CDC questions and patterns with Cassandra and Pulsar

You can check the CDC status of a table by running the following CQL query:

SELECT * FROM system_distributed.cdc_local WHERE keyspace_name = 'keyspace_name' AND table_name = 'table_name';

If the CDC status is enabled, then CDC is enabled on the table. If the CDC status is disabled then CDC is disabled on the table. If the CDC status is null then CDC is not enabled on the table.

If the CDC status is null, then you can enable CDC on the table by running the following CQL query:

ALTER TABLE keyspace_name.table_name WITH cdc = {'enabled': true};

If the CDC status is enabled, then you can disable CDC on the table by running the following CQL query:

ALTER TABLE keyspace_name.table_name WITH cdc = {'enabled': false};

If the CDC status is disabled, then you can enable CDC on the table by running the following CQL query:

ALTER TABLE keyspace_name.table_name WITH cdc = {'enabled': true};

You can check the status of the Cassandra agent by running the following CQL query:

SELECT * FROM system_distributed.cdc_local WHERE keyspace_name = 'cdc' AND table_name = 'raw_cdc';

The status column will be running if the agent is running. If the status column is null then the agent is not running. If the status column is stopped then the agent is not running.

If the status column is stopped then you can start the agent by running the following CQL query:

ALTER TABLE cdc.raw_cdc WITH cdc = {'enabled': true};

If the status column is null then you can start the agent by running the following CQL query:

ALTER TABLE cdc.raw_cdc WITH cdc = {'enabled': true};

If the status column is running then you can stop the agent by running the following CQL query:

ALTER TABLE cdc.raw_cdc WITH cdc = {'enabled': false};

Unacknowledged messages mean the CDC agent was not able to produce the event message in Pulsar. If this is the case the table row mutation will fail which the Cassandra client will then see an exception. So data will get committed to Cassandra and no event will be created.

Another scenario might be the Pulsar broker is too busy to process messages and a backlog has been created. In this case, Pulsar’s backlog policies take effect and event messages are handled accordingly. The data will be committed to Cassandra but there might be some additional latency to the event message creation.

The design of CDC in Cassandra assumed that when table changes are sync’d to the raw_cdc log, another process will be draining that log. There is a max log size setting that will disable writes to the table when the set threshold is reached. If a connection to the Pulsar cluster is needed for the log to be drained, and it’s not responsive, the log will being to fill, which can impact a table’s write availability.

For more, see the Scaling up your configuration section in the official documentation.

A dead letter topic is used when a message can’t be delivered to a consumer. Maybe the message acknowledgment time expired (no consumer acknowledged receipt of the message), or a consumer negatively acknowledged the message, or a retry letter topic is in use and retries were exhausted.

The Cassandra Source Connector creates a consumer to receive new event messages from the CDC agent, but does not configure a dead letter topic. It is assumed that parallel instances, broker compute, and function worker compute will be sized to handle the workload.

There are 3 areas of scalability to focus on. First are the hosts in the Cassandra cluster. The CDC agent is running on each host in its own JVM. If you are administering your own Cassandra cluster, then you can tune the JVM compute properties to handle the appropriate workload. If you are using Cassandra in a serverless environment, then the JVM is already set to handle significant load.

The second area of focus is the number of Cassandra Source Connector instances running. This is initially set when the Source Connector is created, and can be updated throughout the life of the running connector. Depending on your Pulsar configuration, an instance can represent a process thread on the broker or a function worker. If using Kubernetes, this could be a pod. Each represents different scaling strategies like increasing compute, adding more workers, and more K8s nodes.

Finally, the third area focuses on managing the broker backlog size and throughput tolerances. There are potentially a large amount of messages being created, so you must ensure the Pulsar cluster is sized correctly. Our Luna Streaming Production Cluster Sizing can help you understand this better.

Transformation functions are a great way to manipulate messages on CDC data (with no code required!) Put them inline to watch the data topic and write to a different topic. Call the topic something memorable like "filtered-data" topic.

Learn more about transformation functions here.

One of the requirements of CDC is that both the Cassandra and Pulsar clusters need to be in the same cloud region (or on-premise data center). If you are using geo-replication, you need the change data to be replicated across multiple clusters. The most manageable way to handle this is to use Pulsar’s Cassandra sink to "watch" the CDC data topic and write the change to a different Cassandra table (in another Org).

The Cassandra sink requires the following provisions:

Migrating data between tables solves quite a few different challenges. The basic approach is to use a Cassandra sink to watch the Cassandra source and write to another table while mapping columns appropriately. As the original table is phased out, the number of messages will decrease to none, while consumers are watching the new table’s CDC data topic. Refer to the "Multi-region CDC" question above for more detail.