Kafka Connect ArangoDB Sink Connector

Supported versions

This connector is compatible with:

• Kafka 3.4 and higher versions
• JDK 8 and higher versions
• ArangoDB 3.11.1 and higher versions

Installation

Download the Jar file from Maven Central  and copy it into one of the directories that are listed in the Kafka Connect worker’s plugin.path configuration property. This must be done on each of the installations where Kafka Connect will be run.

Once installed, you can then create a connector configuration file with the connector’s settings, and deploy that to a Connect worker. See the configuration documentation for the available options.

For more detailed plugin installation instructions, see the Confluent Documentation .

Connection handling

Task connections to the database are evenly distributed across all available ArangoDB Coordinators. If a connectivity error occur, a connection is re-established with a different Coordinator.

Available Coordinators can be periodically monitored by setting connection.acquireHostList.enabled to true and optionally configuring the monitoring interval.

Over time, due to connection failovers, the distribution of task connections across the Coordinators may become uneven. To address this, connections are periodically re-balanced across the available Coordinators. The rebalancing interval can be configured via the connection.rebalance.interval.ms property. The default is 30 minutes.

Delivery guarantees

This connector guarantees that each record in the Kafka topic is delivered at least once. For example, the same record could be delivered multiple times in the following scenarios:

• Transient errors in the communication between the connector and the database system, leading to retries
• Errors in the communication between the connector and Kafka, preventing to commit offsets of already written records
• Abrupt termination of connector task

When restarted, the connector resumes reading from the Kafka topic at an offset prior to where it stopped. As a result, at least in the cases mentioned above, some records might get written to ArangoDB more than once. Even if configured for idempotent writes (e.g. with insert.overwriteMode=replace), writing the same record multiple times still updates the _rev field of the document.

Note that in case of retries, Ordering Guarantees are still provided.

To improve the likelihood that every write survives even in case of a DB-Server failover, consider configuring the configuration property insert.waitForSync (default false), which determines whether the write operations are synced to disk before returning.

Error handling

The connector categorizes all the possible errors into two types, data errors and transient errors.

Data errors

Data errors are unrecoverable and caused by the data being processed. For example:

• Conversion errors:
  • Illegal key type
  • Illegal value type
• Server validation errors:
  • Illegal _key, _from, _to values
  • JSON schema validation errors
• Server constraints violations
  • Unique index violations
  • Key conflicts (in case of insert.overwriteMode=conflict)

The configuration property data.errors.tolerance allows you to configure the behavior for tolerating data errors:

• none: data errors result in an immediate connector task failure (default)
• all: changes the behavior to skip over records generating data errors. If DLQ is configured, then the record is reported (see Dead Letter Queue).

Data errors detection can be further customized via the extra.data.error.nums configuration property. In addition to the cases listed above, the server errors reporting errorNums listed by this configuration property are considered data errors.

Transient errors

Transient errors are recoverable and may succeed if retried with some delay (see Retries). If all retries fail, then the connector task fails.

All errors that are not data errors are considered transient errors.

Retries

In case of transient errors, the max.retries configuration property determines how many times the connector retries.

The retry.backoff.ms configuration property allows you to set the time in milliseconds to wait following an error before a retry attempt is made.

Data errors are never retried.

Dead Letter Queue

This connector supports the Dead Letter Queue (DLQ) functionality. For information about accessing and using the DLQ, see Confluent Platform Dead Letter Queue .

Only data errors can be reported to the DLQ. Transient errors, after potential retries, always make the task fail.

You can enable DLQ support for data errors by setting data.errors.tolerance=all and errors.deadletterqueue.topic.name.

Multiple tasks

The connector can scale out by running multiple tasks in parallel as specified by the tasks.max  configuration parameter. In this case, each connector task handles the records from different Kafka topics partitions.

Data mapping

The sink connector optionally supports schemas. For example, the Avro converter that comes with Schema Registry, the JSON converter with schemas enabled, or the Protobuf converter.

Kafka record keys and Kafka record value field _key, if present, must be a primitive type of either:

• string
• integral numeric (integer)

The record value must be either:

• struct (Kafka Connect structured record)
• map
• null (tombstone record)

Record values are converted to JSON objects using Kafka Connect JsonConverter, ensuring compatibility with Kafka Connect data types and support to schemas. If the data in the topic is not of a compatible format, applying an SMT  or implementing a custom converter may be necessary.

Key handling

The _key of the documents inserted into ArangoDB is derived in the following way:

1. Use the Kafka record value field _key if present and not null, else
2. Use the Kafka record key if not null, else
3. Use the Kafka record coordinates (topic-partition-offset)

Delete mode

The connector can delete documents in a database collection when it consumes a tombstone record, which is a Kafka record that has a non-null key and a null value. This behavior is disabled by default, meaning that any tombstone records results in a failure of the connector.

You can enable deletes with delete.enabled=true.

Delete operations are idempotent and do not throw errors if the target document does not exist.

Enabling delete mode does not affect the insert.overwriteMode.

Write modes

The insert.overwriteMode configuration parameter allow you to set the write behavior in case a document with the same _key already exists:

• conflict: the new document value is not written and an exception is thrown (default)
• ignore: the new document value is not written
• replace: the existing document is overwritten with the new document value
• update: the existing document is patched (partially updated) with the new document value

Idempotent writes

All the write modes supported are idempotent, with the exception that the document revision field (_rev) changes every time a document is written. See Document revisions for more details.

If there are failures, the Kafka offset used for recovery may not be up-to-date with what was committed as of the time of the failure, which can lead to re-processing during recovery. In case of insert.overwriteMode=conflict (default), this can lead to constraint violations errors if records need to be re-processed.

Ordering guarantees

Kafka records in the same Kafka topic partition that are mapped to documents with the same _key (see Key handling) are written to ArangoDB in the same order as they are in the Kafka topic partition.

The order of the writes for records in the same Kafka partition that are mapped to documents with different _key is not guaranteed.

The order between writes for records in different Kafka partitions is not guaranteed.

To guarantee documents in ArangoDB are eventually consistent with the records in the Kafka topic, it is recommended deriving the document _key from Kafka record keys and using a key-based partitioner that assigns the same partition to records with the same key (i.e. Kafka default partitioner).

Otherwise, in case the document _key is assigned from Kafka record value field _key, the same could be achieved using a field partitioner on _key.

When restarted, the connector may resume reading from the Kafka topic at an offset prior to where it stopped. This can lead to reprocessing of batches containing multiple Kafka records that are mapped to documents with the same _key. In such case, it is possible to observe the related document in the database being temporarily updated to older versions and eventually to newer ones.

Monitoring

The Kafka Connect framework exposes basic status information over a REST interface. Fine-grained metrics, including the number of processed messages and the rate of processing, are available via JMX. For more information, see Monitoring Kafka Connect and Connectors  (published by Confluent, but equally applies to a standard Apache Kafka distribution).

SSL

To connect to ArangoDB using an SSL connection, you must set the ssl.enabled configuration property to true.

Certificate from file

The connector can load the trust store to be used from file. The following configuration properties can be used:

• ssl.truststore.location: the location of the trust store file
• ssl.truststore.password: the password for the trust store file

Note that the trust store file path needs to be accessible at the same given location from all Kafka Connect workers.

Certificate from configuration property value

The connector can accept the SSL certificate value from configuration property encoded as Base64 string. The following configuration properties can be used:

• ssl.cert.value: Base64-encoded SSL certificate

See SSL configuration for further options.

Limitations

• Record values are required to be object-like structures (DE-644)
• Auto-creation of ArangoDB collection is not supported (DE-653)
• Batch writes are not guaranteed to be executed serially (FRB-300)
• Batch writes may succeed for some documents while failing for others (FRB-300) This has two important consequences:
  • Transient errors might be retried and succeed at a later point
  • Data errors might be asynchronously reported to the DLQ