Memgraph can connect to existing Kafka, Redpanda, and Pulsar sources to ingest the data, which you can then query with the power of MAGE algorithms or your own custom procedures.
To use streams, a user must:
- Create a transformation module
- Load the transformation module into Memgraph
- Create the stream with a query and optionally set its offset with
- Start the stream with a query
You can write Python transformation modules, create and start streams using the Stream section in the Memgraph Lab, check out how.
Create a stream
The syntax for creating a stream depends on the type of the source because each specific type supports a different set of configuration options.
There is no strict order for specifying the configuration options.
Kafka and Redpanda
To check the list of possible configuration options and their values, please check the documentation of librdkafka library, which is used in Memgraph. At the time of writing this documentation Memgraph uses version 1.7.0 of librdkafka.
Pulsar
The transformation procedure is called if either the or the is reached, and at least one message is received.
The starts when the:
- the stream is started
- the processing of the previous batch is completed
- the previous batch interval ended without receiving any messages
After each message is processed, the stream will acknowledge them. If the stream is stopped, the next time it starts, it will continue processing the message from the last acknowledged message.
The user who executes the query is the owner of the stream.
Memgraph Community doesn’t support authentication and authorization, so the owner is always , and the privileges are not checked.
In Memgraph Enterprise, owner privileges are checked upon executing the queries returned from the transformation procedures. If the owner doesn’t have the required privileges, the execution of the queries will fail. Find more information about how the owner affects the stream in the reference guide.
Triggers created by users logged in via SSO cannot be used beyond the session they were created in.
Start a stream
The following query will start a specific stream with name to consume number of batches for a maximum duration of milliseconds.
The stream will automatically stop after consuming the given number of batches or reaching the timeout. If number of batches are not processed within the specified , probably because not enough messages was received, an exception is thrown. is measured in milliseconds, and its default value is 30000. It can only be used in combination with the option.
If (and ) is not provided, the stream will run for an infinite number of batches without a timeout limit.
The following query will start all streams for an infinite number of batches and without a timeout limit.
When a stream is started, it resumes ingesting data from the last committed offset. If no offset is committed for the consumer group, the largest offset will be used. Therefore, only the new messages will be consumed.
Stop a stream
The following queries stop a specific stream or all streams.
Delete a stream
The following query drops a stream with the name .
Show streams
To show streams, use the following query:
It shows a list of existing streams with the following information:
- stream name
- stream type
- batch interval
- batch size
- transformation procedure name
- the owner of the streams
- whether the stream is running or not
Check stream
To perform a dry-run on the stream and get the results of the transformation, use the following query:
The clause will do a dry-run on the stream with number of batches and return the result of the transformation, that is, the queries and parameters that would be executed in a normal run. If number of batches are not processed within the specified , probably because not enough messages were received, an exception is thrown.
The default value of is 1. is measured in milliseconds, and its default value is 30000.
Get stream information
To get more information about a specific Kafka or Redpanda stream, use the following query:
This procedure will return information about the bootstrap server, set configuration, consumer group, credentials, and topics.
To get more information about a specific Pulsar stream, use the following query:
The procedure will return the service URL and topics.
Kafka producer delivery semantics
In stream processing, it is important to consider how failures are handled. When connecting an external application such as Memgraph to a Kafka stream, there are two possible ways to handle failures during message processing:
- Every message is processed at least once: the message offsets are committed to the Kafka cluster after processing. If the committing fails, the messages can get processed multiple times.
- Every message is processed at most once: the message offsets are committed to the Kafka cluster right after they are received before the processing is started. If the processing fails, the same messages won’t be processed again.
Missing a message can result in missing an edge that would connect two independent components of a graph. Therefore, the general opinion in Memgraph is that missing some information is a bigger problem in graphs databases than having duplicated information, so Memgraph uses at least once semantics, i.e., the queries returned by the transformations are first executed and committed to the database for every batch of messages, and only then is the message offset committed to the Kafka cluster.
However, even though Memgraph cannot guarantee exactly once semantics, it tries to minimize the possibility of processing messages multiple times. This means committing the message offsets to the Kafka cluster happens right after the transaction is committed to the database.
Configuring stream transactions
A stream can fail for various reasons. One important type of failure is when a transaction (in which the returned queries of the transformation are executed) fails to commit because of another conflicting transaction. This is a side effect of isolation levels and can be remedied by the following Memgraph flag:
By default, Memgraph will always try to execute a transaction once. However, for streams, if Memgraph fails because of transaction conflicts, it will retry to execute the transaction again for up to times (default value is 30).
Moreover, the interval of retries is also important and can be configured with the following Memgraph flag:
The is measured in and the default value is .
Setting a stream offset
When using a Kafka stream, you can manually set the offset of the next consumed message with a call to the query procedure :
- An offset of denotes the start of the stream, i.e., the beginning offset available for the given topic/partition.
- An offset of denotes the end of the stream, i.e., for each topic/partition, its logical end such that only the next produced message will be consumed.
Stream can consume messages from multiple topics with multiple partitions. Therefore, when setting the offsets to an arbitrary number be aware that setting the offset of a stream internally sets all of the associated offsets of that stream (topics/partitions) to that value.