New Streamlet API for Heron

As of version 0.16.0, Heron offers a new Streamlet API that you can use to write topologies in a more declarative, functional manner, without needing to specify spout and bolt logic directly. The Streamlet API is currently in beta and available for Java. The Streamlet API for Python will be available soon.

More information on the Streamlet API can be found below.

A Heron topology is a directed acyclic graph (DAG) used to process streams of data. Topologies can be stateless or stateful depending on your use case.

Heron topologies consist of two basic components:

  • Spouts inject data into Heron topologies, potentially from external sources like pub-sub messaging systems (Apache Kafka, Apache Pulsar, etc.)
  • Bolts apply user-defined processing logic to data supplied by spouts

Spouts and bolts are connected to one another via streams of data. Below is a visual illustration of a simple Heron topology:

Heron topology

In the diagram above, spout S1 feeds data to bolts B1 and B2 for processing; in turn, bolt B1 feeds processed data to bolts B3 and B4, while bolt B2 feeds processed data to bolt B4. This is just a simple example; you can create arbitrarily complex topologies in Heron.

Creating topologies

There are currently two APIs available that you can use to build Heron topologies:

  1. The higher-level Heron Streamlet API (recommended for new topologies), which enables you to create topologies in a declarative, developer-friendly style inspired by functional programming concepts (such as map, flatMap, and filter operations)
  2. The lower-level topology API (not recommended for new topologies), based on the original Apache Storm API, which requires you to specify spout and bolt logic directly

Topology Lifecycle

Once you’ve set up a Heron cluster, you can use Heron’s CLI tool to manage the entire lifecycle of a topology, which typically goes through the following stages:

  1. Submit the topology to the cluster. The topology is not yet processing streams but is ready to be activated.
  2. Activate the topology. The topology will begin processing streams in accordance with the topology architecture that you’ve created.
  3. Restart an active topology if, for example, you need to update the topology configuration.
  4. Deactivate the topology. Once deactivated, the topology will stop processing but remain running in the cluster.
  5. Kill a topology to completely remove it from the cluster. It is no longer known to the Heron cluster and can no longer be activated. Once killed, the only way to run that topology is to re-submit and re-activate it.

Logical Plan

A topology’s logical plan is analagous to a database query plan in that it maps out the basic operations associated with a topology. Here’s an example logical plan for the example Streamlet API topology below:

Topology logical Plan

Whether you use the Heron Streamlet API or the topology API, Heron automatically transforms the processing logic that you create into both a logical plan and a physical plan.

Physical Plan

A topology’s physical plan is related to its logical plan but with the crucial difference that a physical plan determines the “physical” execution logic of a topology, i.e. how topology processes are divided between containers. Here’s a basic visual representation of a physical plan:

Topology Physical Plan

In this example, a Heron topology consists of one spout and five different bolts (each of which has multiple instances) that have automatically been distributed between five different containers.

The Heron Streamlet API

When Heron was first created, the model for creating topologies was deeply indebted to the Apache Storm model. Under that model, developers creating topologies needed to explicitly define the behavior of every spout and bolt in the topology. Although this provided a powerful low-level API for creating topologies, that approach presented a variety of drawbacks for developers:

  • Verbosity — In both the Java and Python topology APIs, creating spouts and bolts involved substantial boilerplate, requiring developers to both provide implementations for spout and bolt classes and also specify the connections between those spouts and bolts. This often led to the problem of…
  • Difficult debugging — When spouts, bolts, and the connections between them need to be created “by hand,” a great deal of cognitive load
  • Tuple-based data model — In the older topology API, spouts and bolts passed tuples and nothing but tuples within topologies. Although tuples are a powerful and flexible data type, the topology API forced all spouts and bolts to serialize or deserialize tuples.

In contrast with the topology API, the Heron Streamlet API offers:

  • Boilerplate-free code — Instead of re to implement spout and bolt classes, the Heron Streamlet API enables you to write functions, such as map, flatMap, join, and filter functions, instead.
  • Easy debugging — With the Heron Streamlet API, you don’t have to worry about spouts and bolts, which means that you can more easily surface problems with your processing logic.
  • Completely flexible, type-safe data model — Instead of requiring that all processing components pass tuples to one another (which implicitly requires serialization to and deserializaton from your application-specific types), the Heron Streamlet API enables you to write your processing logic in accordance with whatever types you’d like—including tuples, if you wish. In the Streamlet API for Java, all streamlets are typed (e.g. Streamlet<MyApplicationType>), which means that type errors can be caught at compile time rather than runtime.

Heron Streamlet API topologies

With the Heron Streamlet API you still create topologies, but only implicitly. Heron automatically performs the heavy lifting of converting the streamlet-based processing logic that you create into spouts and bolts and, from there, into containers that are then deployed using whichever scheduler your Heron cluster is using.

From the standpoint of both operators and developers managing topologies’ lifecycles, the resulting topologies are equivalent. From a development workflow standpoint, however, the difference is profound.


The core construct underlying the Heron Streamlet API is that of the streamlet. A streamlet is a potentially unbounded, ordered collection of tuples. Streamlets can originate from a wide variety of sources, such as pub-sub messaging systems like Apache Kafka and Apache Pulsar (incubating), random generators, or static files like CVS or Parquet files.

Streamlet operations

In the Heron Streamlet API, processing data means transforming streamlets into other streamlets. This can be done using a wide variety of available operations, including many that you may be familiar with from functional programming:

Operation Description
map Returns a new streamlet by applying the supplied mapping function to each element in the original streamlet
flatMap Like a map operation but with the important difference that each element of the streamlet is flattened
join Joins two separate streamlets into a single streamlet
filter Returns a new streamlet containing only the elements that satisfy the supplied filtering function

Streamlet API example

You can see an example streamlet-based processing graph in the diagram below:

Streamlet-based processing graph for Heron

Here’s the corresponding Java code for the processing logic shown in the diagram:

package heron.streamlet.example;

import com.twitter.heron.streamlet.*;
import com.twitter.heron.streamlet.impl.StreamletImpl;

import java.util.concurrent.ThreadLocalRandom;

public final class ExampleStreamletAPITopology {
    public ExampleStreamletAPITopology() {}

    private int randomInt(int lower, int upper) {
        return ThreadLocalRandom.current().nextInt(lower, upper + 1);

    public static void main(String[] args) {
        Builder builder = Builder.CreateBuilder();

        builder.newSource(() -> 0)

        builder.newSource(() -> randomInt(1, 10))
                .map(i -> i + 1)
                .filter(i -> i != 2)

        Config conf = new Config();

        new Runner().run("ExampleStreamletAPITopology", conf, builder);

That Java code will produce this logical plan:

Heron Streamlet API logical plan

Key-value streamlets

In order to perform some operations, such as streamlet joins and streamlet reduce operations, you’ll need to create key-value streamlets.


In the topology API, processing parallelism can be managed via adjusting the number of spouts and bolts performing different operations, enabling you to, for example, increase the relative parallelism of a bolt by using three of that bolt instead of two.

The Heron Streamlet API provides a different mechanism for controlling parallelism: partitioning. To understand partitioning, keep in mind that rather than physical spouts and bolts, the core processing construct in the Heron Streamlet API is the processing step. With the Heron Streamlet API, you can explicitly assign a number of partitions to each processing step in your graph (the default is one partition).

The example topology above, for example, has five steps:

  • the random integer source
  • the “add one” map operation
  • the union operation
  • the filtering operation
  • the logging operation.

You could apply varying numbers of partitions to each step in that topology like this:

Builder builder = Builder.CreateBuilder();

builder.newSource(() -> 0)

builder.newSource(() -> ThreadLocalRandom.current().nextInt(1, 11))
        .map(i -> i + 1)
        .filter(i -> i != 2)

The number of partitions to assign to each processing step when using the Streamlet API depends on a variety of factors.

Window operations

Windowed computations gather results from a topology or topology component within a specified finite time frame rather than, say, on a per-tuple basis.

Here are some examples of window operations:

  • Counting how many customers have purchased a product during each one-hour period in the last 24 hours.
  • Determining which player in an online game has the highest score during a 15-minute period.

Sliding windows

Sliding windows are windows that overlap, as in this figure:

Sliding time window

For sliding windows, you need to specify two things:

  1. The length or duration of the window (length if the window is a count window, duration if the window is a time window).
  2. The sliding interval, which determines when the window slides, i.e. at what point during the current window the new window begins.

In the figure above, the duration of the window is 10 seconds, while the sliding interval is 5 seconds. Each new window begins five seconds into the current window.

With sliding time windows, data can be processed in more than one window. Tuples 3, 4, and 5 above are processed in both window 1 and window 2 while tuples 6, 7, and 8 are processed in both window 2 and window 3.

Setting the duration of a window to 16 seconds and the sliding interval to 12 seconds would produce this window arrangement:

Sliding time window with altered time interval

Here, the sliding interval determines that a new window is always created 12 seconds into the current window.

Tumbling windows

Tumbling windows are windows that don’t overlap, as in this figure:

Tumbling time window

Tumbling windows don’t overlap because a new window doesn’t begin until the current window has elapsed. For tumbling windows, you only need to specify the length or duration of the window but no sliding interval.

With tumbling windows, data are never processed in more than one window because the windows never overlap.

Count windows

Count windows are specified on the basis of the number of operations rather than a time interval. A count window of 100 would mean that a window would elapse after 100 tuples have been processed, with no relation to clock time.

With count windows, this scenario (for a count window of 50) would be completely normal:

Window Tuples processed Clock time
1 50 10 seconds
2 50 12 seconds
3 50 1 hour, 12 minutes
4 50 5 seconds

Time windows

Time windows differ from count windows because you need to specify a time duration (in seconds) rather than a number of tuples processed.

With time windows, this scenario (for a time window of 30 seconds) would be completely normal:

Window Tuples processed Clock time
1 150 30 seconds
2 50 30 seconds
3 0 30 seconds
4 375 30 seconds

All window types

As explained above, windows differ along two axes: sliding (overlapping) vs. tumbling (non overlapping) and count vs. time. This produces four total types:

  1. Sliding time windows
  2. Sliding count windows
  3. Tumbling time windows
  4. Tumbling count windows

Resource allocation with the Heron Streamlet API

When creating topologies using the Streamlet API, there are three types of resources that you can specify:

  1. The number of containers into which the topology’s physical plan will be split
  2. The total number of CPUs allocated to be used by the topology
  3. The total amount of RAM allocated to be used by the topology

For each topology, there are defaults for each resource type:

Resource Default Minimum
Number of containers 1 1
CPU 1.0 1.0
RAM 512 MB 192MB

Allocating resources to topologies

For instructions on allocating resources to topologies, see the language-specific documentation for:


A Heron spout is a source of streams, responsible for emitting tuples into the topology. A spout may, for example, read data from a Kestrel queue or read tweets from the Twitter API and emit tuples to one or more bolts.

Information on building spouts can be found in Building Spouts.


A Heron bolt consumes streams of tuples emitted by spouts and performs some set of user-defined processing operations on those tuples, which may include performing complex stream transformations, performing storage operations, aggregating multiple streams into one, emitting tuples to other bolts within the topology, and much more.

Information on building bolts can be found in Building Bolts.

Data Model

Heron’s original topology API required using a fundamentally tuple-driven data model. You can find more information in Heron’s Data Model.