Skip to content

A library enabling DAG structuring of data processing programs such as ETLs

License

Notifications You must be signed in to change notification settings

noleme/noleme-flow

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Noleme Flow

Maven Build Maven Central Repository javadoc coverage GitHub FOSSA Status

This library provides features enabling DAG structuring of data processing programs such as ETLs.

Implementations found in this package shouldn't be tied to any specific Noleme project.

Note: This library is considered as "in beta" and as such significant API changes may occur without prior warning.

I. Installation

Add the following in your pom.xml:

<dependency>
    <groupId>com.noleme</groupId>
    <artifactId>noleme-flow</artifactId>
    <version>0.18</version>
</dependency>

II. Notes on Structure and Design

The core idea behind this library is to structure the program as a DAG of "actions" to perform, then compile that graph into concrete runnable instance which properties will depend on the implementation.

These actions can be of three different types, mirroring an ETL process:

  • Extractor for introducing data into the flow, they would typically be connectors for a variety of data sources
  • Transformer for manipulating data and returning an altered version of it, or new data inferred from the input
  • Loader for dumping data out of the flow

Additionally, there are two actions related to stream flows:

  • Generator for gradually introducing data into the flow, they can be iterating through a Collection, reading off an InputStream or generating data on-the-fly
  • Accumulator for doing the reverse operation, they accumulate all outputs from a stream flow and continue with a standard (ie. non-stream) flow

Once a DAG has been defined, a FlowCompiler will be responsible for transforming the DAG representation into a runnable instance, a FlowRuntime.

At the time of this writing, there are two available implementations:

  • the serial PipelineRuntime which will run one node after another, making sure each one can satisfy its input
  • the parallel ParallelRuntime which will attempt to run any node that can satisfy its input in a parallel fashion

Once a FlowRuntime has been produced, we can simply run it.

TODO

III. Usage

First, let us start by the end and have a look at what it can look like "in practice".

Starting with a CSV located at data/my.csv such that:

key,value,metadata,flag
0,234,interesting,false
1,139,not_interesting,false
3,982,interesting,true
4,389,interesting,false
5,093,not_interesting,false

Below is a flow that will leverage tablesaw for transforming this local CSV, perform some transformations, and dump it back on the filesystem.

var flow = Flow
    .from(new FileStreamer(), "data/my.csv")
    .pipe(new TablesawCSVParser(tableProperties)) //tableProperties is a tablesaw-specific configuration classs, don't mind it
    .pipe(table -> table.where(t -> t.stringColumn("metadata").isEqualTo("interesting")))
    .pipe(table -> table.where(t -> t.booleanColumn("flag").isFalse()))
    .sink(new TablesawCSVWrite("data/my-filtered.csv"))
;

The overarching goal for noleme-flow is to have a simple yet flexible API that can enable both:

  • simplistic scenarios like this one, where ease of use and not being locked-in by a heavy ecosystem is paramount: noleme-flow aims to remain first and foremost a lightweight library enabling quick drafts
  • intermediate and complex scenarios joining multiple data-sources where noleme-flow's main goal shifts towards enabling better code reuse, with the help of noleme-vault for configuration management, by making it easy to bundle flow sequences for reuse into larger flow graphs

Implementations mentioned above can be found over at noleme-flow-connectors.

Going back, here is a very basic example of pipeline we could create:

/* We initialize a flow */
var flow = Flow
    .from(() -> 1)
    .pipe(i -> i + 1)
    .pipe(i -> i * 2)
    .sink(System.out::println)
;

/* We run it as a Pipeline */
Flow.runAsPipeline(flow);

Which, upon running should print 4.

Another example:

/* We initialize a flow */
var flow = Flow
    .from(() -> 2)
    .pipe(i -> i * 2)
;

/* We branch the flow in two branchs */
var branchA = flow.pipe(i -> i * i);
var branchB = flow.pipe(i -> i * 5);

/* We join the two branchs and collect the end result */
var recipient = branchA
    .join(branchB, Integer::sum)
    .pipe(i -> i * 2)
    .collect()
;

var output = Flow.runAsPipeline(flow);

System.out.println(output.get(recipient));

Upon running this should print 72 (2*((2*2)^2)+((2*2)*5)).

Now a final example with a stream flow going on:

/* Let's have a "standard" flow doing its thing */
var branch = Flow
    .from(() -> 2)
    .pipe(i -> i + 1)
;

/* Create a "stream" flow from a list of integers  */
var flow = Flow
    .from(() -> List.of(0, 1, 2, 3, 4, 5, 6, 7, 8, 9))
    .stream(IterableGenerator::new)
    .pipe(i -> i * i)
    .join(branch, (f, b) -> f * b) /* All values in the main flow will be multiplied by the output from the branch flow */
    .accumulate(values -> values.stream() /* Once the generator is exhausted and all stream nodes have run, we gather the output integers and sum them ; note that accumulation is optional (you could also end the stream with a sink) */
        .reduce(Integer::sum)
        .orElseThrow(() -> new AccumulationException("Could not sum data."))
    )
    .pipe(i -> i + 1) /* After the accumulation step, the flow is back to being a "standard" flow so we can queue further transformations */
    .sink(System.out::println)
;

Flow.runAsPipeline(flow);

Upon running this should print 856.

Note that noleme-flow itself doesn't provide any Generator implementation, but the IterableGenerator class mentioned above is part of the noleme-flow-connect-commons library (over there).

Other features that will need to be documented include:

  • the complete set of DAG building methods (including alternate flavours of from, stream, as well as driftSink, after and the generic into)
  • control-flow with partial DAG interruption (interrupt and interruptIf, nonFatal helpers)
  • runtime input management (dynamic from and the Input component)
  • runtime output management, sampling/collection features (collect, sample and the Output component)
  • stream flows and parallelization (setMaxParallelism and implementation-specific considerations)
  • flow slices (SourceSlice, PipeSlice and SinkSlice) for flow DAG fragments code reuse
  • ParallelRuntime service executor lifecycle and other considerations
  • DAG node naming for debugging purposes (appears in traces, logs)

TODO

IV. Dev Installation

This project will require you to have the following:

  • Java 11+
  • Git (versioning)
  • Maven (dependency resolving, publishing and packaging)

License

FOSSA Status