kill, Jobs, sub workflows, etc.

Job: Extends Work. This is the parent of the typical oozie jobs. (HiveJob, MapReduceJob, DistCpJob, etc). Specific jobs are easy to extend and customize in order to fit the client's needs.

In practice, each line of developer-specified Scoozie code will return a Node, which other Nodes will depend on (Decisions are a special case of this, which will simply return a Dependency).

Tutorial:

Hello World!
The following workflow will create an Oozie workflow that will create a Hello World directory on HDFS.
```
def HelloWorldWorkflow = {
    val mkHelloWorld = FsJob(
        name = "make-hello-world",
        tasks = List(
            MkDir("${nameNode}/users/tmp/oozie-fun/hello-world_${wf:id()}")
        )
    ) dependsOn Start
    val done = End dependsOn mkHelloWorld
    Workflow("hello-world-wf", done)
}
```
This is an example of the basic structure of every scoozie workflow. Likely every scoozie workflow you will write will begin with a job that depends on Start and ends with an "End" job. The dependency chain makes it possible to specify the entire workflow by only referencing the end node in the Workflow object creation.
Getting more complicated - using forks & joins.
Scoozie is smart enough to figure out fork and join structures by looking at the dependencies between workflow nodes. This reduces cognitive load as for each node the developer only needs to think about what that node depends on.
For example, consider the following workflow:
```
def SimpleForkJoin = {
    val first = MapReduceJob("first") dependsOn Start
    val secondA = MapReduceJob("secondA") dependsOn first
    val secondB = MapReduceJob("secondB") dependsOn first
    val third = MapReduceJob("third") dependsOn (secondA, secondB)
    Workflow("simple-fork-join", third)
}
```
Scoozie will automatically create a workflow structured as first -> fork -> secondA/secondB -> join -> third

In addition, Scoozie will alert you by throwing an error if the workflow you have specified is not allowed by Oozie (for example, two nodes from different threads being joined)
Making things interesting - Decisions
Often a developer needs to specify variable paths in a workflow that are run conditionally. This is allowed in Scoozie via Decision and OneOf.
```
def DecisionExample = {
	val first = MapReduceJob("first") dependsOn Start
	val decision = Decision(
	    "route1" -> Predicates.BooleanProperty("${doRoute1}")
	) dependsOn first 
	val route1Start = MapReduceJob("r1Start") dependsOn (decision option "route1")
	val route1End = MapReduceJob("r1End") dependsOn route1Start
	val route2Start = MapReduceJob("r2Start") dependsOn (decision default)
	val route2End = MapReduceJob("r2End") dependsOn route2Start
	val last = MapReduceJob("last") dependsOn OneOf(route1End, route2End)
	val done = End dependsOn last
	Workflow("simple-decision", done)
}
```
In this example, route1 (route1Start -> route1End) will be run if ${doRoute1} is evaluated to true. Otherwise, route2 (route2Start -> route2End) will be run. Again, this reduces developer cognitive overhead as the developer only needs to consider the dependencies for one node at a time. Clearly "last" depends on only one of route1 and route2.
More complex decision structures are supported by Oozie as well. Verification will be automatically performed on the decisions to make sure that the developer has included all routes and default routes for each decision.
1. Pour some sugar on me - "doIf" and "Optional"
  A common design pattern for decisions is to have an "optional" route that may or may not be inserted into the workflow. Scoozie has provided some sugar for making this special case easy to think about and specify in code.
```
def SugarOption = {
    val first = MapReduceJob("first") dependsOn Start
    val option = MapReduceJob("option") dependsOn first doIf "doOption"
    val second = MapReduceJob("second") dependsOn Optional(option)
    val done = End dependsOn second
    Workflow("sugar-option-decision", done)
}
```
This syntax makes the semantics behind the optional node more clear than thinking about a separate decision. In this example, "option" is run if the "${doOption}" argument is evaluated to true (scoozie will figure out the brackets), and "second" must come after "first" and optionally "option".
Scalability, Modularity - Sub Workflows
Scoozie allows the developer to control the granularity of his project by using and defining sub-workflows at will. Using the previous SugarOption workflow as an example, we can easily create a new, more complex workflow without worrying about its specifics.
```
def SubWfExample = {
	val begin = MapReduceJob("begin") dependsOn Start
	val someWork = MapReduceJob("someWork") dependsOn begin
	val subwf = SugarOption dependsOn someWork
	val end = End dependsOn subwf
	Workflow("sub-wf-example", end)
}
```
Misc. Cool Things
1. Specifying custom error-to paths on nodes. Scoozie also makes it easy to provide customizations such as error-to paths on nodes (rather than erroring to "kill" or "fail")
```
def CustomErrorTo = {
    val first = MapReduceJob("first") dependsOn Start
    val errorPath = MapReduceJob("error") dependsOn (first error)
    val second = MapReduceJob("second") dependsOn first
    val end = End dependsOn OneOf(second, errorPath)
}
```
This scoozie code will create a workflow containing three nodes. The "first" node will proceed to "second" in the optimal case, but will now proceed to "errorPath" in the case of an error. Additionally, the OneOf clause is used again here and has an intuitive meaning.
Creating and Running Workflows in Scoozie
Scoozie makes running workflows easy. Simply create an object that extends ScoozieApp, pass in your workflow and any parameters, and you've got a runable main class that will generate and run an oozie workflow.
```
object RunExample extends ScoozieApp(SubWfExample)
```
Scoozie uses sbt-assembly in order to create an executable jar. Simply type
```
 > assembly
```
from sbt and all of the necessary dependencies and code will be assembled into a single jar: "scoozie-assembly-0.x.jar" You can run your workflow by typing
```
java -cp scoozie-assembly-0.x.jar RunExample -arg1=foo -arg2=bar ...
```
More examples of creating runnable scoozie objects can be found in runSamples.scala

Additional Help

More example scoozie code can be found in samples.scala. Available job definitions can be found in jobs.scala. These jobs can be easily extended to fit your specific use cases.

maheshsv/scoozie