Debezium connector for MongoDB

Although the Debezium MongoDB connector does not become part of a replica set, it uses a similar replication mechanism to obtain oplog data. The main difference is that the connector does not read the oplog directly. Instead, it delegates the capture and decoding of oplog data to the MongoDB change streams feature. With change streams, the MongoDB server exposes the changes that occur in a collection as an event stream. The Debezium connector monitors the stream and then delivers the changes downstream. The first time that the connector detects a replica set, it examines the oplog to obtain the last recorded transaction, and then performs a snapshot of the primary’s databases and collections. After the connector finishes copying the data, it creates a change stream beginning from the oplog position that it read earlier.

As the MongoDB connector processes changes, it periodically records the position at which the event originated in the oplog stream. When the connector stops, it records the last oplog stream position that it processed, so that after a restart it can resume streaming from that position. In other words, the connector can be stopped, upgraded or maintained, and restarted some time later, and always pick up exactly where it left off without losing a single event. Of course, MongoDB oplogs are usually capped at a maximum size, so if the connector is stopped for long periods, operations in the oplog might be purged before the connector has a chance to read them. In this case, after a restart the connector detects the missing oplog operations, performs a snapshot, and then proceeds to stream changes.

The MongoDB connector is also quite tolerant of changes in membership and leadership of the replica sets, of additions or removals of shards within a sharded cluster, and network problems that might cause communication failures. The connector always uses the replica set’s primary node to stream changes, so when the replica set undergoes an election and a different node becomes primary, the connector will immediately stop streaming changes, connect to the new primary, and start streaming changes using the new primary node. Similarly, if connector is unable to communicate with the replica set primary, it attempts to reconnect (using exponential backoff so as to not overwhelm the network or replica set). After connection is reestablished, the connector continues to stream changes from the last event that it captured. In this way the connector dynamically adjusts to changes in replica set membership, and automatically handles communication disruptions.

You can specify MongoDB read preferences for a connection in the connector properties. The method that you use to set read preferences depends on the MongoDB topology, and the mongodb.connection.mode.

The MongoDB connector supports the following MongoDB topologies:

To capture data from MongoDB, Debezium attaches to the database as a MongoDB user. The MongoDB user account that you create for Debezium requires specific database permissions to read from the database. The connector user requires the following permissions:

The connector user might also require the following permission:

The connector user must be able to read from all databases, or to read from a specific database, depending on the value of the connector’s capture.scope property. Assign one of the following permissions to the user, depending on the capture.scope setting:

Regardless of the capture.scope setting, the user requires permission to run the MongoDB ping command.

For connectors that capture changes from a sharded MongoDB cluster, and for which the mongodb.connection.mode property is set to replica_set, you must grant the user permission to read the config.shards system collection.

The connector configuration property topic.prefix serves as a logical name for the MongoDB replica set or sharded cluster. The connector uses the logical name in a number of ways: as the prefix for all topic names, and as a unique identifier when recording the change stream position of each replica set.

You should give each MongoDB connector a unique logical name that meaningfully describes the source MongoDB system. We recommend logical names begin with an alphabetic or underscore character, and remaining characters that are alphanumeric or underscore.

When a Debezium task starts to use a replica set, it uses the connector’s logical name and the replica set name to find an offset that describes the position where the connector previously stopped reading changes. If an offset can be found and it still exists in the oplog, then the task immediately proceeds with streaming changes, starting at the recorded offset position.

However, if no offset is found, or if the oplog no longer contains that position, the task must first obtain the current state of the replica set contents by performing a snapshot. This process starts by recording the current position of the oplog and recording that as the offset (along with a flag that denotes a snapshot has been started). The task then proceeds to copy each collection, spawning as many threads as possible (up to the value of the snapshot.max.threads configuration property) to perform this work in parallel. The connector records a separate read event for each document it sees. Each read event contains the object’s identifier, the complete state of the object, and source information about the MongoDB replica set where the object was found. The source information also includes a flag that denotes that the event was produced during a snapshot.

This snapshot will continue until it has copied all collections that match the connector’s filters. If the connector is stopped before the tasks' snapshots are completed, upon restart the connector begins the snapshot again.

By default, a connector runs an initial snapshot operation only after it starts for the first time. Following this initial snapshot, under normal circumstances, the connector does not repeat the snapshot process. Any future change event data that the connector captures comes in through the streaming process only.

However, in some situations the data that the connector obtained during the initial snapshot might become stale, lost, or incomplete. To provide a mechanism for recapturing collection data, Debezium includes an option to perform ad hoc snapshots. You might want to perform an ad hoc snapshot after any of the following changes occur in your Debezium environment:

You can re-run a snapshot for a collection for which you previously captured a snapshot by initiating a so-called ad-hoc snapshot. Ad hoc snapshots require the use of signaling collections. You initiate an ad hoc snapshot by sending a signal request to the Debezium signaling collection.

When you initiate an ad hoc snapshot of an existing collection, the connector appends content to the topic that already exists for the collection. If a previously existing topic was removed, Debezium can create a topic automatically if automatic topic creation is enabled.

Ad hoc snapshot signals specify the collections to include in the snapshot. The snapshot can capture the entire contents of the database, or capture only a subset of the collections in the database.

You specify the collections to capture by sending an execute-snapshot message to the signaling collection. Set the type of the execute-snapshot signal to incremental or blocking, and provide the names of the collections to include in the snapshot, as described in the following table:

You initiate an ad hoc incremental snapshot by adding an entry with the execute-snapshot signal type to the signaling collection. After the connector processes the message, it begins the snapshot operation. The snapshot process reads the first and last primary key values and uses those values as the start and end point for each collection. Based on the number of entries in the collection, and the configured chunk size, Debezium divides the collection into chunks, and proceeds to snapshot each chunk, in succession, one at a time.

For more information, see Incremental snapshots.

You initiate an ad hoc blocking snapshot by adding an entry with the execute-snapshot signal type to the signaling collection. After the connector processes the message, it begins the snapshot operation. The connector temporarily stops streaming, and then initiates a snapshot of the specified collection, following the same process that it uses during an initial snapshot. After the snapshot completes, the connector resumes streaming.

For more information, see Blocking snapshots.

To provide flexibility in managing snapshots, Debezium includes a supplementary snapshot mechanism, known as incremental snapshotting. Incremental snapshots rely on the Debezium mechanism for sending signals to a Debezium connector. Incremental snapshots are based on the DDD-3 design document.

In an incremental snapshot, instead of capturing the full state of a database all at once, as in an initial snapshot, Debezium captures each collection in phases, in a series of configurable chunks. You can specify the collections that you want the snapshot to capture and the size of each chunk. The chunk size determines the number of rows that the snapshot collects during each fetch operation on the database. The default chunk size for incremental snapshots is 1024 rows.

As an incremental snapshot proceeds, Debezium uses watermarks to track its progress, maintaining a record of each collection row that it captures. This phased approach to capturing data provides the following advantages over the standard initial snapshot process:

When you run an incremental snapshot, Debezium sorts each collection by primary key and then splits the collection into chunks based on the configured chunk size. Working chunk by chunk, it then captures each collection row in a chunk. For each row that it captures, the snapshot emits a READ event. That event represents the value of the row when the snapshot for the chunk began.

As a snapshot proceeds, it’s likely that other processes continue to access the database, potentially modifying collection records. To reflect such changes, INSERT, UPDATE, or DELETE operations are committed to the transaction log as per usual. Similarly, the ongoing Debezium streaming process continues to detect these change events and emits corresponding change event records to Kafka.

In some cases, the UPDATE or DELETE events that the streaming process emits are received out of sequence. That is, the streaming process might emit an event that modifies a collection row before the snapshot captures the chunk that contains the READ event for that row. When the snapshot eventually emits the corresponding READ event for the row, its value is already superseded. To ensure that incremental snapshot events that arrive out of sequence are processed in the correct logical order, Debezium employs a buffering scheme for resolving collisions. Only after collisions between the snapshot events and the streamed events are resolved does Debezium emit an event record to Kafka.

To assist in resolving collisions between late-arriving READ events and streamed events that modify the same collection row, Debezium employs a so-called snapshot window. The snapshot windows demarcates the interval during which an incremental snapshot captures data for a specified collection chunk. Before the snapshot window for a chunk opens, Debezium follows its usual behavior and emits events from the transaction log directly downstream to the target Kafka topic. But from the moment that the snapshot for a particular chunk opens, until it closes, Debezium performs a de-duplication step to resolve collisions between events that have the same primary key..

For each data collection, the Debezium emits two types of events, and stores the records for them both in a single destination Kafka topic. The snapshot records that it captures directly from a table are emitted as READ operations. Meanwhile, as users continue to update records in the data collection, and the transaction log is updated to reflect each commit, Debezium emits UPDATE or DELETE operations for each change.

As the snapshot window opens, and Debezium begins processing a snapshot chunk, it delivers snapshot records to a memory buffer. During the snapshot windows, the primary keys of the READ events in the buffer are compared to the primary keys of the incoming streamed events. If no match is found, the streamed event record is sent directly to Kafka. If Debezium detects a match, it discards the buffered READ event, and writes the streamed record to the destination topic, because the streamed event logically supersede the static snapshot event. After the snapshot window for the chunk closes, the buffer contains only READ events for which no related transaction log events exist. Debezium emits these remaining READ events to the collection’s Kafka topic.

The connector repeats the process for each snapshot chunk.

To use incremental snapshots with sharded MongoDB clusters, you must set specific values for the following properties:

After the connector task for a replica set records an offset, it uses the offset to determine the position in the oplog where it should start streaming changes. The task then (depending on the configuration) either connects to the replica set’s primary node or connects to a replica-set-wide change stream and starts streaming changes from that position. It processes all of create, insert, and delete operations, and converts them into Debezium change events. Each change event includes the position in the oplog where the operation was found, and the connector periodically records this as its most recent offset. The interval at which the offset is recorded is governed by offset.flush.interval.ms, which is a Kafka Connect worker configuration property.

When the connector is stopped gracefully, the last offset processed is recorded so that, upon restart, the connector will continue exactly where it left off. If the connector’s tasks terminate unexpectedly, however, then the tasks may have processed and generated events after it last records the offset but before the last offset is recorded; upon restart, the connector begins at the last recorded offset, possibly generating some the same events that were previously generated just prior to the crash.

As mentioned earlier, the connector tasks always use the replica set’s primary node to stream changes from the oplog, ensuring that the connector sees the most up-to-date operations as possible and can capture the changes with lower latency than if secondaries were to be used instead. When the replica set elects a new primary, the connector immediately stops streaming changes, connects to the new primary, and starts streaming changes from the new primary node at the same position. Likewise, if the connector experiences any problems communicating with the replica set members, it tries to reconnect, by using exponential backoff so as to not overwhelm the replica set, and once connected it continues streaming changes from where it last left off. In this way, the connector is able to dynamically adjust to changes in replica set membership and automatically handle communication failures.

To summarize, the MongoDB connector continues running in most situations. Communication problems might cause the connector to wait until the problems are resolved.

In MongoDB 6.0 and later, you can configure change streams to emit the pre-image state of a document to populate the before field for MongoDB change events. To enable the use of pre-images in MongoDB, you must set the changeStreamPreAndPostImages for a collection by using db.createCollection(), create, or collMod. To enable the Debezium MongoDB to include pre-images in change events, set the capture.mode for the connector to one of the *_with_pre_image options.

The MongoDB connector writes events for all insert, update, and delete operations to documents in each collection to a single Kafka topic. The name of the Kafka topics always takes the form logicalName.databaseName.collectionName, where logicalName is the logical name of the connector as specified with the topic.prefix configuration property, databaseName is the name of the database where the operation occurred, and collectionName is the name of the MongoDB collection in which the affected document existed.

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

Notice that the topic names do not incorporate the replica set name or shard name. As a result, all changes to a sharded collection (where each shard contains a subset of the collection’s documents) all go to the same Kafka topic.

You can set up Kafka to auto-create the topics as they are needed. If not, then you must use Kafka administration tools to create the topics before starting the connector.

The MongoDB connector does not make any explicit determination about how to partition topics for events. Instead, it allows Kafka to determine how to partition topics based on event keys. You can change Kafka’s partitioning logic by defining the name of the Partitioner implementation in the Kafka Connect worker configuration.

Kafka maintains total order only for events written to a single topic partition. Partitioning the events by key does mean that all events with the same key always go to the same partition. This ensures that all events for a specific document are always totally ordered.

Debezium can generate events that represents transaction metadata boundaries and enrich change data event messages.

For every transaction BEGIN and END, Debezium generates an event that contains the following fields:

The following example shows a typical message:

Unless overridden via the topic.transaction option, transaction events are written to the topic named <topic.prefix>.transaction.

When transaction metadata is enabled, the data message Envelope is enriched with a new transaction field. This field provides information about every event in the form of a composite of fields:

Following is an example of what a message looks like:

A change event’s key contains the schema for the changed document’s key and the changed document’s actual key. For a given collection, both the schema and its corresponding payload contain a single id field. The value of this field is the document’s identifier represented as a string that is derived from MongoDB extended JSON serialization strict mode.

Consider a connector with a logical name of fulfillment, a replica set containing an inventory database, and a customers collection that contains documents such as the following.

Every change event that captures a change to the customers collection has the same event key schema. For as long as the customers collection has the previous definition, every change event that captures a change to the customers collection has the following key structure. In JSON, it looks like this:

This example uses a document with an integer identifier, but any valid MongoDB document identifier works the same way, including a document identifier. For a document identifier, an event key’s payload.id value is a string that represents the updated document’s original _id field as a MongoDB extended JSON serialization that uses strict mode. The following table provides examples of how different types of _id fields are represented.

The value in a change event is a bit more complicated than the key. Like the key, the value has a schema section and a payload section. The schema section contains the schema that describes the Envelope structure of the payload section, including its nested fields. Change events for operations that create, update or delete data all have a value payload with an envelope structure.

Consider the same sample document that was used to show an example of a change event key:

The value portion of a change event for a change to this document is described for each event type:

The following example shows the value portion of a change event that the connector generates for an operation that creates data in the customers collection:

The value in a delete change event has the same schema portion as create and update events for the same collection. The payload portion in a delete event contains values that are different from create and update events for the same collection. In particular, a delete event contains neither an after value nor a updateDescription value. Here is an example of a delete event for a document in the customers collection:

MongoDB connector events are designed to work with Kafka log compaction. Log compaction enables removal of some older messages as long as at least the most recent message for every key is kept. This lets Kafka reclaim storage space while ensuring that the topic contains a complete data set and can be used for reloading key-based state.

All MongoDB connector events for a uniquely identified document have exactly the same key. When a document is deleted, the delete event value still works with log compaction because Kafka can remove all earlier messages that have that same key. However, for Kafka to remove all messages that have that key, the message value must be null. To make this possible, after Debezium’s MongoDB connector emits a delete event, the connector emits a special tombstone event that has the same key but a null value. A tombstone event informs Kafka that all messages with that same key can be removed.

You can use the Debezium connector for MongoDB with MongoDB Atlas. Note that MongoDB Atlas only supports secure connections via SSL, i.e. the +mongodb.ssl.enabled connector option must be set to true.

The Debezium MongoDB connector reads change streams to obtain oplog data for a replica set. Because the oplog is a fixed-sized, capped collection, if it exceeds its maximum configured size, it begins to overwrite its oldest entries. If the connector is stopped for any reason, when it restarts, it attempts to resume streaming from the last oplog stream position. However, if last stream position was removed from the oplog, depending on the value specified in the connector’s snapshot.mode property, the connector might fail to start, reporting an invalid resume token error. In the event of a failure, you must create a new connector to enable Debezium to continue capturing records from the database. For more information, see Connector fails after it is stopped for a long interval if snapshot.mode is set to initial.

To ensure that the oplog retains the offset values that Debezium requires to resume streaming, you can use either of the following approaches:

To help prevent failures that are related to missing oplog entries, it’s important to track metrics that report replication behavior, and to optimize the oplog size to support Debezium. In particular, you should monitor the values of Oplog GB/Hour and Replication Oplog Window. If Debezium is offline for an interval that exceeds the value of the replication oplog window, and the primary oplog grows faster than Debezium can consume entries, a connector failure can result.

For information about how to monitor these metrics, see the MongoDB documentation.

It’s best to set the maximum oplog size to a value that is based on the anticipated hourly growth of the oplog (Oplog GB/Hour), multiplied by the time that might be required to address a Debezium failure.

That is,

Oplog GB/Hour X average reaction time to Debezium failure

For example, if the oplog size limit is set to 1GB, and the oplog grows by 3GB per hour, oplog entries are cleared three times per hour. If Debezium were to fail during this time, its last oplog position is likely to be removed.

If the oplog grows at the rate of 3GB/hour, and Debezium is offline for two hours, you would thus set the oplog size to 3GB/hour X 2 hours, or 6GB.

Following is an example of the configuration for a connector instance that captures data from a MongoDB replica set rs0 at port 27017 on 192.168.99.100, which we logically name fullfillment. Typically, you configure the Debezium MongoDB connector in a JSON file by setting the configuration properties that are available for the connector.

You can choose to produce events for a particular MongoDB replica set or sharded cluster. Optionally, you can filter out collections that are not needed.

For the complete list of the configuration properties that you can set for the Debezium MongoDB connector, see MongoDB connector configuration properties.

You can send this configuration with a POST command to a running Kafka Connect service. The service records the configuration and starts one connector task that performs the following actions:

To start running a Debezium MongoDB connector, create a connector configuration, and add the configuration to your Kafka Connect cluster.

After the connector starts, it completes the following actions:

The Debezium MongoDB connector has numerous configuration properties that you can use to achieve the right connector behavior for your application. Many properties have default values. Information about the properties is organized as follows:

The following configuration properties are required unless a default value is available.

The following advanced configuration properties have good defaults that will work in most situations and therefore rarely need to be specified in the connector’s configuration.

The MBean is debezium.mongodb:type=connector-metrics,context=snapshot,server=<topic.prefix>,task=<task.id>.

Snapshot metrics are not exposed unless a snapshot operation is active, or if a snapshot has occurred since the last connector start.

The following table lists the shapshot metrics that are available.

The Debezium MongoDB connector also provides the following custom snapshot metrics:

The MBean is debezium.mongodb:type=connector-metrics,context=streaming,server=<topic.prefix>,task=<task.id>.

The following table lists the streaming metrics that are available.

The Debezium MongoDB connector also provides the following custom streaming metrics:

In the following situations, the connector fails when trying to start, reports an error or exception in the log, and stops running:

After a failure, the connector attempts to reconnect by using exponential backoff. You can configure the maximum number of reconnection attempts.

In these cases, the error will have more details about the problem and possibly a suggested work around. The connector can be restarted when the configuration has been corrected or the MongoDB problem has been addressed.

Once the connector is running, if the primary node of any of the MongoDB replica sets become unavailable or unreachable, the connector will repeatedly attempt to reconnect to the primary node, using exponential backoff to prevent saturating the network or servers. If the primary remains unavailable after the configurable number of connection attempts, the connector will fail.

The attempts to reconnect are controlled by three properties:

Each delay is double that of the prior delay, up to the maximum delay. Given the default values, the following table shows the delay for each failed connection attempt and the total accumulated time before failure.

A connector that is stopped for a long period fails to start, and reports the following exception:

The preceding exception indicates that the entry that corresponds to the connector’s resume token is no longer present in the oplog. Because the oplog no longer contains the last offset that the connector processed, the connector cannot resume streaming.

You can use either of the following options to recover from the failure:

To help prevent failures related to missing resume tokens, optimize configuration of the oplog.

If Kafka Connect is being run in distributed mode, and a Kafka Connect process is stopped gracefully, then prior to shutdown of that processes Kafka Connect will migrate all of the process' connector tasks to another Kafka Connect process in that group, and the new connector tasks will pick up exactly where the prior tasks left off. There is a short delay in processing while the connector tasks are stopped gracefully and restarted on the new processes.

If the group contains only one process and that process is stopped gracefully, then Kafka Connect will stop the connector and record the last offset for each replica set. Upon restart, the replica set tasks will continue exactly where they left off.

If the Kafka Connector process stops unexpectedly, then any connector tasks it was running will terminate without recording their most recently-processed offsets. When Kafka Connect is being run in distributed mode, it will restart those connector tasks on other processes. However, the MongoDB connectors will resume from the last offset recorded by the earlier processes, which means that the new replacement tasks may generate some of the same change events that were processed just prior to the crash. The number of duplicate events depends on the offset flush period and the volume of data changes just before the crash.

As the connector generates change events, the Kafka Connect framework records those events in Kafka using the Kafka producer API. Kafka Connect will also periodically record the latest offset that appears in those change events, at a frequency that you have specified in the Kafka Connect worker configuration. If the Kafka brokers become unavailable, the Kafka Connect worker process running the connectors will simply repeatedly attempt to reconnect to the Kafka brokers. In other words, the connector tasks will simply pause until a connection can be reestablished, at which point the connectors will resume exactly where they left off.

If the connector is gracefully stopped, users might continue to perform operations on replica set members. Changes that occur while the connector is offline continue to be recorded in MongoDB’s oplog. In most cases, after the connector is restarted, it reads the offset value in the oplog to determine the last operation that it streamed for each replica set, and then resumes streaming changes from that point. After the restart, database operations that occurred while the connector was stopped are emitted to Kafka as usual, and after some time, the connector catches up with the database. The amount of time required for the connector to catch up depends on the capabilities and performance of Kafka and the volume of changes that occurred in the database.

However, if the connector remains stopped for a long enough interval, it can occur that MongoDB purges the oplog during the time that the connector is inactive, resulting in the loss of information about the connector’s last position. After the connector restarts, it cannot resume streaming, because the oplog no longer contains the previous offset value that marks the last operation that the connector processed. The connector also cannot perform a snapshot, as it typically would when the snapshot.mode property is set to initial, and no offset value is present. In this case, a mismatch exists, because the oplog does not contain the value of the previous offset, but the offset value is present in the connector’s internal Kafka offsets topic. An error results and the connector fails.

To recover from the failure, delete the failed connector, and create a new connector with the same configuration but with a different connector name. When you start the new connector, it performs a snapshot to ingest the state of database, and then resumes streaming.

In certain failure situations, MongoDB can lose commits, which results in the MongoDB connector being unable to capture the lost changes. For example, if the primary crashes suddenly after it applies a change and records the change to its oplog, the oplog might become unavailable before secondary nodes can read its contents. As a result, the secondary node that is elected as the new primary node might be missing the most recent changes from its oplog.

At this time, there is no way to prevent this side effect in MongoDB.