Skip to main content
Version: latest


As Zeebe processes jobs and processes, or performs internal maintenance (e.g. raft failover), it generates an ordered stream of records.


Exporters are not available in Camunda Cloud Software-as-a-Service (SaaS).


While the clients provide no way to inspect this stream directly, Zeebe can load and configure user code that can process each and every one of those records in the form of an exporter.

An exporter provides a single entry point to process every record written on a stream.

  • Persist historical data by pushing it to an external data warehouse
  • Export records to a visualization tool (e.g. zeebe-simple-monitor)

Zeebe only loads exporters configured through the main Zeebe YAML configuration file.

Once an exporter is configured, the next time Zeebe starts, the exporter starts receiving records. Note that it is only guaranteed to see records produced from that point on.

Find a reference implementation in the form of the Zeebe-maintained Elasticsearch exporter.

The main impact exporters have on a Zeebe cluster is that they remove the burden of persisting data indefinitely.

Once data is not needed by Zeebe anymore, it queries its exporters to know if it can be safely deleted, and if so, permanently erases it, thereby reducing disk usage.


If no exporters are configured at all, Zeebe automatically erases data when it is not necessary anymore. If you need historical data, you must configure an exporter to stream records into your external data warehouse.

Regardless of how an exporter is loaded (whether through an external JAR or not), all exporters interact in the same way with the broker, which is defined by the Exporter interface.


Once configured, exporters are loaded as part of the broker startup phase, before any processing is done.

During the loading phase, the configuration for each exporter is validated, such that the broker will not start if:

  • An exporter ID is not unique
  • An exporter points to a non-existent/non-accessible JAR
  • An exporter points to a non-existent/non-instantiable class
  • An exporter instance throws an exception in its Exporter#configure method.

The last point is there to provide individual exporters to perform lightweight validation of their configuration (e.g. fail if missing arguments).

One caveat is that an instance of an exporter is created and immediately thrown away. Therefore, exporters should not perform any computationally heavy work during instantiation/configuration.

Note: Zeebe creates an isolated class loader for every JAR referenced by exporter configurations - that is, only once per JAR. If the same JAR is reused to define different exporters, these will share the same class loader.

This has some nice properties, primarily that different exporters can depend on the same third party libraries without having to worry about versions, or class name collisions.

Additionally, exporters use the system class loader for system classes, or classes packaged as part of the Zeebe JAR.

Exporter-specific configuration is handled through the exporter's [exporters.args] nested map. This provides a simple Map<String, Object> passed directly in the form of a Configuration object when the broker calls the Exporter#configure(Configuration) method.

Configuration occurs at two different phases: during the broker startup phase, and once every time a leader is elected for a partition.


At any given point, there is exactly one leader node for a given partition.

Whenever a node becomes the leader for a partition, it runs an instance of an exporter stream processor.

This stream processor creates exactly one instance of each configured exporter, and forwards every record written on the stream to each of these in turn.


This implies there will be exactly one instance of every exporter for every partition. If you have four partitions, and at least four threads for processing, there are potentially four instances of your exporter exporting simultaneously.

Zeebe only guarantees at-least-once semantics. That is, a record is seen at least once by an exporter, and maybe more. Cases where this may happen include:

  • During reprocessing after raft failover (i.e. new leader election)
  • On error if the position is not yet updated

To reduce the amount of duplicate records an exporter processes, the stream processor keeps track of the position of the last successfully exported record for every single exporter. The position is sufficient since a stream is an ordered sequence of records whose position is monotonically increasing. This position is set by the exporter once it can guarantee a record is successfully updated.

Although Zeebe tries to reduce the amount of duplicate records an exporter has to handle, it is likely it will have to. Therefore, it is necessary that export operations be idempotent. This can be implemented either in the exporter itself, but if it exports to an external system, it is recommended you perform deduplication there to reduce the load on Zeebe itself. Refer to the exporter-specific documentation for how this is meant to be achieved. :::

Error handling#

If an error occurs during the Exporter#open(Context) phase, the stream processor fails and is restarted, potentially fixing the error. Worst case scenario, this means no exporter runs until these errors stop.

If an error occurs during the Exporter#close phase, it is logged, but will still allow other exporters to gracefully finish their work.

If an error occurs during processing, we retry infinitely the same record until no error is produced. Worst case scenario, this means a failing exporter could bring all exporters to a halt. Currently, exporter implementations are expected to implement their own retry/error handling strategies, though this may change in the future.

Performance impact#

Zeebe naturally incurs a performance impact for each loaded exporter. A slow exporter slows down all other exporters for a given partition, and, in the worst case, could completely block a thread.

It's therefore recommended to keep exporters as simple as possible, and perform any data enrichment or transformation through the external system.