Apache Beam Core Concepts

The Beam Model

Evolved from Google's MapReduce, FlumeJava, and Millwheel projects. Originally called the "Dataflow Model."

Protos

The /model directory defines the official, language-agnostic Protocol Buffer (.proto) and gRPC service specifications that establish the Beam Model and the Beam Portability Framework.

Why /model Exists (Portability & Decoupling)

Without a standardized model representation, supporting $N$ SDK languages across $M$ execution runners would require $N \times M$ separate translation layers. By defining all core pipeline concepts, data encodings, metrics, and worker RPC protocols as Protobuf messages and gRPC services, /model acts as the universal lingua franca:

• SDKs compile user pipelines into standardized Runner API protobuf graphs.

• Runners inspect, optimize, and distribute these graphs without needing SDK-specific language runtimes.

• Workers (SDK Harnesses) execute user code (DoFns) and communicate with runners over standardized Fn API gRPC channels.

Core Directories & What They Do

1. /model/pipeline (Runner API & Core Model): Defines the SDK- and runner-independent representation of pipelines (Pipeline, Components, PTransform, PCollection, Coder), timestamps/constants, Beam Schemas (Row, Field), and execution metrics (MonitoringInfo).
2. /model/fn-execution (Fn API & Provisioning): Defines bidirectional gRPC services between runners and worker SDK harnesses for bundle execution (Control), element streaming (Data), state/timer access (State), log forwarding (Logging), and container initialization (Provisioning).
3. /model/job-management (Job, Expansion, & Artifact APIs): Defines gRPC interfaces for submitting and monitoring jobs on remote servers (JobService), resolving cross-language transforms in remote SDKs (ExpansionService), and staging dependency artifacts or container images (ArtifactService).
4. /model/interactive (Interactive API): Defines metadata and stream headers for recording and replaying data in Interactive Beam notebooks.

What Agents Need to Pay Special Attention To

• Conservative Proto Changes: Proto changes are generally conservative and accepted only when there is a compelling reason and community consensus. Every addition introduces a new obligation that each SDK and runner must support; adding new Beam model elements (portable types, capabilities) increases the compatibility gap between SDK and runner capabilities.
• URNs Are the API Contract: Transforms, coders, windowing strategies, environments, and metrics are bound together by standardized string URNs (e.g., beam:transform:pardo:v1, beam:coder:bytes:v1). When inspecting or creating transforms across languages, always verify URN mappings and registry handlers in both the SDK and Runner runtimes.
• Strict Backward & Wire Compatibility:
  • Never renumber, delete, or modify existing field IDs or URN strings in .proto files, as they are used across distributed RPC boundaries and persisted checkpoints.
• Build System & Naming Collisions:
  • Modifying files in /model requires re-generating language bindings (e.g., ./gradlew :model:pipeline:generateProto).
  • Avoid protobuf field names that conflict with reserved keywords in target languages (e.g., class in Java or output in Python, as noted in beam_fn_api.proto comments).

Key Abstractions

Pipeline

A Pipeline encapsulates the entire data processing task, including reading, transforming, and writing data.

// Java
Pipeline p = Pipeline.create(options);
p.apply(...)
 .apply(...)
 .apply(...);
p.run().waitUntilFinish();

# Python
with beam.Pipeline(options=options) as p:
    (p | 'Read' >> beam.io.ReadFromText('input.txt')
       | 'Transform' >> beam.Map(process)
       | 'Write' >> beam.io.WriteToText('output'))

PCollection

A distributed dataset that can be bounded (batch) or unbounded (streaming).

Properties

• Immutable - Once created, cannot be modified
• Distributed - Elements processed in parallel
• May be bounded or unbounded
• Timestamped - Each element has an event timestamp
• Windowed - Elements assigned to windows

PTransform

A data processing operation that transforms PCollections.

// Java
PCollection<String> output = input.apply(MyTransform.create());

# Python
output = input | 'Name' >> beam.ParDo(MyDoFn())

Core Transforms

ParDo

General-purpose parallel processing.

// Java
input.apply(ParDo.of(new DoFn<String, Integer>() {
    @ProcessElement
    public void processElement(@Element String element, OutputReceiver<Integer> out) {
        out.output(element.length());
    }
}));

# Python
class LengthFn(beam.DoFn):
    def process(self, element):
        yield len(element)

input | beam.ParDo(LengthFn())
# Or simpler:
input | beam.Map(len)

GroupByKey

Groups elements by key.

PCollection<KV<String, Integer>> input = ...;
PCollection<KV<String, Iterable<Integer>>> grouped = input.apply(GroupByKey.create());

CoGroupByKey

Joins multiple PCollections by key.

Combine

Combines elements (sum, mean, etc.).

// Global combine
input.apply(Combine.globally(Sum.ofIntegers()));

// Per-key combine
input.apply(Combine.perKey(Sum.ofIntegers()));

Flatten

Merges multiple PCollections.

PCollectionList<String> collections = PCollectionList.of(pc1).and(pc2).and(pc3);
PCollection<String> merged = collections.apply(Flatten.pCollections());

Partition

Splits a PCollection into multiple PCollections.

Windowing

Types

• Fixed Windows - Regular, non-overlapping intervals
• Sliding Windows - Overlapping intervals
• Session Windows - Gaps of inactivity define boundaries
• Global Window - All elements in one window (default)

input.apply(Window.into(FixedWindows.of(Duration.standardMinutes(5))));

input | beam.WindowInto(beam.window.FixedWindows(300))

Triggers

Control when results are emitted.

input.apply(Window.<T>into(FixedWindows.of(Duration.standardMinutes(5)))
    .triggering(AfterWatermark.pastEndOfWindow()
        .withEarlyFirings(AfterProcessingTime.pastFirstElementInPane()
            .plusDelayOf(Duration.standardMinutes(1))))
    .withAllowedLateness(Duration.standardHours(1))
    .accumulatingFiredPanes());

Side Inputs

Additional inputs to ParDo.

PCollectionView<Map<String, String>> sideInput =
    lookupTable.apply(View.asMap());

mainInput.apply(ParDo.of(new DoFn<String, String>() {
    @ProcessElement
    public void processElement(ProcessContext c) {
        Map<String, String> lookup = c.sideInput(sideInput);
        // Use lookup...
    }
}).withSideInputs(sideInput));

Pipeline Options

Configure pipeline execution.

public interface MyOptions extends PipelineOptions {
    @Description("Input file")
    @Required
    String getInput();
    void setInput(String value);
}

MyOptions options = PipelineOptionsFactory.fromArgs(args).as(MyOptions.class);

Schema

Strongly-typed access to structured data.

@DefaultSchema(AutoValueSchema.class)
@AutoValue
public abstract class User {
    public abstract String getName();
    public abstract int getAge();
}

PCollection<User> users = ...;
PCollection<Row> rows = users.apply(Convert.toRows());

Error Handling

Dead Letter Queue Pattern

TupleTag<String> successTag = new TupleTag<>() {};
TupleTag<String> failureTag = new TupleTag<>() {};

PCollectionTuple results = input.apply(ParDo.of(new DoFn<String, String>() {
    @ProcessElement
    public void processElement(ProcessContext c) {
        try {
            c.output(process(c.element()));
        } catch (Exception e) {
            c.output(failureTag, c.element());
        }
    }
}).withOutputTags(successTag, TupleTagList.of(failureTag)));

results.get(successTag).apply(WriteToSuccess());
results.get(failureTag).apply(WriteToDeadLetter());

Cross-Language Pipelines

Use transforms from other SDKs.

# Use Java Kafka connector from Python
from apache_beam.io.kafka import ReadFromKafka

result = pipeline | ReadFromKafka(
    consumer_config={'bootstrap.servers': 'localhost:9092'},
    topics=['my-topic']
)

Best Practices

1. Prefer built-in transforms over custom DoFns
2. Use schemas for type-safe operations
3. Minimize side inputs for performance
4. Handle late data explicitly
5. Test with DirectRunner before deploying
6. Use TestPipeline for unit tests