Changefeed Messages

On this page

Changefeeds emit messages as changes happen to watched tables. CockroachDB changefeeds have an at-least-once delivery guarantee as well as message ordering guarantees. You can also configure the format of changefeed messages with different options (e.g., format=avro).

This page describes the format and behavior of changefeed messages. You will find the following information on this page:

• Responses: The general format of changefeed messages.
• Message envelopes: The structure of the changefeed message.
• Ordering and delivery guarantees: CockroachDB's guarantees for a changefeed's message ordering.
• Delete messages: The format of messages when a row is deleted.
• Resolved messages: The resolved timestamp option and how to configure it.
• Duplicate messages: The causes of duplicate messages from a changefeed.
• Schema changes: The effect of schema changes on a changefeed.
• Garbage collection: How protected timestamps and garbage collection interacts with running changefeeds.
• Message formats: The limitations and type mapping when creating a changefeed with different message formats.

To filter the data that a changefeed emits in each message, refer to the Change Data Capture Queries page.

Responses

By default, changefeed messages emitted to a sink contain keys and values of the watched table rows that have changed. The message will contain the following fields depending on the type of emitted change and the options you specified to create the changefeed:

• Key: An array composed of the row's PRIMARY KEY field(s) (e.g., [1] for JSON or {"id":{"long":1}} for Avro).
• Value:
  • One of four possible top-level fields:
    • after, which contains the state of the row after the update (or null for DELETEs).
    • updated, which contains the updated timestamp.
    • resolved, which is emitted for records representing resolved timestamps. These records do not include an after value since they only function as checkpoints.
    • before, which contains the state of the row before an update. Changefeeds must use the diff option with the default wrapped envelope to emit the before field. When a row did not previously have any data, the before field will emit null.
  • For INSERT and UPDATE, the current state of the row inserted or updated.
  • For DELETE, null.

For example, changefeeds emitting to a sink will have the default message format:

When a changefeed targets a table with multiple column families, the family name is appended to the table name as part of the topic. Refer to Tables with columns families in changefeeds for guidance.

For webhook sinks, the response format arrives as a batch of changefeed messages with a payload and length.

{"payload": [{"after" : {"a" : 1, "b" : "a"}, "key": [1], "topic": "foo"}, {"after": {"a": 1, "b": "b"}, "key": [1], "topic": "foo" }], "length":2}

Webhook message batching is subject to the same key ordering guarantee as other sinks. Therefore, as messages are batched, you will not receive two batches at the same time with overlapping keys. You may receive a single batch containing multiple messages about one key, because ordering is maintained for a single key within its batch.

Refer to changefeed files for more detail on the file naming format for Enterprise changefeeds.

Message envelopes

The envelope defines the structure of a changefeed message. You can use the envelope option to manipulate the changefeed envelope. The values that the envelope option accepts are compatible with different changefeed sinks, and the structure of the message will vary depending on the sink.

The following sections provide examples of changefeed messages that are emitted when you specify each of the supported envelope options. Other changefeed options can affect the message envelope and what messages are emitted. Therefore, the examples are a guide for what you can expect when only the envelope option is specified.

wrapped

wrapped is the default envelope structure for changefeed messages. This envelope contains an array of the primary key (or the key as part of the message metadata), a top-level field for the type of message, and the current state of the row (or null for deleted rows).

The message envelope contains a primary key array when your changefeed is emitting to a sink that does not have a message key as part of its protocol, (e.g., cloud storage, webhook sinks, or Google Pub/Sub). By default, messages emitted to Kafka sinks do not have the primary key array, because the key is part of the message metadata. If you would like messages emitted to Kafka sinks to contain a primary key array, you can use the key_in_value option. Refer to the following message outputs for examples of this.

• Cloud storage sink:

  CREATE CHANGEFEED FOR TABLE vehicles INTO 'external://cloud';
  

  {"after": {"city": "seattle", "creation_time": "2019-01-02T03:04:05", "current_location": "86359 Jeffrey Ranch", "ext": {"color": "yellow"}, "id": "68ee1f95-3137-48e2-8ce3-34ac2d18c7c8", "owner_id": "570a3d70-a3d7-4c00-8000-000000000011", "status": "in_use", "type": "scooter"}, "key": ["seattle", "68ee1f95-3137-48e2-8ce3-34ac2d18c7c8"]}
  

• Kafka sink:

  • Default when envelope=wrapped or envelope is not specified:

    CREATE CHANGEFEED FOR TABLE vehicles INTO 'external://kafka';
    

    {"after": {"city": "washington dc", "creation_time": "2019-01-02T03:04:05", "current_location": "24315 Elizabeth Mountains", "ext": {"color": "yellow"}, "id": "dadc1c0b-30f0-4c8b-bd16-046c8612bbea", "owner_id": "034075b6-5380-4996-a267-5a129781f4d3", "status": "in_use", "type": "scooter"}}
    

  • Kafka sink message with key_in_value provided:

    CREATE CHANGEFEED FOR TABLE vehicles INTO 'external://kafka' WITH key_in_value, envelope=wrapped;
    

    {"after": {"city": "washington dc", "creation_time": "2019-01-02T03:04:05", "current_location": "46227 Jeremy Haven Suite 92", "ext": {"brand": "Schwinn", "color": "red"}, "id": "298cc7a0-de6b-4659-ae57-eaa2de9d99c3", "owner_id": "beda1202-63f7-41d2-aa35-ee3a835679d1", "status": "in_use", "type": "bike"}, "key": ["washington dc", "298cc7a0-de6b-4659-ae57-eaa2de9d99c3"]}
    

wrapped and diff

To include a before field in the changefeed message that contains the state of a row before an update in the changefeed message, use the diff option with wrapped:

CREATE CHANGEFEED FOR TABLE rides INTO 'external://kafka' WITH diff, envelope=wrapped;

{"after": {"city": "seattle", "end_address": null, "end_time": null, "id": "f6c02fe0-a4e0-476d-a3b7-91934d15dce2", "revenue": 25.00, "rider_id": "14067022-6e9b-427b-bd74-5ef48e93da1f", "start_address": "2 Michael Field", "start_time": "2023-06-02T15:14:20.790155", "vehicle_city": "seattle", "vehicle_id": "55555555-5555-4400-8000-000000000005"}, "before": {"city": "seattle", "end_address": null, "end_time": null, "id": "f6c02fe0-a4e0-476d-a3b7-91934d15dce2", "revenue": 25.00, "rider_id": "14067022-6e9b-427b-bd74-5ef48e93da1f", "start_address": "5 Michael Field", "start_time": "2023-06-02T15:14:20.790155", "vehicle_city": "seattle", "vehicle_id": "55555555-5555-4400-8000-000000000005"}, "key": ["seattle", "f6c02fe0-a4e0-476d-a3b7-91934d15dce2"]}

bare

bare removes the after key from the changefeed message and stores any metadata in a crdb field. When used with avro format, record will replace the after key.

• Cloud storage sink:

  CREATE CHANGEFEED FOR TABLE vehicles INTO 'external://cloud' WITH envelope=bare;
  

  {"__crdb__": {"key": ["washington dc", "cd48e501-e86d-4019-9923-2fc9a964b264"]}, "city": "washington dc", "creation_time": "2019-01-02T03:04:05", "current_location": "87247 Diane Park", "ext": {"brand": "Fuji", "color": "yellow"}, "id": "cd48e501-e86d-4019-9923-2fc9a964b264", "owner_id": "a616ce61-ade4-43d2-9aab-0e3b24a9aa9a", "status": "available", "type": "bike"}
  

CDC queries use envelope=bare message format by default. The bare message envelope places the output of the SELECT clause at the top level of the message instead of under an "after" key. When there is additional information that the changefeed is sending, such as updated or resolved timestamps, the messages will include a crdb field containing this information.

• In CDC queries:

  • A changefeed containing a SELECT clause without any additional options:

    CREATE CHANGEFEED INTO 'external://kafka' AS SELECT city, type FROM movr.vehicles;
    

    {"city": "los angeles", "type": "skateboard"}
    

  • A changefeed containing a SELECT clause with the topic_in_value option specified:

    CREATE CHANGEFEED INTO 'external://kafka' WITH topic_in_value AS SELECT city, type FROM movr.vehicles;
    

    {"__crdb__": {"topic": "vehicles"}, "city": "los angeles", "type": "skateboard"}
    

key_only

key_only emits only the key and no value, which is faster if you only need to know the key of the changed row. This envelope option is only supported for Kafka sinks or sinkless changefeeds.

• Kafka sink:

  CREATE CHANGEFEED FOR TABLE users INTO 'external://kafka' WITH envelope=key_only;
  

  ["boston", "22222222-2222-4200-8000-000000000002"]
  

  Note:

  It is necessary to set up a Kafka consumer to display the key because the key is part of the metadata in Kafka messages, rather than in its own field. When you start a Kafka consumer, you can use --property print.key=true to have the key print in the changefeed message.

• Sinkless changefeeds:

  CREATE CHANGEFEED FOR TABLE users WITH envelope=key_only;
  

  {"key":"[\"seattle\", \"fff726cc-13b3-475f-ad92-a21cafee5d3f\"]","table":"users","value":""}
  

row

row emits the row without any additional metadata fields in the message. This envelope option is only supported for Kafka sinks or sinkless changefeeds. row does not support avro format—if you are using avro, refer to the bare envelope option.

• Kafka sink:

  CREATE CHANGEFEED FOR TABLE vehicles INTO 'external://kafka' WITH envelope=row;
  

  {"city": "washington dc", "creation_time": "2019-01-02T03:04:05", "current_location": "85551 Moore Mountains Apt. 47", "ext": {"color": "red"}, "id": "d3b37607-1e9f-4e25-b772-efb9374b08e3", "owner_id": "4f26b516-f13f-4136-83e1-2ea1ae151c20", "status": "available", "type": "skateboard"}
  

Ordering and delivery guarantees

Changefeeds provide the following guarantees for message delivery to changefeed sinks:

• Per-key ordering
• At-least-once delivery

Per-key ordering

Changefeeds provide a per-key ordering guarantee for messages emitted to the sink. Once the changefeed has emitted a row with a timestamp, the changefeed will not emit any previously unseen versions of that row with a lower timestamp. Therefore, you will never receive a new change for that row at an earlier timestamp.

For example, a changefeed can emit updates to rows A at timestamp T1, B at T2, and C at T3 in any order.

When there are updates to rows A at T1, B at T2, and A at T3, the changefeed will always emit A at T3 after A at T1. However, A at T3 could precede or follow B at T2. This is because there is no timestamp ordering between keys.

As an example, you run the following sequence of SQL statements to create a changefeed:

1. Create a table:

   CREATE TABLE employees (
       id INT PRIMARY KEY,
       name STRING,
       office STRING
   );
   

2. Create a changefeed targeting the employees table:

   CREATE CHANGEFEED FOR TABLE employees INTO 'external://sink' WITH updated;
   

3. Insert and update values in employees:

   INSERT INTO employees VALUES (1, 'Terry', 'new york city');
   INSERT INTO employees VALUES (2, 'Alex', 'los angeles');
   UPDATE employees SET name = 'Terri' WHERE id = 1;
   INSERT INTO employees VALUES (3, 'Ash', 'london');
   UPDATE employees SET name = 'Terrence' WHERE id = 1;
   UPDATE employees SET office = 'new york city' WHERE id = 2;
   INSERT INTO employees VALUES (4, 'Danny', 'los angeles');
   INSERT INTO employees VALUES (5, 'Robbie', 'london');
   

   Note:

   In a transaction, if a row is modified more than once in the same transaction, the changefeed will only emit the last change.

4. The sink will receive messages of the inserted rows emitted per timestamp:

   {"after": {"id": 1, "name": "Terry", "office": "new york city"}, "key": [1], "updated": "1701102296662969433.0000000000"}
   {"after": {"id": 1, "name": "Terri", "office": "new york city"}, "key": [1], "updated": "1701102311425045162.0000000000"}
   {"after": {"id": 2, "name": "Alex", "office": "los angeles"}, "key": [2], "updated": "1701102305519323705.0000000000"}
   {"after": {"id": 3, "name": "Ash", "office": "london"}, "key": [3], "updated": "1701102316388801052.0000000000"}
   {"after": {"id": 1, "name": "Terrence", "office": "new york city"}, "key": [1], "updated": "1701102320607990564.0000000000"}
   {"after": {"id": 2, "name": "Alex", "office": "new york city"}, "key": [2], "updated": "1701102325724272373.0000000000"}
   {"after": {"id": 5, "name": "Robbie", "office": "london"}, "key": [5], "updated": "1701102330377135318.0000000000"}
   {"after": {"id": 4, "name": "Danny", "office": "los angeles"}, "key": [4], "updated": "1701102561022789676.0000000000"}
   

   The messages received at the sink are in order by timestamp for each key. Here, the update for key [1] is emitted before the insertion of key [2] even though the timestamp for the update to key [1] is higher. That is, if you follow the sequence of updates for a particular key at the sink, they will always be in the correct timestamp order.

   The updated option adds an updated timestamp to each emitted row. You can also use the resolved option to emit a resolved timestamp message to each Kafka partition, or to a separate file at a cloud storage sink. A resolved timestamp guarantees that no (previously unseen) rows with a lower update timestamp will be emitted on that partition.

   Note:

   Depending on the workload, you can use resolved timestamp notifications on every Kafka partition to provide strong ordering and global consistency guarantees by buffering records in between timestamp closures. Use the resolved timestamp to see every row that changed at a certain time.

Define a key column

Typically, changefeeds that emit to Kafka sinks shard rows between Kafka partitions using the row's primary key, which is hashed. The resulting hash remains the same and ensures a row will always emit to the same Kafka partition.

In some cases, you may want to specify another column in a table as the key by using the key_column option, which will determine the partition your messages will emit to. However, if you implement key_column with a changefeed, consider that other columns may have arbitrary values that change. As a result, the same row (i.e., by primary key) may emit to any partition at the sink based upon the column value. A changefeed with a key_column specified will still maintain per-key and at-least-once delivery guarantees.

To confirm that messages may emit the same row to different partitions when an arbitrary column is used, you must include the unordered option:

CREATE CHANGEFEED FOR TABLE employees INTO 'external://kafka-sink'
    WITH key_column='office', unordered;

At-least-once delivery

Changefeeds also provide an at-least-once delivery guarantee, which means that each version of a row will be emitted once. Under some infrequent conditions a changefeed will emit duplicate messages. This happens when the changefeed was not able to emit all messages before reaching a checkpoint. As a result, it will re-emit all messages starting from the previous checkpoint to ensure every message is delivered at least once, but could be emitted more than once.

Refer to Duplicate messages for causes of messages repeating at the sink.

For example, the checkpoints and changefeed pauses marked in this output show how messages may be duplicated, but always delivered:

{"after": {"id": 1, "name": "Terry", "office": "new york city"}, "key": [1], "updated": "1701102296662969433.0000000000"}
{"after": {"id": 1, "name": "Terri", "office": "new york city"}, "key": [1], "updated": "1701102311425045162.0000000000"}
{"after": {"id": 2, "name": "Alex", "office": "los angeles"}, "key": [2], "updated": "1701102305519323705.0000000000"}

[checkpoint]

{"after": {"id": 3, "name": "Ash", "office": "london"}, "key": [3], "updated": "1701102316388801052.0000000000"}
{"after": {"id": 1, "name": "Terrence", "office": "new york city"}, "key": [1], "updated": "1701102320607990564.0000000000"}

[changefeed pauses before the next checkpoint was reached]

[changefeed resumes and re-emits the messages after the previous checkpoint to ensure the sink received the messages]

{"after": {"id": 3, "name": "Ash", "office": "london"}, "key": [3], "updated": "1701102316388801052.0000000000"}
{"after": {"id": 1, "name": "Terrence", "office": "new york city"}, "key": [1], "updated": "1701102320607990564.0000000000"}
{"after": {"id": 2, "name": "Alex", "office": "new york city"}, "key": [2], "updated": "1701102325724272373.0000000000"}

[changefeed continues to emit new events]

{"after": {"id": 5, "name": "Robbie", "office": "london"}, "key": [5], "updated": "1701102330377135318.0000000000"}
{"after": {"id": 4, "name": "Danny", "office": "los angeles"}, "key": [4], "updated": "1701102561022789676.0000000000"}

[checkpoint]

With duplicates removed, an individual row is emitted in the same order as the transactions that updated it. However, this is not true for updates to two different rows, even two rows in the same table. (Refer to Per-key ordering.)

To compare two different rows for happens-before, compare the updated timestamp. This works across anything in the same cluster (tables, nodes, etc.).

The complexity with timestamps is necessary because CockroachDB supports transactions that can affect any part of the cluster, and it is not possible to horizontally divide the transaction log into independent changefeeds. For more information about this, read our blog post on CDC.

Delete messages

Deleting a row will result in a changefeed outputting the primary key of the deleted row and a null value. For example, with default options, deleting the row with primary key 5 will output:

[5] {"after": null}

In some unusual situations you may receive a delete message for a row without first seeing an insert message. For example, if an attempt is made to delete a row that does not exist, you may or may not get a delete message because the changefeed behavior is undefined to allow for optimizations at the storage layer. Similarly, if there are multiple writes to a row within a single transaction, only the last one will propagate to a changefeed. This means that creating and deleting a row within the same transaction will never result in an insert message, but may result in a delete message.

Resolved messages

When you create a changefeed with the resolved option, the changefeed will emit resolved timestamp messages in a format dependent on the connected sink. The resolved timestamp is the high-water mark that guarantees that no previously unseen rows with an earlier update timestamp will be emitted to the sink. That is, resolved timestamp messages do not emit until all ranges in the changefeed have progressed to a specific point in time.

When you specify the resolved option at changefeed creation, the job's coordinating node will send the resolved timestamp to each endpoint at the sink. For example, each Kafka partition will receive a resolved timestamp message, or a cloud storage sink will receive a resolved timestamp file.

There are three different ways to configure resolved timestamp messages:

• If you do not specify the resolved option at all, then the changefeed coordinator node will not send resolved timestamp messages.
• If you include WITH resolved in your changefeed creation statement without specifying a value, the coordinator node will emit resolved timestamps as the high-water mark advances. Note that new Kafka partitions may not receive resolved messages right away.
• If you specify a duration like WITH resolved={duration}, the changefeed will use it as the minimum duration between resolved messages that the changefeed coordinator sends. The changefeed will only emit a resolved timestamp message if the timestamp has advanced and at least the optional duration has elapsed.

Duplicate messages

Under some circumstances, changefeeds will emit duplicate messages to ensure the sink is receiving each message at least once. The following can cause or increase duplicate messages:

• The changefeed job encounters an error and pauses, or is manually paused.
• A node in the cluster restarts or fails.
• The changefeed job has the min_checkpoint_frequency option set, which can potentially increase duplicate messages.
• A target table undergoes a schema change. Schema changes may also cause the changefeed to emit the whole target table. Refer to Schema changes for detail on duplicates in this case.

A changefeed job cannot confirm that a message has been received by the sink unless the changefeed has reached a checkpoint. As a changefeed job runs, each node will send checkpoint progress to the job's coordinator node. These progress reports allow the coordinator to update the high-water mark timestamp confirming that all changes before (or at) the timestamp have been emitted.

When a changefeed must pause and then it resumes, it will return to the last checkpoint (A), which is the last point at which the coordinator confirmed all changes for the given timestamp. As a result, when the changefeed resumes, it will re-emit the messages that had sent after the last checkpoint, but were not confirmed in the next checkpoint.

Changefeed encounters an error

By default, changefeeds treat errors as retryable except for some specific terminal errors. When a changefeed encounters a retryable or non-retryable error, the job will pause until a successful retry or you resume the job once the error is solved. This can cause duplicate messages at the sink as the changefeed returns to the last checkpoint.

We recommend monitoring for changefeed retry errors and failures. Refer to the Monitor and Debug Changefeeds page.

Node restarts

When a node restarts, the changefeed will emit duplicates since the last checkpoint. During a rolling restart of nodes, a changefeed can fall behind as it tries to catch up during each node restart. For example, as part of a rolling upgrade or cluster maintenance, a node may drain every 5 minutes and the changefeed job checkpoints every 5 minutes.

To prevent the changefeed from falling too far behind, pause changefeed jobs before performing rolling node restarts.

min_checkpoint_frequency option

The min_checkpoint_frequency option controls how often nodes flush their progress to the coordinating changefeed node. Therefore, changefeeds will wait for at least the min_checkpoint_frequency duration before flushing to the sink. If a changefeed pauses and then resumes, the min_checkpoint_frequency duration is the amount of time that the changefeed will need to catch up since its previous checkpoint. During this catch-up time, you could receive duplicate messages.

Schema Changes

In v22.1, CockroachDB introduced the declarative schema changer. When schema changes happen that use the declarative schema changer by default, changefeeds will not emit duplicate records for the table that is being altered. It will only emit a copy of the table using the new schema. Refer to Schema changes with column backfill for examples of this.

Avro schema changes

To ensure that the Avro schemas that CockroachDB publishes will work with the schema compatibility rules used by the Confluent schema registry, CockroachDB emits all fields in Avro as nullable unions. This ensures that Avro and Confluent consider the schemas to be both backward- and forward-compatible, because the Confluent Schema Registry has a different set of rules than Avro for schemas to be backward- and forward-compatible.

The original CockroachDB column definition is also included within a doc field __crdb__ in the schema. This allows CockroachDB to distinguish between a NOT NULL CockroachDB column and a NULL CockroachDB column.

Schema changes with column backfill

When schema changes with column backfill (e.g., adding a column with a default, adding a stored computed column, adding a NOT NULL column, dropping a column) are made to watched rows, CockroachDB emits a copy of the table using the new schema.

The following example demonstrates the messages you will receive after creating a changefeed and then applying a schema change to the watched table:

CREATE TABLE office_dogs (
     id INT PRIMARY KEY,
     name STRING);

INSERT INTO office_dogs VALUES
   (1, 'Petee H'),
   (2, 'Carl'),
   (3, 'Ernie');

CREATE CHANGEFEED FOR TABLE office_dogs INTO 'external://cloud';

You receive each of the rows at the sink:

[1] {"id": 1, "name": "Petee H"}
[2] {"id": 2, "name": "Carl"}
[3] {"id": 3, "name": "Ernie"}

For example, add a column to the watched table:

ALTER TABLE office_dogs ADD COLUMN likes_treats BOOL DEFAULT TRUE;

After the schema change, the changefeed will emit a copy of the table with the new schema:

[1] {"id": 1, "name": "Petee H"}
[2] {"id": 2, "name": "Carl"}
[3] {"id": 3, "name": "Ernie"}
[1] {"id": 1, "likes_treats": true, "name": "Petee H"}
[2] {"id": 2, "likes_treats": true, "name": "Carl"}
[3] {"id": 3, "likes_treats": true, "name": "Ernie"}

If the schema change does not use the declarative schema changer by default, the changefeed will emit a copy of the altered table and a copy of the table using the new schema:

[1] {"id": 1, "name": "Petee H"}
[2] {"id": 2, "name": "Carl"}
[3] {"id": 3, "name": "Ernie"}
[1] {"id": 1, "name": "Petee H"}  # Duplicate
[2] {"id": 2, "name": "Carl"}     # Duplicate
[3] {"id": 3, "name": "Ernie"}    # Duplicate
[1] {"id": 1, "likes_treats": true, "name": "Petee H"}
[2] {"id": 2, "likes_treats": true, "name": "Carl"}
[3] {"id": 3, "likes_treats": true, "name": "Ernie"}

To prevent the changefeed from emitting a copy of the table with the new schema, use the schema_change_policy = nobackfill option. In the preceding two output blocks, the new schema messages that include the "likes_treats" column will not emit.

Refer to the CREATE CHANGEFEED option table for detail on the schema_change_policy option. You can also use the schema_change_events option to define the type of schema change event that triggers the behavior specified in schema_change_policy.

Garbage collection and changefeeds

By default, protected timestamps will protect changefeed data from garbage collection up to the time of the checkpoint.

Protected timestamps will protect changefeed data from garbage collection in the following scenarios:

• The downstream changefeed sink is unavailable. Protected timestamps will protect changes until you either cancel the changefeed or the sink becomes available once again.
• You pause a changefeed with the protect_data_from_gc_on_pause option enabled. Or, a changefeed with protect_data_from_gc_on_pause pauses from a retryable error. Protected timestamps will protect changes until you resume the changefeed.

However, if the changefeed lags too far behind, the protected changes could lead to an accumulation of garbage. This could result in increased disk usage and degraded performance for some workloads. To release the protected timestamps and allow garbage collection to resume, you can:

• Cancel the changefeed job.
• Resume a paused changefeed job.

• New in v23.1: Set the gc_protect_expires_after option, which will automatically expire the protected timestamp records that are older than your defined duration and cancel the changefeed job.

  For example, if the following changefeed is paused or runs into an error and then pauses, protected timestamps will protect changes for up to 24 hours. After this point, if the changefeed does not resume, the protected timestamp records will expire and the changefeed job will be cancelled. This releases the protected timestamp records and allows garbage collection to resume:

  CREATE CHANGEFEED FOR TABLE db.table INTO 'external://sink' WITH on_error='pause', protect_data_from_gc_on_pause, gc_protect_expires_after='24h';
  

We recommend monitoring storage and the number of running changefeeds. If a changefeed is not advancing and is retrying, it will (without limit) accumulate garbage while it retries to run.

When protect_data_from_gc_on_pause is unset, pausing the changefeed will release the existing protected timestamp record. As a result, you could lose the changes if the changefeed remains paused longer than the garbage collection window.

The only ways for changefeeds to not protect data are:

• You pause the changefeed without protect_data_from_gc_on_pause set.
• You cancel the changefeed.
• The changefeed fails without on_error=pause set.

Message formats

By default, changefeeds emit messages in JSON format. You can use a different format by creating a changefeed with the format option and specifying one of the following:

• json
• csv
• avro
• parquet (in Preview)

The following sections outline the limitations and type mapping for relevant formats.

Avro

The following sections provide information on Avro usage with CockroachDB changefeeds. Creating a changefeed using Avro is available in Core and Enterprise changefeeds with the confluent_schema_registry option.

Avro limitations

Below are clarifications for particular SQL types and values for Avro changefeeds:

• Decimals must have precision specified.
• BYTES (or its aliases BYTEA and BLOB) are often used to store machine-readable data. When you stream these types through a changefeed with format=avro, CockroachDB does not encode or change the data. However, Avro clients can often include escape sequences to present the data in a printable format, which can interfere with deserialization. A potential solution is to hex-encode BYTES values when initially inserting them into CockroachDB. This will ensure that Avro clients can consistently decode the hexadecimal. Note that hex-encoding values at insertion will increase record size.

• BIT and VARBIT types are encoded as arrays of 64-bit integers.

  For efficiency, CockroachDB encodes BIT and VARBIT bitfield types as arrays of 64-bit integers. That is, base-2 (binary format) BIT and VARBIT data types are converted to base 10 and stored in arrays. Encoding in CockroachDB is big-endian, therefore the last value may have many trailing zeroes. For this reason, the first value of each array is the number of bits that are used in the last value of the array.

  For instance, if the bitfield is 129 bits long, there will be 4 integers in the array. The first integer will be 1; representing the number of bits in the last value, the second integer will be the first 64 bits, the third integer will be bits 65–128, and the last integer will either be 0 or 9223372036854775808 (i.e., the integer with only the first bit set, or 1000000000000000000000000000000000000000000000000000000000000000 when base 2).

  This example is base-10 encoded into an array as follows:

  {"array": [1, <first 64 bits>, <second 64 bits>, 0 or 9223372036854775808]}
  

  For downstream processing, it is necessary to base-2 encode every element in the array (except for the first element). The first number in the array gives you the number of bits to take from the last base-2 number — that is, the most significant bits. So, in the example above this would be 1. Finally, all the base-2 numbers can be appended together, which will result in the original number of bits, 129.

  In a different example of this process where the bitfield is 136 bits long, the array would be similar to the following when base-10 encoded:

  {"array": [8, 18293058736425533439, 18446744073709551615, 13690942867206307840]}
  

  To then work with this data, you would convert each of the elements in the array to base-2 numbers, besides the first element. For the above array, this would convert to:

  [8, 1111110111011011111111111111111111111111111111111111111111111111, 1111111111111111111111111111111111111111111111111111111111111111, 1011111000000000000000000000000000000000000000000000000000000000]
  

  Next, you use the first element in the array to take the number of bits from the last base-2 element, 10111110. Finally, you append each of the base-2 numbers together — in the above array, the second, third, and truncated last element. This results in 136 bits, the original number of bits.

• A changefeed in Avro format will not be able to serialize user-defined composite (tuple) types. Tracking GitHub Issue

Avro types

Below is a mapping of CockroachDB types to Avro types:

CSV

You can use the format=csv option to emit CSV format messages from your changefeed. However, there are the following limitations with this option:

• It only works in combination with the initial_scan = 'only' option.
• It does not work when used with the diff or resolved options.
• A changefeed emitting CSV will include AS labels in the message format when the changefeed serializes a user-defined composite type. Tracking GitHub Issue

Changefeeds emit the same CSV format as EXPORT. In v22.1, changefeeds emitted CSV data that wrapped some values in single quotes, which were not wrapped when exporting data with the EXPORT statement.

See Export Data with Changefeeds for detail on using changefeeds to export data from CockroachDB.

The following shows example CSV format output:

4ccccccc-cccc-4c00-8000-00000000000f,washington dc,Holly Williams,95153 Harvey Street Suite 5,2165526885
51eb851e-b851-4c00-8000-000000000010,washington dc,Ryan Hickman,21187 Dennis Village,1635328127
56242e0e-4935-4d21-a8cd-915f4002e53c,washington dc,Joshua Smith,80842 Edwards Bridge,1892482054
5707febd-0278-4e55-8715-adbb35f09759,washington dc,Preston Fisher,5603 David Mission Apt. 93,5802323725
576546de-d59c-429b-9251-be79472643d4,washington dc,Anna Underwood,81246 Lee Knoll,2838348371
596c1cf8-d59f-4ad6-9379-6aba82648ca9,washington dc,Gerald Good,59876 Wang Neck,6779715200
5d30f838-e24c-46cb-bb0c-4a5643ddc2b1,washington dc,Lawrence Lucas,67248 Robinson Way Apt. 46,6167017463
65c398b9-7cce-45c5-9a5b-9561569ae030,washington dc,Mr. Xavier Waters,85393 Diaz Camp,1783482816
7a78fb0b-d368-46f6-b530-f9c74c19ba25,washington dc,Christopher Owens,7460 Curtis Centers,1470959770
80696ab6-7ec9-4e55-afee-4f468478fe82,washington dc,Patricia Gibson,77436 Vaughn Ville,3271633253
93750763-f992-4018-8a11-bf15ebfecc06,washington dc,Alison Romero,15878 Grant Forks Suite 16,2742488244
9cc3f995-0a91-4612-a079-e81ca28257ab,washington dc,Corey Dunn,15958 Jenna Locks,2358457606
9efd7047-c5e5-4501-9fcd-cff2d27efc34,washington dc,Patricia Gray,16139 Nicholas Wells Suite 64,8935020269
a253a15c-8e0a-4d25-aa87-1a0839935005,washington dc,Samantha Lee,90429 Russell Coves,2990967825
a3081762-9841-4275-ad7a-75a7e8d5f69d,washington dc,Preston Fisher,5603 David Mission Apt. 93,5802323725
aebb80a6-eceb-4d10-9d9a-f26270188114,washington dc,Kenneth Miller,52393 Stephen Mill Apt. 7,3966083325

See also

• Online Schema Changes
• Change Data Capture Overview
• Create and Configure Changefeeds

Was this helpful?