Streaming execution engine


The streaming execution engine is experimental, and a stable API is not yet guaranteed.


For many complex computations, successive direct invocation of compute functions is not feasible in either memory or computation time. Doing so causes all intermediate data to be fully materialized. To facilitate arbitrarily large inputs and more efficient resource usage, Arrow also provides a streaming query engine with which computations can be formulated and executed.

An example graph of a streaming execution workflow.

ExecNode is provided to reify the graph of operations in a query. Batches of data (ExecBatch) flow along edges of the graph from node to node. Structuring the API around streams of batches allows the working set for each node to be tuned for optimal performance independent of any other nodes in the graph. Each ExecNode processes batches as they are pushed to it along an edge of the graph by upstream nodes (its inputs), and pushes batches along an edge of the graph to downstream nodes (its outputs) as they are finalized.


(2018). Push versus pull-based loop fusion in query engines.
Journal of Functional Programming, 28.



Each node in the graph is an implementation of the ExecNode interface.


A set of ExecNode is contained and (to an extent) coordinated by an ExecPlan.


Instances of ExecNode are constructed by factory functions held in a ExecFactoryRegistry.


Heterogenous parameters for factories of ExecNode are bundled in an ExecNodeOptions.


dplyr-inspired helper for efficient construction of an ExecPlan.


A lightweight container for a single chunk of data in the Arrow format. In contrast to RecordBatch, ExecBatch is intended for use exclusively in a streaming execution context (for example, it doesn’t have a corresponding Python binding). Furthermore columns which happen to have a constant value may be represented by a Scalar instead of an Array. In addition, ExecBatch may carry execution-relevant properties including a guaranteed-true-filter for Expression simplification.

An example ExecNode implementation which simply passes all input batches through unchanged:

class PassthruNode : public ExecNode {
  // InputReceived is the main entry point for ExecNodes. It is invoked
  // by an input of this node to push a batch here for processing.
  void InputReceived(ExecNode* input, ExecBatch batch) override {
    // Since this is a passthru node we simply push the batch to our
    // only output here.
    outputs_[0]->InputReceived(this, batch);

  // ErrorReceived is called by an input of this node to report an error.
  // ExecNodes should always forward errors to their outputs unless they
  // are able to fully handle the error (this is rare).
  void ErrorReceived(ExecNode* input, Status error) override {
    outputs_[0]->ErrorReceived(this, error);

  // InputFinished is used to signal how many batches will ultimately arrive.
  // It may be called with any ordering relative to InputReceived/ErrorReceived.
  void InputFinished(ExecNode* input, int total_batches) override {
    outputs_[0]->InputFinished(this, total_batches);

  // ExecNodes may request that their inputs throttle production of batches
  // until they are ready for more, or stop production if no further batches
  // are required.  These signals should typically be forwarded to the inputs
  // of the ExecNode.
  void ResumeProducing(ExecNode* output) override { inputs_[0]->ResumeProducing(this); }
  void PauseProducing(ExecNode* output) override { inputs_[0]->PauseProducing(this); }
  void StopProducing(ExecNode* output) override { inputs_[0]->StopProducing(this); }

  // An ExecNode has a single output schema to which all its batches conform.
  using ExecNode::output_schema;

  // ExecNodes carry basic introspection for debugging purposes
  const char* kind_name() const override { return "PassthruNode"; }
  using ExecNode::label;
  using ExecNode::SetLabel;
  using ExecNode::ToString;

  // An ExecNode holds references to its inputs and outputs, so it is possible
  // to walk the graph of execution if necessary.
  using ExecNode::inputs;
  using ExecNode::outputs;

  // StartProducing() and StopProducing() are invoked by an ExecPlan to
  // coordinate the graph-wide execution state.  These do not need to be
  // forwarded to inputs or outputs.
  Status StartProducing() override { return Status::OK(); }
  void StopProducing() override {}
  Future<> finished() override { return inputs_[0]->finished(); }

Note that each method which is associated with an edge of the graph must be invoked with an ExecNode* to identify the node which invoked it. For example, in an ExecNode which implements JOIN this tagging might be used to differentiate between batches from the left or right inputs. InputReceived, ErrorReceived, InputFinished may only be invoked by the inputs of a node, while ResumeProducing, PauseProducing, StopProducing may only be invoked by outputs of a node.

ExecPlan contains the associated instances of ExecNode and is used to start and stop execution of all nodes and for querying/awaiting their completion:

// construct an ExecPlan first to hold your nodes
ARROW_ASSIGN_OR_RAISE(auto plan, ExecPlan::Make(default_exec_context()));

// ... add nodes to your ExecPlan

// start all nodes in the graph

SetUserCancellationCallback([plan] {
  // stop all nodes in the graph

// Complete will be marked finished when all nodes have run to completion
// or acknowledged a StopProducing() signal. The ExecPlan should be kept
// alive until this future is marked finished.
Future<> complete = plan->finished();

Constructing ExecPlan objects


The following will be superceded by construction from Compute IR, see ARROW-14074.

None of the concrete implementations of ExecNode are exposed in headers, so they can’t be constructed directly outside the translation unit where they are defined. Instead, factories to create them are provided in an extensible registry. This structure provides a number of benefits:

  • This enforces consistent construction.

  • It decouples implementations from consumers of the interface (for example: we have two classes for scalar and grouped aggregate, we can choose which to construct within the single factory by checking whether grouping keys are provided)

  • This expedites integration with out-of-library extensions. For example “scan” nodes are implemented in the separate library.

  • Since the class is not referencable outside the translation unit in which it is defined, compilers can optimize more aggressively.

Factories of ExecNode can be retrieved by name from the registry. The default registry is available through arrow::compute::default_exec_factory_registry() and can be queried for the built-in factories:

// get the factory for "filter" nodes:
ARROW_ASSIGN_OR_RAISE(auto make_filter,

// factories take three arguments:
ARROW_ASSIGN_OR_RAISE(ExecNode* filter_node, *make_filter(
    // the ExecPlan which should own this node

    // nodes which will send batches to this node (inputs)

    // parameters unique to "filter" nodes

// alternative shorthand:
ARROW_ASSIGN_OR_RAISE(filter_node, MakeExecNode("filter",
    plan.get(), {scan_node}, FilterNodeOptions{filter_expression});

Factories can also be added to the default registry as long as they are convertible to std::function<Result<ExecNode*>( ExecPlan*, std::vector<ExecNode*>, const ExecNodeOptions&)>.

To build an ExecPlan representing a simple pipeline which reads from a RecordBatchReader then filters, projects, and writes to disk:

std::shared_ptr<RecordBatchReader> reader = GetStreamOfBatches();
ExecNode* source_node = *MakeExecNode("source", plan.get(), {},

ExecNode* filter_node = *MakeExecNode("filter", plan.get(), {source_node},
                                        greater(field_ref("score"), literal(3))

ExecNode* project_node = *MakeExecNode("project", plan.get(), {filter_node},
                                         {add(field_ref("score"), literal(1))},
                                         {"score + 1"}

MakeExecNode("write", plan.get(), {project_node},
             WriteNodeOptions{/*base_dir=*/"/dat", /*...*/});

Declaration is a dplyr-inspired helper which further decreases the boilerplate associated with populating an ExecPlan from C++:


std::shared_ptr<RecordBatchReader> reader = GetStreamOfBatches();
                  {"source", SourceNodeOptions::FromReader(
                  {"filter", FilterNodeOptions{
                       greater(field_ref("score"), literal(3))}},
                  {"project", ProjectNodeOptions{
                       {add(field_ref("score"), literal(1))},
                       {"score + 1"}}},
                  {"write", WriteNodeOptions{/*base_dir=*/"/dat", /*...*/}},

Note that a source node can wrap anything which resembles a stream of batches. For example, PR#11032 adds support for use of a DuckDB query as a source node. Similarly, a sink node can wrap anything which absorbs a stream of batches. In the example above we’re writing completed batches to disk. However we can also collect these in memory into a Table or forward them to a RecordBatchReader as an out-of-graph stream. This flexibility allows an ExecPlan to be used as streaming middleware between any endpoints which support Arrow formatted batches.

An arrow::dataset::Dataset can also be wrapped as a source node which pushes all the dataset’s batches into an ExecPlan. This factory is added to the default registry with the name "scan" by calling arrow::dataset::internal::Initialize():


std::shared_ptr<Dataset> dataset = GetDataset();

                  {"scan", ScanNodeOptions{dataset,
                     /* push down predicate, projection, ... */}},
                  {"filter", FilterNodeOptions{/* ... */}},
                  // ...

Datasets may be scanned multiple times; just make multiple scan nodes from that dataset. (Useful for a self-join, for example.) Note that producing two scan nodes like this will perform all reads and decodes twice.