Debezium Connector for Cassandra

    Since Cassandra 3.0, a is introduced. The CDC feature can be enabled on the table level by setting the table property , after which any commit log containing data for a CDC-enabled table will be moved to the CDC directory specified in cassandra.yaml on discard.

    The Cassandra connector resides on each Cassandra node and monitors the cdc_raw directory for change. It processes all local commit log segments as they are detected, produces a change event for every row-level insert, update, and delete operations in the commit log, publishes all change events for each table in a separate Kafka topic, and finally deletes the commit log from the cdc_raw directory. This last step is important because once CDC is enabled, Cassandra itself cannot purge the commit logs. If the cdc_free_space_in_mb fills up, writes to CDC-enabled tables will be rejected.

    The connector is tolerant of failures. As the connector reads commit logs and produces events, it records each commit log segment’s filename and position along with each event. If the connector stops for any reason (including communication failures, network problems, or crashes), upon restart it simply continues reading the commit log where it last left off. This includes snapshots: if the snapshot was not completed when the connector is stopped, upon restart it will begin a new snapshot. We’ll talk later about how the connector behaves when things go wrong.

    The following features are currently not supported by the Cassandra connector. Changes resulted from any of these features are ignored:

    • TTL on collection-type columns

    • Range deletes

    • Static columns

    • Triggers

    • Materialized views

    • Secondary indices

    • Light-weight transactions

    Before the Debezium Cassandra connector can be used to monitor the changes in a Cassandra cluster, CDC must be enabled on the node level and table level.

    To enable CDC, update the following CDC config in cassandra.yaml:

    Additional CDC configs are have the following default values:

    1. cdc_raw_directory: $CASSANDRA_HOME/data/cdc_raw
    2. cdc_free_space_in_mb: 4096
    3. cdc_free_space_check_interval_ms: 250

    • cdc_enabled enables or disables CDC operations node-wide

    • cdc_raw_directory determines the destination for commit log segments to be moved after all corresponding memtables are flushed

    • cdc_free_space_in_mb is the maximum capacity allocated to store commit log segments, and defaults to the minimum of 4096 MB and 1/8 of volume space.

    • cdc_free_space_check_interval_ms is frequency with which we re-calculate the space taken up by cdc_raw_directory to prevent burning CPU cycles unnecessarily when at capacity.

    Enabling CDC on Table

    Once CDC is enabled on the Cassandra node, each table must be be explicitly enabled for CDC as well via the CREATE TABLE or ALTER TABLE command. For example:

    1. CREATE TABLE foo (a int, b text, PRIMARY KEY(a)) WITH cdc=true;
    3. ALTER TABLE foo WITH cdc=true;

    This section goes into detail about how the Cassandra connector performs snapshots, transforms commit log events into Debezium change events, handles commit log life cycle, records events into Kafka, manages schema evolution, and behaves when things go wrong.

    Snapshots

    When the Cassandra connector first starts on a Cassandra node, it will by default perform an initial snapshot of the cluster. This is the default mode, since most of the time CDC is enabled on non-empty tables and commit logs do not contain the complete history.

    The snapshot reader issues a SELECT statement to query all the columns in a table. Cassandra allows consistency level to be set either globally or on the statement level. For snapshotting, the consistency level is set on the statement level to ALL by default to provide the highest consistency. This implies if one node goes down during the snapshot, the snapshot would not be able to continue and a subsequent re-snapshot is required once the node has been brought back online. You can adjust the consistency level of the snapshot to a lower consistency level in order to increase availability, provided that you understand the tradeoff with consistency.

    In Cassandra 3.X, it is not possible to read strictly from the local Cassandra node. Starting in Cassandra 4.0, a NODE_LOCAL consistency level will be added. This will allow the Cassandra connector to read from the node it resides in only (which would be consistent with the way commit logs are processed).

    Unlike relational databases, there is no read lock applied during a snapshot, so writes to Cassandra are not blocked during that snapshot. If the queried data has been modified by another client during the snapshot, those changes may be reflected in the snapshot result set.

    If the connector fails or stops before the snapshot is completed, the connector will begin a new snapshot upon restarts. In the default snapshot mode (initial), once the connector completes its initial snapshot, it will no longer perform any additional snapshots. The only exception would be during a connector restart: if cdc is enabled on a table, and then the connector is restarted, that table would be snapshotted.

    The second snapshot mode (always) allows the connector to perform snapshot whenever necessary. it will check periodically for newly cdc-enabled tables, and snapshot them as soon as they are detected.

    The third snapshot mode (‘never’) ensures the connector never performs snapshots. When a new connector is configured this way, it will only read the commit log in the CDC directory. This is not the default behavior because starting a new connector in this mode (without a snapshot) requires the commit logs to contain the entire history of all cdc-enabled tables, which is often not the case. Another use case for this mode is if there is one connector already doing the snapshotting, you can disable snapshot on others to avoid duplicated work.

    Reading the Commit Log

    The Cassandra connector will typically spend the vast majority of its time reading local commit logs on the Cassandra node.

    Commit logs’ binary data are deserialized with Cassandra’s CommitLogReader and CommitLogReadHandler. Each deserialized object is called a mutation in Cassandra. A mutation contains one or more change events.

    As the Cassandra connector reads the commit log, it transform the log events into Debezium create, update, or delete events that include the position in the commit log where the event was found. The Cassandra connector encode these change events with Kafka Connect converters and publish them to the appropriate Kafka topics.

    Cassandra’s commit logs come with a set of limitations, which are critical for interpreting CDC events correctly:

    • Commit logs only arrive in cdc_raw directory when it is full, in which case it would be flushed/discarded. This implies there is a delay between when the event is logged and when the event is captured.

    • Commit logs on an individual Cassandra node do not reflect all writes to the cluster, they only reflect writes stored on that node. This is why it is necesssary to monitor changes on all nodes in a Cassandra cluster. However, due to replication factor, this also implies it is necessary for downstream consumers of these events to handle deduplication.

    • Writes to an individual Cassandra node are logged as they arrive. However, these events may arrive out-of-order from which they are issued. Downstream consumers of these events must understand and implement logic similar to Cassandra’s read path to get the correct output.

    • Schema changes of tables are not recorded in commit logs, only data changes are recorded. Therefore changes in schema are detected on a best-effort basis. To avoid data loss, it is recommended to pause writes to the table during schema change.

    • Cassandra does not perform read-before-write, as a result commit logs do not record the value of every column in the changed row, it only records the values of columns that have been modified (except for partition key columns, which are always recorded as they are required in Cassandra DML commands).

    • Due to the nature of CQL, insert DMLs can result in a row insertion or update; update DMLs can result in a row insertion, update, or deletion; delete DMLs can result in a row update or deletion. Since queries are not recorded in commit logs, CDC event type is classified based on the effect on the row in a relational database sense.

    TODO: is there a way to determine event type which corresponds to the actual Cassandra DML statement? and if so, is that preferred over the semantic of these events?

    Managing Commit Log Lifecycle

    By default, Cassandra connector will delete commit logs which have been processed. It is not recommended to start the connector while deletion of commit logs is disabled, as this could bloat up disk storage and prevent further writes to the Cassandra cluster. To manage the commit logs in a custom manner (i.e. upload it to a cloud provider), the CommitLogTransfer interface can be implemented.

    Topics names

    The Cassandra connector writes events for all insert, update, and delete uperations on a single table to a single Kafka topic. The name of the Kafka topics always take the form clusterName.keyspaceName.tableName, where clusterName is the logical name of the connector as specified with the kafka.topic.prefix configuraiton property, keyspaceName is the name of the keyspace where the operation occurred, and tableName is the name of the table on which the operation occurred.

    For example, consider a Cassandra installation with an inventory keyspace that contains four tables: products, products_on_hand, customers, and orders. If the connector monitoring this database were given a logical server name of fulfillment, then the connector would produce events on these four Kafka topics:

    • fulfillment.inventory.products

    • fulfillment.inventory.products_on_hand

    • fulfillment.inventory.customers

    • fulfillment.inventory.orders

    TODO: for topic name, is clusterName.keyspaceName.tableName okay? or should it be connectorName.keyspaceName.tableName or connectorName.clusterName.keyspaceName.tableName?

    Schema Evolution

    DDLs are not recorded in commit logs. When the schema of a table change, this change is issued from one of the Cassandra node and propagated to other nodes via Gossip Protocol.

    Schema changes in Cassandra will be detected by an implemented SchemaChangeListener with latency less than 1s, which will then update the schema instance loaded from Cassandra as well as the Kafka key value schemas cached for each table.

    Please note that with the current schema evolution approach, the Cassandra connector won’t be able to provide accurate data change information for a small period of time in the following cases:

    1. If CDC gets disabled for a table, data changes which have happened before CDC got disabled will be skipped.

    2. If a column is removed from a table, data changes involving this column before it’s removed cannot be deserialized correctly and will be skipped.

    All data change events produced by the Cassandra connector have a key and a value, although the structure of the key and value depends on the table from which the change events originated (see Topic Names).

    Change Event’s Key

    For a given table, the change event’s key will have a structure that contains a field for each column in the primary key of the table at the time the event was created. Consider an inventory database with a customers table defined as:

    1. CREATE TABLE customers (
    2.   id bigint,
    3.   registration_date timestamp,
    4.   first_name text,
    5.   last_name text,
    6.   email text,
    7.   PRIMARY KEY (id, registration_date)
    8. );

    Every change event for the customers table while it has this definition will feature the same key schema, which in JSON representation looks like this:

    1. {
    2.   "type": "record",
    3.   "name": "cassandra-cluster-1.inventory.customers.Key",
    4.   "namespace": "io.debezium.connector.cassandra",
    5.   "fields": [
    6.     {
    7.       "name": "id",
    8.       "type": "long"
    9.     },
    10.     {
    11.       "name": "registration_date",
    12.       "type": "long",
    13.       "logicalType": "timestamp-millis"
    14.   ]
    15. }

    For id = 1001 and registration_date = 1562202942545, the key payload in JSON representation would look like this:

    1. {
    2.   "id": 1001,
    3.   "registration_date": 1562202942545
    4. }

    Although the column.exclude.list configuration property allows you to remove columns from the event values, all columns in a primary key are always included in the event’s key.

    Change event’s value

    The value of the change event message is a bit more complicated. Every change event value produced by Cassandra connector has an envelope structure with the following fields:

    • op is a mandatory field that contains a string value describing the type of operation. Values for the Cassandra connector are i for insert, u for update, and d for delete.

    • after is an optional field that if present contains the state of the row after the event occurred. The structure will be described by the cassandra-cluster-1.inventory.customers.Value Kafka Connect schema, which represent the cluster, keyspace, and table the event is referring to.

    • source is a mandatory field that contains a structure describing the source metadata for the event, which in the case of Cassandra contains several fields: the Debezium version, the connector name, the Cassandra cluster name, the name of the commit log file where the event was recorded, the position in that commit log file where the event appeared, whether this event was part of a snapshot, name of the affected keyspace and table, and the maximum timestamp of the partition update in microseconds.

    • ts_ms is optional and if present contains the time (using the system clock in the JVM running the Cassandra connector) at which the connector processed the event.

    Note that there is no before field. This is because Cassandra does not perform a read-before-write, therefore the commit log does not contain row values before the change is applied.

    The following is a JSON representation of a value schema for a create event for our customers table:

    TODO: verify max timestamp != deletion timestamp in case of deletion DDLs

    Given the following insert DML:

    1. INSERT INTO customers (
    2.   id,
    3.   registration_date,
    4.   first_name,
    5.   last_name,
    6.   email)
    7. VALUES (
    8.   1001,
    9.   now(),
    10.   "Anne",
    11.   "Kretchmar",
    12.   "annek@noanswer.org"
    13. );

    The value payload in JSON representation would look like this:

    1. {
    2.   "op": "c",
    3.   "ts_ms": 1562202942832,
    4.   "after": {
    5.     "id": {
    6.     "value": 1001,
    7.     "deletion_ts": null,
    8.     "set": true
    9.   },
    10.     "registration_date": {
    11.     "value": 1562202942545,
    12.     "deletion_ts": null,
    13.     "set": true
    14.   },
    15.   "first_name": {
    16.     "value": "Anne",
    17.     "deletion_ts": null,
    18.     "set": true
    19.   },
    20.   "last_name": {
    21.     "value": "Kretchmar",
    22.     "deletion_ts": null,
    23.     "set": true
    24.   },
    25.   "email": {
    26.     "value": "annek@noanswer.org",
    28.     "set": true
    29.   }
    30.   },
    31.   "source": {
    32.     "version": "1.9.6.Final",
    33.     "connector": "cassandra",
    34.     "cluster": "cassandra-cluster-1",
    35.     "snapshot": false,
    36.     "keyspace": "inventory",
    37.     "table": "customers",
    38.     "file": "commitlog-6-123456.log",
    39.     "pos": 54,
    40.     "ts_ms": 1562202942666382
    41.   }
    42. }

    Given the following update DML:

    1. UPDATE customers
    2. SET email = "annek_new@noanswer.org"
    3. WHERE id = 1001 AND registration_date = 1562202942545

    The value payload in JSON representation would look like this:

    1. {
    2.   "op": "u",
    3.   "ts_ms": 1562202942912,
    4.   "after": {
    5.     "id": {
    6.     "value": 1001,
    7.     "deletion_ts": null,
    8.     "set": true
    9.   },
    10.     "registration_date": {
    11.     "value": 1562202942545,
    12.     "set": true
    13.   },
    14.   "first_name": null,
    15.   "last_name": null,
    16.   "email": {
    17.     "value": "annek_new@noanswer.org",
    18.     "deletion_ts": null,
    19.     "set": true
    20.   }
    21.   },
    22.   "source": {
    23.     "version": "1.9.6.Final",
    24.     "connector": "cassandra",
    25.     "cluster": "cassandra-cluster-1",
    26.     "snapshot": false,
    27.     "keyspace": "inventory",
    28.     "table": "customers",
    29.     "file": "commitlog-6-123456.log",
    30.     "pos": 102,
    31.     "ts_ms": 1562202942666490
    32.   }
    33. }

    When we compare this to the value in the insert event, we see a couple differences:

    • The op field value is now u, signifying that this row changed because of an update.

    • The after field now has the updated state of the row, and here we can see that the email value is now annek_new@noanswer.org. Notice that first_name and last_name are null, this is because these fields did not change during this update. However, id and registration_date are still included, because these are the primary keys of this table.

    • The source field structure has the same fields as before, but the values are different since this event is from a different position in the commit log.

    • The ts_ms shows the timestamp milliseconds which the connector processed this event.

    Finally, given the following delete DML:

    1. DELETE FROM customers
    2. WHERE id = 1001 AND registration_date = 1562202942545;

    The value payload in JSON representation would look like this:

    When we compare this to the value in the insert and update event, we see a couple differences:

    • The op field value is now d, signifying that this row changed because of a deletion.

    • The after field only contains values for id and registration_date because this is a deletion by primary keys.

    • The source field structure has the same fields as before, but the values are different since this event is from a different position in the commit log.

    • The ts_ms shows the timestamp milliseconds which the connector processed this event.

    TODO: given TTL is not currently support, would it be better to remove delete_ts? would it also be okay to derive whether a field is set or not by looking at the each column to see if it is null?

    TODO: discuss tombstone events in Cassandra connector

    Data Types

    As described above, the Cassandra connector represents the changes to rows with events that are structured like the table in which the row exist. The event contains a field for each column value, and how that value is represented in the event depends on the Cassandra data type of the column. This section describes this mapping.

    The following table describes how the connector maps each of the Cassandra data types to an Kafka Connect data type.

    TODO: add logical types

    If the default data type conversions do not meet your needs, you can create a custom converter for the connector.

    When Things Go Wrong

    Configuration And Startup Errors

    The Cassandra connector will fail upon startup, report error or exception in the log, and stop running if the configurations are invalid or if the connector cannot successfully connector to Cassandra using the specified connectivity parameters. In this case, the error will have more details about the problem and possibly suggest a work around. The connector can be restarted when the configuration has been corrected.

    Cassandra Becomes Unavailable

    Once the connector is running, if the Cassandra node becomes unavailable for any reason, the connector will fail and stop. In this case, restart the connector once the server becomes available. If this happened during snapshot, it will rebootstrap the entire table from the beginning of the table.

    Cassandra Connector Stops Gracefully

    If the Cassandra connector is gracefully shut down, prior to stopping the process it will make sure to flush all events in the ChangeEventQueue to Kafka. The Cassandra connector keeps track of the filename and offset each time a streamed record is send to Kafka. So when the connector is restarted, it will resume from where it left off. It does this by searching for the oldest commit log in the directory, start processing that commitlog, skipping the already-read records, until it finds the most recent record that hasn’t been processed. If the Cassandra connector is stopped during snapshot, it will pick up from that table, but will rebootstrap the entire table.

    Cassandra Connector Crashes

    If the Cassandra connector crashes unexpected, then the Cassandra connector would likely have terminated without recording the most-recently processed offset. In this case, when the connector is restarted, it will resume from the most recent recorded offset. This means duplicates is likely (which is trivial since we already be get duplicates from RF). Note that since the offset is only updated when a record has been successfully send to Kafka, it is okay to lose the un-emitted data in the ChangeEventQueue during a crash, as these events will be recreated.

    Kafka Becomes Unavailable

    As the connector generate change event, it will publish those events to Kafka using Kafka producer API. If Kafka broker becomes unavailable (producer encounters TimeoutException), the Cassandra connector will repeatedly attempt to reconnect to the broker once per second until a successful retry.

    Cassandra connector is Stopped for a Duration

    Depending on the write load of a table, when a Cassandra connector is stopped for a long time, it risks into hitting the cdc_total_space_in_mb capacity. Once this upper limit is reached, Cassandra will stop accepting writes for this table; which means it is important to monitor this space while running the Cassandra connector. In the worst case scenario when this happens, the best thing to do is to (1) turn off Cassandra connector (2) disable cdc for the table so it stops generating additional writes (although writes to other cdc-enabled tables on the same node could still affect the commitlog file generation given the commit logs are not filtered) (3) remove the recorded offset from the offset file (4) once the capacity is increased or the directory used space is under control, restart the connector so it will rebootstrap the table.

    Cassandra Table CDC is Enabled, Then Temporarily Disabled, And Then Enabled Again

    If a Cassandra table temporarily disables CDC and then re-enables it again after some time, it must be re-bootstrapped. To re-bootstrap an individual table, you can manually remove the recorded offset line corresponding to that table from snapshot_offset.properties file.

    The Cassandra connector should be deployed each Cassandra node in a Cassandra cluster. The Cassandra connector Jar file takes in a cdc configuration (.properties) file. See see example configurations for reference.

    Example configuration

    The following represents an example .properties configuration file for running and testing the Cassandra Connector locally:

    1. connector.name=test_connector
    2. commit.log.relocation.dir=/Users/test_user/debezium-connector-cassandra/test_dir/relocation/
    3. http.port=8000
    5. cassandra.config=/usr/local/etc/cassandra/cassandra.yaml
    6. cassandra.hosts=127.0.0.1
    7. cassandra.port=9042
    9. kafka.producer.bootstrap.servers=127.0.0.1:9092
    10. kafka.producer.retries=3
    11. kafka.producer.retry.backoff.ms=1000
    12. kafka.topic.prefix=test_prefix
    14. key.converter=io.confluent.connect.avro.AvroConverter
    15. key.converter.schema.registry.url: http://localhost:8081
    16. value.converter=io.confluent.connect.avro.AvroConverter
    17. value.converter.schema.registry.url: http://localhost:8081
    19. offset.backing.store.dir=/Users/test_user/debezium-connector-cassandra/test_dir/
    21. snapshot.consistency=ONE
    22. snapshot.mode=ALWAYS
    24. latest.commit.log.only=true

    Cassandra connector has built-in support for JMX metrics. The Cassandra driver also publishes a number of metrics about the driver’s activities that can be monitored through JMX. The connector has two types of metrics. Snapshot metrics help you monitor the snapshot activity and are available when the connector is performing a snapshot. Binlog metrics help you monitor the progress and activity while the connector reads the Cassandra commit logs.

    Snapshot Metrics

    Attribute Name

    Type

    Description

    total-table-count

    int

    The total number of tables that are being included in the snapshot.

    remaining-table-count

    int

    The number of tables that the snapshot has yet to copy.

    snapshot-running

    boolean

    Whether the snapshot was started.

    snapshot-aborted

    boolean

    Whether the snapshot was aborted.

    snapshot-completed

    boolean

    Whether the snapshot completed.

    snapshot-during-in-seconds

    long

    The total number of seconds that the snapshot has taken so far, even if not complete.

    rows-scanned

    Map<String, Long>

    Map containing the number of rows scanned for each table in the snapshot. Tables are incrementally added to the Map during processing. Updates every 10,000 rows scanned and upon completing a table.

    Commitlog Metrics

    Attribute Name

    Type

    Description

    commitlog-filename

    string

    The name of the commit log filename that the connector has most recently read.

    commitlog-position

    long

    number-of-processed-mutations

    long

    The number of mutations that have been processed.

    number-of-unrecoverable-errors

    long

    The number of unrecoverable errors while processing commit logs.

    Connector properties

    If the Cassandra agent use SSL to connect to Cassandra node, an SSL config file is required. The following example shows how to write the SSL config file:

    1. keyStore.location=/var/private/ssl/cassandra.keystore.jks
    2. keyStore.password=cassandra
    3. keyStore.type=JKS
    4. trustStore.location=/var/private/ssl/cassandra.truststore.jks
    5. trustStore.password=cassandra
    6. trustStore.type=JKS
    7. keyManager.algorithm=SunX509
    8. trustManager.algorithm=SunX509
    9. cipherSuites=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384

    The cipherSuites field is not mandatory, it simply allows you to add one (or more) ciphers that are not present. The default value of trustStore.type and keyStore.type is JKS. The default value of keyManager.algorithm and trustManager.algorithm is SunX509.

    The connector also supports pass-through configuration properties that are used when creating the Kafka producer. Specifically, all connector configuration properties that begin with the kafka.producer. prefix are used (without the prefix) when creating the Kafka producer that writes events to Kafka.

    For example, the follwoing connector configuration properties can be used to secure connections to the Kafka broker:

    1. kafka.producer.security.protocol=SSL
    2. kafka.producer.ssl.keystore.location=/var/private/ssl/kafka.server.keystore.jks
    3. kafka.producer.ssl.keystore.password=test1234
    4. kafka.producer.ssl.truststore.location=/var/private/ssl/kafka.server.truststore.jks
    5. kafka.producer.ssl.truststore.password=test1234
    6. kafka.producer.ssl.key.password=test1234
    7. kafka.consumer.security.protocol=SSL
    8. kafka.consumer.ssl.keystore.location=/var/private/ssl/kafka.server.keystore.jks
    9. kafka.consumer.ssl.keystore.password=test1234
    10. kafka.consumer.ssl.truststore.location=/var/private/ssl/kafka.server.truststore.jks
    11. kafka.consumer.ssl.truststore.password=test1234

    Be sure to consult the for all of the configuration properties for Kafka producers.