TypeScript Beam SDK

A library for writing Apache Beam pipelines in Typescript.

As well as being a fully-functioning SDK, it serves as a cleaner, more modern template for building SDKs in other languages (see README-dev.md for more details).

Getting started

The Typescript SDK can be installed with

npm install apache_beam

Due to its extensive use of cross-language transforms, it is recommended that Python 3 and Java be available on the system as well.

A fully working setup is provided as a clonable starter project on github.

Running a pipeline

Beam pipelines can be run on a variety of runners. The typical way to create a runner is with beam.runners.runner.create_runner({runner: "runnerType", ...}), as seen in the wordcount example.

After building, to run locally one can execute:

node path/to/main.js --runner=direct

To run against Flink, where the local infrastructure is automatically downloaded and set up:

node path/to/main.js --runner=flink

To run on Dataflow:

node path/to/main.js \
    --runner=dataflow \
    --project=${PROJECT_ID} \
    --tempLocation=gs://${GCS_BUCKET}/wordcount-js/temp --region=${REGION}

API

We generally try to apply the concepts from the Beam API in a TypeScript idiomatic way, but it should be noted that few of the initial developers have extensive (if any) JavaScript/TypeScript development experience, so feedback is greatly appreciated.

In addition, some notable departures are taken from the traditional SDKs:

• We take a "relational foundations" approach, where schema'd data is the primary way to interact with data, and we generally eschew the key-value requiring transforms in favor of a more flexible approach naming fields or expressions. JavaScript's native Object is used as the row type.

• As part of being schema-first we also de-emphasize Coders as a first-class concept in the SDK, relegating it to an advanced feature used for interop. Though we can infer schemas from individual elements, it is still TBD to figure out if/how we can leverage the type system and/or function introspection to regularly infer schemas at construction time. A fallback coder using BSON encoding is used when we don't have sufficient type information.

• We have added additional methods to the PCollection object, notably map and flatmap, rather than only allowing apply. In addition, apply can accept a function argument (PCollection) => ... as well as a PTransform subclass, which treats this callable as if it were a PTransform's expand.

• In the other direction, we have eliminated the problematic Pipeline object from the API, instead providing a Root PValue on which pipelines are built, and invoking run() on a Runner. We offer a less error-prone Runner.run which finishes only when the pipeline is completely finished as well as Runner.runAsync which returns a handle to the running pipeline.

• Rather than introduce PCollectionTuple, PCollectionList, etc. we let PValue literally be an array or object with PValue values which transforms can consume or produce. These are applied by wrapping them with the P operator, e.g. P([pc1, pc2, pc3]).apply(new Flatten()).

• Like Python, flatMap and ParDo.process return multiple elements by yielding them from a generator, rather than invoking a passed-in callback. TBD how to output to multiple distinct PCollections. There is currently an operation to split a PCollection into multiple PCollections based on the properties of the elements, and we may consider using a callback for side outputs.

• The map, flatMap, and ParDo.process methods take an additional (optional) context argument, which is similar to the keyword arguments used in Python. These are javascript objects whose members may be constants (which are passed as is) or special DoFnParam objects which provide getters to element-specific information (such as the current timestamp, window, or side input) at runtime.

• Rather than introduce multiple-output complexity into the map/do operations themselves, producing multiple outputs is done by following with a new Split primitive that takes a PCollection<{a?: AType, b: BType, ... }> and produces an object {a: PCollection<AType>, b: PCollection<BType>, ...}.

• JavaScript supports (and encourages) an asynchronous programing model, with many libraries requiring use of the async/await paradigm. As there is no way (by design) to go from the asynchronous style back to the synchronous style, this needs to be taken into account when designing the API. We currently offer asynchronous variants of PValue.apply(...) (in addition to the synchronous ones, as they are easier to chain) as well as making Runner.run asynchronous. TBD to do this for all user callbacks as well.

An example pipeline can be found at https://github.com/apache/beam/blob/master/sdks/typescript/src/apache_beam/examples/wordcount.ts and more documentation can be found in the beam programming guide.