Aiven's GCS Sink Connector for Apache Kafka®

Important

The Aiven GCS Connector for Apache Kafka development has been moved to https://github.com/Aiven-Open/commons-for-apache-kafka-connect/

This is a sink Apache Kafka Connect connector that stores Kafka messages in a Google Cloud Storage (GCS) bucket.

The connector requires Java 11 or newer for development and production.

How It Works

The connector subscribes to the specified Kafka topics and collects messages coming in them and periodically dumps the collected data to the specified bucket in GCS.

Sometimes—for example, on reprocessing of some data—the connector will overwrite files that are already in the bucket. You need to ensure the bucket doesn't have a retention policy that prohibits overwriting.

The following object permissions must be enabled in the bucket:

• storage.objects.create;
• storage.objects.delete (needed for overwriting).

File name format

The connector uses the following format for output files (blobs): <prefix><filename>.

<prefix> is the optional prefix that can be used, for example, for subdirectories in the bucket.

<filename> is the file name. The connector has the configurable template for file names. It supports placeholders with variable names: {{ variable_name }}. Currently, supported variables are:

• topic - the Kafka topic;
• partition:padding=true|false - the Kafka partition, if padding set to true it will set leading zeroes for offset, the default value is false;
• start_offset:padding=true|false - the Kafka offset of the first record in the file, if padding set to true it will set leading zeroes for offset, the default value is false;
• timestamp:unit=yyyy|MM|dd|HH - the timestamp of when the Kafka record has been processed by the connector.
  • unit parameter values:
    • yyyy - year, e.g. 2020 (please note that YYYY is deprecated and is interpreted as yyyy)
    • MM - month, e.g. 03
    • dd - day, e.g. 01
    • HH - hour, e.g. 24
• key - the Kafka key.

To add zero padding to Kafka offsets, you need to add additional parameter padding in the start_offset variable, which value can be true or false (the default). For example: {{topic}}-{{partition}}-{{start_offset:padding=true}}.gz will produce file names like mytopic-1-00000000000000000001.gz.

To add zero padding to partition number, you need to add additional parameter padding in the partition variable, which value can be true or false (the default). For example: {{topic}}-{{partition:padding=true}}-{{start_offset}}.gz will produce file names like mytopic-0000000001-1.gz.

To add formatted timestamps, use timestamp variable.
For example: {{topic}}-{{partition}}-{{start_offset}}-{{timestamp:unit=yyyy}}{{timestamp:unit=MM}}{{timestamp:unit=dd}}.gz will produce file names like mytopic-2-1-20200301.gz.

To configure the time zone for the timestamp variable, use file.name.timestamp.timezone property. Please see the description of properties in the "Configuration" section.

Only the certain combinations of variables and parameters are allowed in the file name template (however, variables in a template can be in any order). Each combination determines the mode of record grouping the connector will use. Currently, supported combinations of variables and the corresponding record grouping modes are:

• topic, partition, start_offset, and timestamp - grouping by the topic, partition, and timestamp;
• key - grouping by the key.

If the file name template is not specified, the default value is {{topic}}-{{partition}}-{{start_offset}} (+ .gz when compression is enabled).

Record grouping

Incoming records are being grouped until flushed.

Grouping by the topic and partition

In this mode, the connector groups records by the topic and partition. When a file is written, an offset of the first record in it is added to its name.

For example, let's say the template is {{topic}}-part{{partition}}-off{{start_offset}}. If the connector receives records like

topic:topicB partition:0 offset:0
topic:topicA partition:0 offset:0
topic:topicA partition:0 offset:1
topic:topicB partition:0 offset:1
flush

there will be two files topicA-part0-off0 and topicB-part0-off0 with two records in each.

Each flush produces a new set of files. For example:

topic:topicA partition:0 offset:0
topic:topicA partition:0 offset:1
flush
topic:topicA partition:0 offset:2
topic:topicA partition:0 offset:3
flush

In this case, there will be two files topicA-part0-off0 and topicA-part0-off2 with two records in each.

Grouping by the key

In this mode, the connector groups records by the Kafka key. It always puts one record in a file, the latest record that arrived before a flush for each key. Also, it overwrites files if later new records with the same keys arrive.

This mode is good for maintaining the latest values per key as files on GCS.

Let's say the template is k{{key}}. For example, when the following records arrive

key:0 value:0
key:1 value:1
key:0 value:2
key:1 value:3
flush

there will be two files k0 (containing value 2) and k1 (containing value 3).

After a flush, previously written files might be overwritten:

key:0 value:0
key:1 value:1
key:0 value:2
key:1 value:3
flush
key:0 value:4
flush

In this case, there will be two files k0 (containing value 4) and k1 (containing value 3).

The string representation of a key

The connector in this mode uses the following algorithm to create the string representation of a key:

1. If key is null, the string value is "null" (i.e., string literal null).
2. If key schema type is STRING, it's used directly.
3. Otherwise, Java .toString() is applied.

If keys of you records are strings, you may want to use org.apache.kafka.connect.storage.StringConverter as key.converter.

Warning: Single key in different partitions

The group by key mode primarily targets scenarios where each key appears in one partition only. If the same key appears in multiple partitions the result may be unexpected.

For example:

topic:topicA partition:0 key:x value:aaa
topic:topicA partition:1 key:x value:bbb
flush

file kx may contain aaa or bbb, i.e. the behavior is non-deterministic.

Data format

Output files are text files that contain one record per line (i.e., they're separated by \n) except PARQUET format

There are four types of data format available:

• [Default] Flat structure, where field values are separated by comma (csv)

  Configuration: format.output.type=csv. Also, this is the default if the property is not present in the configuration.

• Complex structure, where file is in format of JSON lines. It contains one record per line and each line is a valid JSON object(jsonl)

  Configuration: format.output.type=jsonl.

• Complex structure, where file is a valid JSON array of record objects.

  Configuration: format.output.type=json.

• Complex structure, where file is in Apache Parquet file format.

  Configuration: format.output.type=parquet.

• Complex structure, where file is in Apache Avro Container File file format.

  Configuration: format.output.type=avro.

The connector can output the following fields from records into the output: the key, the value, the timestamp, the offset and headers. (The set of these output fields is configurable.) The field values are separated by comma.

It is possible to control the number of records to be put in a particular output file by setting file.max.records. By default, it is 0, which is interpreted as "unlimited".

CSV Format example

The key and the value—if they're output—are stored as binaries encoded in Base64.

For example, if we output key,value,offset,timestamp, a record line might look like:

a2V5,TG9yZW0gaXBzdW0gZG9sb3Igc2l0IGFtZXQ=,1232155,1554210895

It is possible to control the encoding of the value field by setting format.output.fields.value.encoding to base64 or none.

If the key, the value or the timestamp is null, an empty string will be output instead:

,,,1554210895

NB!

• The key.converter property must be set to org.apache.kafka.connect.converters.ByteArrayConverter or org.apache.kafka.connect.storage.StringConverter for this data format.

• The value.converter property must be set to org.apache.kafka.connect.converters.ByteArrayConverter for this data format.

JSONL Format example

For example, if we output key,value,offset,timestamp, a record line might look like:

{ "key": "k1", "value": "v0", "offset": 1232155, "timestamp":"2020-01-01T00:00:01Z" }

OR

{ "key": "user1", "value": {"name": "John", "address": {"city": "London"}}, "offset": 1232155, "timestamp":"2020-01-01T00:00:01Z" }

It is recommended to use

• org.apache.kafka.connect.storage.StringConverter,
• org.apache.kafka.connect.json.JsonConverter, or
• io.confluent.connect.avro.AvroConverter.

as key.converter and/or value.converter to make output files human-readable.

NB!

• The value of the format.output.fields.value.encoding property is ignored for this data format.
• Value/Key schema will not be presented in output file, even if value.converter.schemas.enable property is true. But, it is still important to set this property correctly, so that connector could read records correctly.

JSON Format example

For example, if we output key,value,offset,timestamp, an output file might look like:

[
{ "key": "k1", "value": "v0", "offset": 1232155, "timestamp":"2020-01-01T00:00:01Z" },
{ "key": "k2", "value": "v1", "offset": 1232156, "timestamp":"2020-01-01T00:00:05Z" }
]

OR

[
{ "key": "user1", "value": {"name": "John", "address": {"city": "London"}}, "offset": 1232155, "timestamp":"2020-01-01T00:00:01Z" }
]

It is recommended to use

• org.apache.kafka.connect.storage.StringConverter,
• org.apache.kafka.connect.json.JsonConverter, or
• io.confluent.connect.avro.AvroConverter.

as key.converter and/or value.converter to make output files human-readable.

NB!

• The value of the format.output.fields.value.encoding property is ignored for this data format.
• Value/Key schema will not be presented in output file, even if value.converter.schemas.enable property is true. But, it is still important to set this property correctly, so that connector could read records correctly.

NB!

For both JSON and JSONL another example could be for a single field output e.g. value, a record line might look like:

{ "value": "v0" }

OR

{ "value": {"name": "John", "address": {"city": "London"}} }

In this case it sometimes make sense to get rid of additional JSON object wrapping the actual value using format.output.envelope. Having format.output.envelope=false can produce the following output:

"v0"

OR

{"name": "John", "address": {"city": "London"}}

Parquet format example

For example, if we output key,offset,timestamp,headers,value, an output Parquet schema might look like this:

{
    "type": "record", "fields": [
      {"name": "key", "type": "RecordKeySchema"},
      {"name": "offset", "type": "long"},
      {"name": "timestamp", "type": "long"},
      {"name": "headers", "type": "map"},
      {"name": "value", "type": "RecordValueSchema"}
  ]
}

where RecordKeySchema - a key schema and RecordValueSchema - a record value schema. This means that in case you have the record and key schema like:

Key schema:

{
  "type": "string"
}

Record schema:

{
    "type": "record", "fields": [
      {"name": "foo", "type": "string"},
      {"name": "bar", "type": "long"}
  ]
}

the final Avro schema for Parquet is:

{
    "type": "record", "fields": [
      {"name": "key", "type": "string"},
      {"name": "offset", "type": "long"},
      {"name": "timestamp", "type": "long"},
      {"name": "headers", "type": "map", "values": "long"},
      { "name": "value",
        "type": "record",
        "fields": [
          {"name": "foo", "type": "string"},
          {"name": "bar", "type": "long"}
        ]
      }
  ]
}

For a single-field output e.g. value, a record line might look like:

{ "value": {"name": "John", "address": {"city": "London"}} }

In this case it sometimes make sense to get rid of additional JSON object wrapping the actual value using format.output.envelope. Having format.output.envelope=false can produce the following output:

{"name": "John", "address": {"city": "London"}}

NB!

• The value of the format.output.fields.value.encoding property is ignored for this data format.
• Due to Avro limitation message headers values must be the same datatype
• If you use org.apache.kafka.connect.json.JsonConverter be sure that you message contains schema. E.g. possible JSON message:

  {
    "schema": {
      "type": "struct",
      "fields": [
        {"type":"string", "field": "name"}
      ]
    }, "payload": {"name":  "foo"}
  }

• Connector works just fine with and without Schema Registry
• format.output.envelope=false is ignored if the value is not of type org.apache.avro.Schema.Type.RECORD or org.apache.avro.Schema.Type.MAP.

Avro format example

The output file is an Avro Object Container File.

For example, if we output key,offset,timestamp,headers,value, an output Avro schema might look like this:

{
    "type": "record", "fields": [
      {"name": "key", "type": "RecordKeySchema"},
      {"name": "offset", "type": "long"},
      {"name": "timestamp", "type": "long"},
      {"name": "headers", "type": "map"},
      {"name": "value", "type": "RecordValueSchema"}
  ]
}

where RecordKeySchema - a key schema and RecordValueSchema - a record value schema. This means that in case you have the record and key schema like:

Key schema:

{
  "type": "string"
}

Record schema:

{
    "type": "record", "fields": [
      {"name": "foo", "type": "string"},
      {"name": "bar", "type": "long"}
  ]
}

the final Avro schema for output is:

{
    "type": "record", "fields": [
      {"name": "key", "type": "string"},
      {"name": "offset", "type": "long"},
      {"name": "timestamp", "type": "long"},
      {"name": "headers", "type": "map", "values": "long"},
      { "name": "value",
        "type": "record",
        "fields": [
          {"name": "foo", "type": "string"},
          {"name": "bar", "type": "long"}
        ]
      }
  ]
}

For a single-field output e.g. value, a record line might look like:

{ "value": {"name": "John", "address": {"city": "London"}} }

In this case it sometimes make sense to get rid of additional object wrapping the actual value using format.output.envelope. Having format.output.envelope=false can produce the following output:

{"name": "John", "address": {"city": "London"}}

NB!

• The value of the format.output.fields.value.encoding property is ignored for this data format.
• Due to Avro limitation message headers values must be the same datatype
• Connector works just fine with and without Schema Registry
• format.output.envelope=false is ignored if the value is not of type org.apache.avro.Schema.Type.RECORD or org.apache.avro.Schema.Type.MAP.
• The Avro Object Container File requires that each value is written with the same schema in the file. When schema evolution happens for the input data, a new output file is created on every schema change. When data with previous and new schema is interleaved in the source topic multiple files will get generated in short duration.
• The schema for output file is derived from the Connect Schema. The Connect Schema is derived from the input records Avro schema by using the Schema Registry.

Retry strategy configuration property

There are six configuration properties to configure retry strategy exist.

Apache Kafka connect retry strategy properties

• kafka.retry.backoff.ms - The retry backoff in milliseconds. This config is used to notify Apache Kafka Connect to retry delivering a message batch or performing recovery in case of transient exceptions. Maximum value is 24 hours.

Google Cloud Storage retry strategy

• gcs.retry.backoff.initial.delay.ms - Initial retry delay in milliseconds. This config controls the delay before the first retry. Subsequent retries will use this value adjusted according to the gcs.retry.backoff.delay.multiplier. The default value is 1000 ms.
• gcs.retry.backoff.delay.multiplier - Retry delay multiplier. This config controls the change in retry delay. The retry delay of the previous call is multiplied by it to calculate the retry delay for the next call. The default value is 2.0.
• gcs.retry.backoff.max.delay.ms - Maximum retry delay in milliseconds. This config puts a limit on the value of the retry delay, so that the gcs.retry.backoff.delay.multiplier value can't increase the retry delay higher than this amount. The default value is 32 000 ms.
• gcs.retry.backoff.total.timeout.ms - Retry total timeout in milliseconds. This config controls over how long the logic should keep trying the remote call until it gives up completely. The default value is 50 000 ms. The maximum value is 24 hours.
• gcs.retry.backoff.max.attempts - Retry max attempts. This config defines the maximum number of attempts to perform. The default value is 6.

Configuration

Here you can read about the Connect workers configuration and here, about the connector Configuration.

Here is an example connector configuration with descriptions:

### Standard connector configuration

## Fill in your values in these:

# Unique name for the connector.
# Attempting to register again with the same name will fail.
name=my-gcs-connector

## These must have exactly these values:

# The Java class for the connector
connector.class=io.aiven.kafka.connect.gcs.GcsSinkConnector

# The key converter for this connector
key.converter=org.apache.kafka.connect.storage.StringConverter

# The value converter for this connector
value.converter=org.apache.kafka.connect.json.JsonConverter

# Identify, if value contains a schema.
# Required value converter is `org.apache.kafka.connect.json.JsonConverter`.
value.converter.schemas.enable=false

# The type of data format used to write data to the GCS output files.
# The supported values are: `csv`, `json`, `jsonl` and `parquet`.
# Optional, the default is `csv`.
format.output.type=jsonl

# A comma-separated list of topics to use as input for this connector
# Also a regular expression version `topics.regex` is supported.
# See https://kafka.apache.org/documentation/#connect_configuring
topics=topic1,topic2

### Connector-specific configuration
### Fill in you values

# The name of the GCS bucket to use
# Required.
gcs.bucket.name=my-gcs-bucket

## The following three options are used to specify GCP credentials.
## See the overview of GCP authentication:
##  - https://cloud.google.com/docs/authentication/
##  - https://cloud.google.com/docs/authentication/production
## If none are present, the connector will default to trying to connect without credentials.
## If only one is present, the connector will use it to get the credentials.
## If more than one is present, this is an error.

# The path to a GCP credentials file.
# Optional, the default is null.
gcs.credentials.path=/some/path/google_credentials.json

# GCP credentials as a JSON object.
# Optional, the default is null.
gcs.credentials.json={"type":"...", ...}

# Autodiscover GCP Credentials from the execution environment
gcs.credentials.default=true
##

# The value of object metadata Content-Encoding.
# This can be used for leveraging storage-side de-compression before download.
# Optional, the default is null.
gcs.object.content.encoding=gzip

# The set of the fields that are to be output, comma separated.
# Supported values are: `key`, `value`, `offset`, `timestamp`, and `headers`.
# Optional, the default is `value`.
format.output.fields=key,value,offset,timestamp,headers

# The option to enable/disable wrapping of plain values into additional JSON object(aka envelope)
# Optional, the default value is `true`.
format.output.envelope=true

# The prefix to be added to the name of each file put on GCS.
# See the GCS naming requirements https://cloud.google.com/storage/docs/naming
# Optional, the default is empty.
file.name.prefix=some-prefix/

# The compression type used for files put on GCS.
# The supported values are: `gzip`, `snappy`, `zstd`, `none`.
# Optional, the default is `none`.
file.compression.type=gzip

# The compression used for Avro Container File blocks.
# The supported values are: `bzip2`, `deflate`, `null`, `snappy`, `zstandard`.
# Optional, the default is `null`.
avro.codec=null

# The time zone in which timestamps are represented.
# Accepts short and long standard names like: `UTC`, `PST`, `ECT`,
# `Europe/Berlin`, `Europe/Helsinki`, or `America/New_York`.
# For more information please refer to https://docs.oracle.com/javase/tutorial/datetime/iso/timezones.html.
# The default is `UTC`.
file.name.timestamp.timezone=Europe/Berlin

# The source of timestamps.
# Supports only `wallclock` which is the default value.
file.name.timestamp.source=wallclock

# The file name template.
# See "File name format" section.
# Optional, the default is `{{topic}}-{{partition:padding=false}}-{{start_offset:padding=false}}` or
# `{{topic}}-{{partition:padding=false}}-{{start_offset:padding=false}}.gz` if the compression is enabled.
file.name.template={{topic}}-{{partition:padding=true}}-{{start_offset:padding=true}}.gz

Getting releases

The connector releases are available in the Releases section.

Release JARs are available in Maven Central:

<dependency>
  <groupId>io.aiven</groupId>
  <artifactId>gcs-connector-for-apache-kafka</artifactId>
  <version>x.y.z</version>
</dependency>

Development

Developing together with Common Module for Apache Kafka Connect library

This project depends on Common Module for Apache Kafka Connect. Normally, an artifact from a globally accessible repository is used. However, if you need to introduce changes to both this connector and Common Module for Apache Kafka Connect library at the same time, you should short-circuit the development loop via locally published artifacts. Please follow these steps:

1. Checkout the main HEAD of Common Module for Apache Kafka Connect.
2. Ensure the version here is with -SNAPSHOT prefix.
3. Make changes to Common Module for Apache Kafka Connect.
4. Publish it locally with ./gradlew publishToMavenLocal.
5. Change the version in the connector's build.gradle (ext.aivenConnectCommonsVersion) to match the published snapshot version of Common Module for Apache Kafka Connect.

After that, the latest changes you've done to Common Module for Apache Kafka Connect will be used.

When you finish developing the feature and is sure Common Module for Apacha Kafka Connect won't need to change:

1. Make a proper release of Common Module for Apache Kafka Connect.
2. Publish the artifact to the currently used globally accessible repository.
3. Change the version of Common Module for Apache Kafka Connect in the connector to the published one.

Integration testing

Integration tests are implemented using JUnit, Gradle and Docker.

To run them, you need:

• a GCS bucket with the read-write permissions;
• Docker installed.

In order to run the integration tests, execute from the project root directory:

./gradlew clean integrationTest -PtestGcsBucket=test-bucket-name

where PtestGcsBucket is the name of the GCS bucket to use.

The default GCP credentials will be used during the test (see the GCP documentation and the comment in GCP SDK code). This can be overridden either by setting the path to the GCP credentials file or by setting the credentials JSON string explicitly. (See Configuration section for details).

To specify the GCS credentials path, use gcsCredentialsPath property:

./gradlew clean integrationTest -PtestGcsBucket=test-bucket-name \
    -PgcsCredentialsPath=/path/to/credentials.json

To specify the GCS credentials JSON, use gcsCredentialsJson property:

./gradlew clean integrationTest -PtestGcsBucket=test-bucket-name \
    -PgcsCredentialsJson='{type":"...", ...}'

Gradle allows setting properties using environment variables, for example, ORG_GRADLE_PROJECT_testGcsBucket=test-bucket-name. See more about the ways to set properties here.

Releasing

TBD

License

This project is licensed under the Apache License, Version 2.0.

Trademarks

Apache Kafka, Apache Kafka Connect are either registered trademarks or trademarks of the Apache Software Foundation in the United States and/or other countries. Google Cloud Storage (GCS) is a trademark and property of their respective owners. All product and service names used in this website are for identification purposes only and do not imply endorsement.