Passing data between pipeline steps

To pass data between the different pipeline steps, across different languages, we make use of the Apache Arrow project. The Orchest SDK provides a convenience wrapper of the project to be used within Orchest.

See the data passing quickstart of the SDK to get started directly.


The image below is a screenshot from the properties pane of step that has incoming steps “A”, “B” and “C”. The order of the list can be changed with a simple drag and drop.

From top to bottom: A -> C -> B

The order of this list is important as it determines the order in which the receiving step obtaines data from the steps A, B and C when calling orchest.transfer.get_inputs(). In the example image above the receiving step would get a list with the data from steps A, C and B respectively.


The Orchest SDK actually infers the order via the pipeline definition. The UI simply stores the order in the pipeline definition file.

Memory data passing

To pass data through memory between steps (which is enabled by default) we make use of the Plasma in-memory object store from the Apache Arrow project. Within Orchest it is wrapped with additional code for object eviction, which we will cover later in this section. Every interactive session gets its own memory store, which is shared between the kernels and interactive runs, for pipeline runs as part of experiments each gets an insolated memory store.

When an object is sent from one step to another (using orchest.transfer.output()) it is actually stored inside the Plasma store and copied into the memory of the receiving step. This is useful in interactive runs as it allows you to rerun a certain step without having to run the steps it depends on (if they have run before) enabling faster iteration on your ideas.

When it comes to clearing the memory store there are two options:

  1. Clearing all objects from memory through the pipeline settings.
  2. Enabling auto eviction also through the pipeline settings, additional information about this setting can be found in pipeline level configurations.