Provided code docs for the relevant language of the I/O. This should also have links to any external sources of information within the Apache Beam site or external documentation location.

• Java doc
• Python doc
• Go doc

Add a new page under I/O connector guides that covers specific tips and configurations. The following shows those for Parquet, Hadoop and others.

Examples:

Formatting of the section headers in your Javadoc/Pythondoc should be consistent throughout such that programmatic information extraction for other pages can be enabled in the future.

Example subset of sections to include in your page in order:

1. Before you start
2. {Connector}IO basics
3. Supported Features
   1. Relational
4. Authentication
5. Reading from {Connector}
6. Writing to {Connector}
7. Resource scalability
8. Limitations
9. Reporting an Issue

Example:

I/O Connectors should include a table under Supported Features subheader that indicates the Relational Features utilized.

Relational Features are concepts that can help improve efficiency and can optionally be implemented by an I/O Connector. Using end user supplied pipeline configuration (SchemaIO) and user query (FieldAccessDescriptor) data, relational theory is applied to derive improvements such as faster pipeline execution, lower operation costs and less data read/written.

Example table:

<div class="table-container-wrapper">
<table class="table table-bordered table-io-standards-relational-features">
   <tr>
      <th>
         <p><strong>Relational Feature</strong>
      </th>
      <th>
         <p><strong>Supported</strong>
      </th>
      <th>
         <p><strong>Notes</strong>
      </th>
   </tr>
   <tr>
      <td>
         <p>Column Pruning
      </td>
      <td>
         <p>Yes/No
      </td>
      <td>
         <p>To Be Filled
      </td>
   </tr>
   <tr>
      <td>
         <p>Filter Pushdown
      </td>
      <td>
         <p>Yes/No
      </td>
      <td>
         <p>To Be Filled
      </td>
   </tr>
   <tr>
      <td>
         <p>Table Statistics
      </td>
      <td>
         <p>Yes/No
      </td>
      <td>
         <p>To Be Filled
      </td>
   </tr>
   <tr>
      <td>
         <p>Partition Metadata
      </td>
      <td>
         <p>Yes/No
      </td>
      <td>
         <p>To Be Filled
      </td>
   </tr>
   <tr>
      <td>
         <p>Metastore
      </td>
      <td>
         <p>Yes/No
      </td>
      <td>
         <p>To Be Filled
      </td>
   </tr>
</table>
</div>

Example implementations:

Add a page under Common pipeline patterns, if necessary, outlining common usage patterns involving your I/O.

Update I/O Connectors with your I/O’s information

Example:

https://beam.apache.org/documentation/io/connectors/#built-in-io-connectors

Provide setup steps to use the I/O, under a Before you start header.

Example:

Include a canonical read/write code snippet after the initial description for each supported language. The below example shows Hadoop with examples for Java.

Example:

Indicate how timestamps for elements are assigned. This includes batch sources to allow for future I/Os which may provide more useful information than current_time().

Outline any temporary resources (for example, files) that the connector will create.

Example:

BigQuery batch loads first create a temp GCS location

Provide, under an Authentication subheader, how to acquire partner authorization material to securely access the source/sink.

Example:

https://beam.apache.org/documentation/io/built-in/snowflake/#authentication

Here BigQuery names it permissions but the topic covers similarities

I/Os should provide links to the Source/Sink documentation within Before you start header.

Example:

Indicate if there is native or X-language support in each language with a link to the docs.

Example:

Indicate known limitations under a Limitations header. If the limitation has a tracking issue, please link it inline.

Example:

https://beam.apache.org/documentation/io/built-in/snowflake/#limitations

I/O (not built-in)

Custom I/Os are not included in the Apache Beam Github repository. Some examples would be SolaceIO.

Update the Other I/O Connectors for Apache Beam table with your information.

The aformentioned table

This section outlines API syntax, semantics and recommendations for features that should be adopted for new as well as existing Apache Beam I/O Connectors.

The I/O Connector development guidelines are written with the following principles in mind:

• Consistency makes an API easier to learn.
  • If there are multiple ways of doing something, we should strive to be consistent first
• With a couple minutes of studying documentation, users should be able to pick up most I/O connectors.
• The design of a new I/O should consider the possibility of evolution.
• Transforms should integrate well with other Beam utilities.

All SDKs

Pipeline Configuration / Execution / Streaming / Windowing semantics guidelines

An I/O should rarely rely on a PipelineOptions subclass to tune internal parameters.

A source must return elements in the GlobalWindow unless explicitly parameterized in the API by the user.

• ReadFromIO(window_by=...)
• ReadFromIO.IntoFixedWindows(...)
• ReadFromIO(apply_windowing=True/False) (e.g. PeriodicImpulse)
• IO.read().withWindowing(...)
• IO.read().windowBy(...)
• IO.read().withFixedWindows(...)

A sink should be Window agnostic and handle elements sent with any Windowing method, unless explicitly parameterized or expressed in its API.

A sink may change windowing of a PCollection internally in any way. However, the metadata that it returns as part of its result object must be:

• must be in the same window as the input, unless explicitly decared otherwise in the API.
• must have accurate timestamps
• may contain additional information about windowing (e.g. a BigQuery job may have a timestamp, but also a window associated with it).

• WriteToIO(triggering_frequency=...) - e.g. WriteToBigQuery (This only sets the windowing within the transform - input data is still in the Global Window).
• WriteBatchesToIO(...)
• WriteWindowsToIO(...)

A streaming sink (or any transform accessing an external service) may implement throttling of its requests to prevent from overloading the external service.

• Per-key fixed throttling
• Adaptive throttling with sink-reported backpressure
• Ramp-up throttling from a start point

TODO: Tracking Issue

Java

General

The primary class used in working with the connector should be named {connector}IO

Example:

The class should be placed in the package org.apache.beam.sdk.io.{connector}

Example:

The unit/integration/performance tests should live under the package org.apache.beam.sdk.io.{connector}.testing. This will cause the various tests to work with the standard user-facing interfaces of the connector.

Unit tests should reside in the same package (i.e. org.apache.beam.sdk.io.{connector}), as they may often test internals of the connector.

An I/O transform should avoid receiving user lambdas to map elements from a user type to a connector-specific type. Instead, they should interface with a connector-specific data type (with schema information when possible).

When necessary, an I/O transform should receive a type parameter that specifies the input type (for sinks) or output type (for sources) of the transform.

It is highly discouraged to directly expose third-party libraries in the public API part of the I/O Connector for the following reasons:

• It reduces Apache Beam’s compatibility guarantees - Changes to third-party libraries can/will directly break existing user’s pipelines.
• It makes code maintainability hard - If libraries are directly exposed at API level, a dependency change will require multiple changes throughout the I/O implementation code
• It forces third-party dependencies onto end users

Instead, we highly recommend exposing Beam-native interfaces and an adaptor that holds mapping logic.

Source and Sinks should be abstracted with a PTransform wrapper, and internal classes be declared protected or private. By doing so implementation details can be added/changed/modified without breaking implementation by dependencies.

Classes / Methods / Properties

Gives access to the class that represents reads within the I/O. The Read class should implement a fluent interface similar to the fluentbuilder pattern (e.g. withX(...).withY(...)). Together with default values, it provide a fail-fast (with immediate validation feedback after each .withX()) that is slightly less verbose than the builder pattern.

A few different sources implement runtime configuration for reading from a data source. This is a valuable pattern because it enables a purely batch source to become a more sophisticated streaming source.

As much as possible, this type of transform should have the type richness of a construction-time-configured transform:

• Support Beam Row output with a schema known at construction-time.
• Extra configuration may be needed (and acceptable) in this case (e.g. a SchemaProvider parameter, a Schema parameter, a Schema Catalog or a utility of that sort).
• The input PCollection should have a fixed type with a schema, so it can be easily manipulated by users.

Example:

Gives access to the class that represents writes within the I/O. The Write class should implement a fluent interface pattern (e.g. withX(...).withY(...)) as described further above for IO.Read.

Some data storage and external systems implement APIs that do not adjust easily to Read or Write semantics (e.g. FhirIO implements several different transforms that fetch or send data to Fhir).

These classes should be added only if it is impossible or prohibitively difficult to encapsulate their functionality as part of extra configuration of Read, Write and ReadAll transforms, to avoid increasing the cognitive load on users.

Some connectors rely on other user-facing classes to set configuration parameters.

(e.g. JdbcIO.DataSourceConfiguration). These classes should be nested within the {Connector}IO class.

The top-level I/O class will provide a static method to start constructing an I/O.Write transform. This returns a PTransform with a single input PCollection, and a Write.Result output.

This method should not specify in its name any of the following:

• Internal data format
• Strategy used to write data
• Input or output data type

The method to start constructing an I/O.Read transform. This returns a PTransform with a single output PCollection.

This method should not specify in its name any of the following:

• Internal data format
• Strategy used to read data
• Output data type

The above should be specified via configuration parameters if possible. If not possible, then a new static method may be introduced, but this must be exceptional, and documented in the I/O header as part of the API.

A Read transform must provide a from method where users can specify where to read from. If a transform can read from different kinds of sources (e.g. tables, queries, topics, partitions), then multiple implementations of this from method can be provided to accommodate this:

• IO.Read from(Query query)
• IO.Read from(Table table) / from(String table)
• IO.Read from (Topic topic)
• IO.Read from(Partition partition)

The input type for these methods can reflect the external source’s API (e.g. Kafka TopicPartition should use a Beam-implemented TopicPartition object).

• IO.Read from(String table)
• IO.Read fromQuery(String query)

A Write transform must provide a to method where users can specify where to write data. If a transform can write to different kinds of sources while still using the same input element type(e.g. tables, queries, topics, partitions), then multiple implementations of this from method can be provided to accommodate this:

• IO.Write to(Query query)
• IO.Write to(Table table) / from(String table)
• IO.Write to(Topic topic)
• IO.Write to(Partition partition)

The input type for these methods can reflect the external sink's API (e.g. Kafka TopicPartition should use a Beam-implemented TopicPartition object).

If different kinds of destinations require different types of input object types, then these should be done in separate I/O connectors.

• IO.Write to(String table)
• IO.Write toTable(String table)

A write transform may enable writing to more than one destination. This can be a complicated pattern that should be implemented carefully (it is the preferred pattern for connectors that will likely have multiple destinations in a single pipeline).

The preferred pattern for this is to define a DynamicDestinations interface (e.g. BigQueryIO.DynamicDestinations) that will allow the user to define all necessary parameters for the configuration of the destination.

class IO.Read.withX

withX provides a method for configuration to be passed to the Read method, where X represents the configuration to be created. With the exception of generic with statements ( defined below ) the I/O should attempt to match the name of the configuration option with that of the option name in the source.

These methods should return a new instance of the I/O rather than modifying the existing instance.

Example:

IO.Read.withConfigObject

Some connectors in Java receive a configuration object as part of their configuration. This pattern is encouraged only for particular cases. In most cases, a connector can hold all necessary configuration parameters at the top level.

To determine whether a multi-parameter configuration object is an appropriate parameter for a high level transform, the configuration object must:

• Hold only properties related to the connection/authentication parameters for the external data store (e.g. JdbcIO.DataSourceConfiguration).
  • Generally, secrets should not be passed as parameters, unless an alternative is not feasible. For secret management, a secret-management service or KMS is the recommended approach.
• Or mirror an API characteristic from the external data source (e.g. KafkaIO.Read.withConsumerConfigUpdates), without exposing that external API in the Beam API.
  • The method should mirror the name of the API object (e.g. given an object SubscriptionStatConfig, a method would be withSubscriptionStatConfig).
• Or when a connector can support different configuration ‘paths’ where a particular property requires other properties to be specified (e.g. BigQueryIO’s method will entail various different properties). (see last examples).

Example:

JdbcIO.DataSourceConfiguration, SpannerConfig, KafkaIO.Read.withConsumerConfigUpdates

BigQueryIO.write()
  .withWriteConfig(FileLoadsConfig.withAvro()
                                 .withTriggeringFrequency()...)

BigQueryIO.write()
  .withWriteConfig(StreamingInsertsConfig.withDetailedError()
                                  .withExactlyOnce().etc..)

Discouraged - except for dynamic destinations

For sources that can receive Beam Row-typed PCollections, the format function should not be necessary, because Beam should be able to format the input data based on its schema.

For sinks providing Dynamic Destination functionality, elements may carry data that helps determine their destination. These data may need to be removed before writing to their final destination.

• Show that it’s not possible to perform data matching automatically.
• Support Dynamic Destinations and need changes to the input data due to that reason.

IO.Read.withCoder

Strongly Discouraged

Sets the coder to use to encode/decode the element type of the output / input PCollection of this connector. In general, it is recommended that sources will:

1. Return Row objects with a schema that is automatically inferred.
2. Automatically set the necessary coder by having fixed output/input types, or inferring their output/input types.

Connector transforms should provide a method to override the interface between themselves and the external system that they communicate with. This can enable various uses:

Sets the coder to use to encode/decode the element type of the output / input PCollection of this connector. In general, it is recommended that sources will:

• Local testing by mocking the destination service
• User-enabled metrics, monitoring, and security handling in the client.
• Integration testing based on emulators

Example:

BigQueryIO.Write.withTestServices(BigQueryServices)

Types

The expand method of a Read transform must return a PCollection object with a type. The type may be parameterized or fixed to a class.

The type of the PCollection will usually be one of the following four options. For each of these option, the encoding / data is recommended to be as follows:

• A pre-defined, basic Java type (e.g. String)
  • This encoding should be simple, and use a simple Beam coder (e.g. Utf8StringCoder)
• A pre-set POJO type (e.g. Metadata) with a schema
  • The preferred strategy for these is to define the output type as an @AutoValue, with @DefaultSchema and @SchemaCreate annotations. This will ensure compact, fast encoding with RowCoder.
• A Beam Row with a specific schema
• A type with a schema that’s not known at construction time

The expand method of any write transform must return a type IO.Write.Result object that extends a PCollectionTuple. This object allows transforms to return metadata about the results of its writing and allows this write to be followed by other PTransforms.

If the Write transform would not need to return any metadata, a Write.Result object is still preferable, because it will allow the transform to evolve its metadata over time.

Examples of metadata:

• Failed elements and errors
• Successfully written elements
• API tokens from calls issued by the transform

Examples:

BigQueryIO’s WriteResult

Evolution

Over time, I/O need to evolve to address new use cases, or use new APIs under the covers. Some examples of necessary evolution of an I/O:

• A new data type needs to be supported within it (e.g. any-type partitioning in JdbcIO.ReadWithPartitions)
• A new backend API needs to be supported

In general, one should resist adding a completely new static method for functionality that can be captured as configuration within an existing method.

An example of too many top-level methods that could be supported via configuration is PubsubIO

A new top-level static method should only be added in the following cases:

• A new PTransform is being added that cannot / does not make sense to be supported as a simple configuration parameter of the existing PTransforms (e.g. FhirIO’s various transforms, any ReadAll transform).
• A feature that represents a new recommended standard to use the transform (e.g. JdbcIO’s ReadWithPartitions)
• A change in the usual way of interacting with this transform, or a parameter that’s necessary to initialize the transform (e.g. BigQueryIO.read(...) method vs read())

Python

General

If the I/O lives in Apache Beam it should be placed in the package apache_beam.io.{connector} or apache_beam.io.{namespace}.{connector}

Example:

There will be a module named {connector}.py which is the primary entry point used in working with the connector in a pipeline apache_beam.io.{connector} or apache_beam.io.{namespace}.{connector}

Example:

apache_beam.io.gcp.bigquery / apache_beam/io/gcp/bigquery.py

If the I/O implementation exists in a single module (a single file), then the file {connector}.py can hold it.

Otherwise, the connector code should be defined within a directory (connector package) with an __init__.py file that documents the public API.

If the connector defines other files containing utilities for its implementation, these files must clearly document the fact that they are not meant to be a public interface.

Classes / Methods / Properties

This gives access to the PTransform to read from a given data source. It allows you to configure it via the arguments that it receives. For long lists of optional parameters, they may be defined as parameters with a default value.

Q. Java uses a builder pattern. Why can’t we do that in Python? Optional parameters can serve the same role in Python.

Example:

As much as possible, this type of transform should have the type richness and safety of a construction-time-configured transform:

• Support output with a schema known at construction-time
  • Extra configuration may be needed (and acceptable) in this case (e.g. a SchemaProvider parameter, a Schema parameter, a Schema Catalog or a utility of that sort).
• The input PCollection should have a fixed type with a schema, so it can be easily manipulated by users.

Example:

This gives access to the PTransform to write into a given data sink. It allows you to configure it via the arguments that it receives. For long lists of optional parameters, they may be defined as parameters with a default value.

Q. Java uses a builder pattern. Why can’t we do that in Python? Optional parameters can serve the same

role in Python.

Example:

parameter ReadFrom{Connector}({source})

The first parameter in a Read or Write I/O connector must specify the source for readers or the destination for writers.

1. Retain a single argument, but auto-infer the source/sink type (e.g. pandas.read_sql(...) supporting tables and queries)
2. Add a new argument for each possible source/sink type (e.g. ReadFromBigQuery having query/table parameters)

A write transform may enable writing to more than one destination. This can be a complicated pattern that should be implemented carefully (it is the preferred pattern for connectors that will likely have multiple destinations in a single pipeline).

The preferred API pattern in Python is to pass callables (e.g. WriteToBigQuery) for all parameters that will need to be configured. In general, examples of callable parameters may be:

• Destination callable → Should receive an element, and return a destination for that element
• Other examples
  • Schema callable → Should receive a destination and return a schema for the destination
  • Format function → Should receive a record (and maybe a destination) and format the record to be inserted.

Using these callables also allows maintainers to add new parameterizable callables over time (with default values to avoid breaking existing users) that will define extra configuration parameters if necessary.

parameter ReadFromIO(param={param_val})

Discouraged

Some connectors in Python may receive a complex configuration object as part of their configuration. This pattern is discouraged, because a connector can hold all necessary configuration parameters at the top level.

To determine whether a multi-parameter configuration object is an appropriate parameter for a high level transform, the configuration object must:

• Hold only properties related to the connection/authentication parameters for the external data store (e.g. ReadFromKafka(consumer_config parameter))
• Or mirror an API characteristic from the external data source (e.g. ReadFromDebezium(connector_properties parameter))

Types

The expand method of a Read transform must return a PCollection object with a type, and be annotated with the type. Preferred PCollection types in Python are (in order of preference):

Simple Python types if the (bytes, str, numbers)

1. A NamedTuple or DataClass with a set schema, encoded with RowCoder
2. A Python dictionary
   1. The dictionaries should be encoded via RowCoder, if possible.
3. A preset Python class, if a schema is not possible

The expand method of any write transform must return a Python object with a fixed class type. The recommended name for the class is WriteTo{IO}Result. This object allows transforms to return metadata about the results of its writing.

If the Write transform would not need to return any metadata, a Python object with a class type is still preferable, because it will allow the transform to evolve its metadata over time.

Examples of metadata:

• Failed elements and errors
• Successfully written elements
• API tokens from calls issued by the transform

Example:

BigQueryIO’s WriteResult

The expand method of a Write transform must return a PCollection object with a type, and be annotated with the type. Preferred PCollection types in Python are the same as the output types for a ReadFromIO referenced in T1.

GoLang

General

If the I/O lives in Apache Beam it should be placed in the package:

{connector}io

Example:

Integration and Performance tests should live under the same package as the I/O itself

{connector}io

Typescript

Classes / Methods / Properties

The method to start constructing an I/O.Write transform.

Testing

An I/O should have unit tests, integration tests, and performance tests. In the following guidance we explain what each type of test aims to achieve, and provide a baseline standard of test coverage. Do note that the actual test cases and business logic of the actual test would vary depending on specifics of each source/sink but we have included some suggested test cases as a baseline.

This guide complements the Apache Beam I/O transform testing guide by adding specific test cases and scenarios. For general information regarding testing Beam I/O connectors, please refer to that guide.

Integration and performance tests should live under the package org.apache.beam.sdk.io.{connector}.testing. This will cause the various tests to work with the standard user-facing interfaces of the connector.

Unit tests should reside in the same package (i.e. org.apache.beam.sdk.io.{connector}), as they may often test internals of the connector.

Unit Tests

I/O unit tests need to efficiently test the functionality of the code. Given that unit tests are expected to be executed many times over multiple test suites (for example, for each Python version) these tests should execute relatively fast and should not have side effects. We recommend trying to achieve 100% code coverage through unit tests.

When possible, unit tests are favored over integration tests due to faster execution time and low resource usage. Additionally, unit tests can be easily included in pre-commit tests suites (for example, Jenkins beam_PreCommit_* test suites) hence has a better chance of discovering regressions early. Unit tests are also preferred for error conditions.

The unit testing class should be part of the same package as the IO and named {connector}IOTest.

Example:

sdks/java/io/cassandra/src/test/java/org/apache/beam/sdk/io/cassandra/CassandraIOTest.java

Suggested Test Cases

BigtableIOTest.testReading

pubsub_test.TestReadFromPubSub.test_read_messages_success

BigtableIOTest.testWriting

BigTableIOTest.testReadWithBigTableOptionsSetsRetryOptions

AvroIOTest.testReadDisplayData

DatastoreV1Test.testReadDisplayData

BigqueryIOReadTest.estBigQueryQuerySourceInitSplit

BigTableIOTest.testReadingSplitAtFractionExhaustive

Reading a PCollection<Row> or writing a PCollection<Row>

BigQueryIOReadTest.testReadTableWithSchema

BigQueryIOWriteTest.testWriteValidatesDataset

SpannerIOReadTest.testReadMetrics

SpannerIOReadTest.readAllPipeline

WatermarkPolicyTest.shouldAdvanceWatermarkWithTheArrivalTimeFromKinesisRecords

Integration Tests

Integration tests test end-to-end interactions between the Beam runner and the data store a given I/O connects to. Since these usually involve remote RPC calls, integration tests take a longer time to execute. Additionally, Beam runners may use more than one worker when executing integration tests. Due to these costs, an integration test should only be implemented when a given scenario cannot be covered by a unit test.

The integration testing class should be part of the same package as the I/O and named {connector}IOIT.

For example:

sdks/java/io/cassandra/src/test/java/org/apache/beam/sdk/io/cassandra/CassandraIOIT.java

Suggested Test Cases

Same as “write then read” but for sources that support reading a PCollection of source configs. All future (SDF) sources are expected to support this.

KafkaIOIT.testKafkaIOReadsAndWritesCorrectlyInStreaming

Performance Tests

Because the Performance testing framework is still in flux, performance tests can be a follow-up submission after the actual I/O code.

The Performance testing framework does not yet support GoLang or Typescript.

Performance benchmarks are a critical part of best practices for I/Os as they effectively address several areas:

• To evaluate if the cost and performance of a specific I/O or dataflow template meets the customer’s business requirements.
• To illustrate performance regressions and improvements to I/O or dataflow templates between code changes.
• To help end customers estimate costs and plan capacity to meet their SLOs.

Dashboard

Google runs performance tests routinely for built-in I/Os and publishes them to an externally viewable dashboard for Java and Python.

Guidance

Include a Resource Scalability section into your page under Built-in I/O connector guides documentation which will indicate the upper bounds which the IO has integration tests for.

For example:

An indication that KafkaIO has integration tests with xxxx topics. The documentation can state if the connector authors believe that the connector can scale beyond the integration test number, however this will make it clear to the user the limits of the tested paths.

Include expected performance characteristics of the I/O based on performance tests that the connector has in place.