Acero Overview#

This page gives an overview of the basic Acero concepts and helps distinguish Acero from other modules in the Arrow code base. It’s intended for users, developers, potential contributors, and for those that would like to extend Acero, either for research or for business use. This page assumes the reader is already familiar with core Arrow concepts. This page does not expect any existing knowledge in relational algebra.

What is Acero?#

Acero is a C++ library that can be used to analyze large (potentially infinite) streams of data. Acero allows computation to be expressed as an “execution plan” (ExecPlan). An execution plan takes in zero or more streams of input data and emits a single stream of output data. The plan describes how the data will be transformed as it passes through. For example, a plan might:

  • Merge two streams of data using a common column

  • Create additional columns by evaluating expressions against the existing columns

  • Consume a stream of data by writing it to disk in a partitioned layout

A sample execution plan that joins three streams of data and writes to disk

Acero is not…#

A Library for Data Scientists#

Acero is not intended to be used directly by data scientists. It is expected that end users will typically be using some kind of frontend. For example, Pandas, Ibis, or SQL. The API for Acero is focused around capabilities and available algorithms. However, such users may be intersted in knowing more about how Acero works so that they can better understand how the backend processing for their libraries operates.

A Database#

A database (or DBMS) is typically a much more expansive application and often packaged as a standalone service. Acero could be a component in a database (most databases have some kind of execution engine) or could be a component in some other data processing application that hardly resembles a database. Acero does not concern itself with user management, external communication, isolation, durability, or consistency. In addition, Acero is focused primarily on the read path, and the write utilities lack any sort of transaction support.

An Optimizer#

Acero does not have an SQL parser. It does not have a query planner. It does not have any sort of optimizer. Acero expects to be given very detailed and low-level instructions on how to manipulate data and then it will perform that manipulation exactly as described.

Creating the best execution plan is very hard. Small details can have a big impact on performance. We do think an optimizer is important but we believe it should be implemented independent of acero, hopefully in a composable way through standards such as Substrait so that any backend could leverage it.

Distributed#

Acero does not provide distributed execution. However, Acero aims to be usable by a distributed query execution engine. In other words, Acero will not configure and coordinate workers but it does expect to be used as a worker. Sometimes, the distinction is a bit fuzzy. For example, an Acero source may be a smart storage device that is capable of performing filtering or other advanced analytics. One might consider this a distributed plan. The key distinction is Acero does not have the capability of transforming a logical plan into a distributed execution plan. That step will need to be done elsewhere.

Acero vs…#

Arrow Compute#

This is described in more detail in Relation to Arrow C++ but the key difference is that Acero handles streams of data and Arrow Compute handles situations where all the data is in memory.

Arrow Datasets#

The Arrow datasets library provides some basic routines for discovering, scanning, and writing collections of files. The datasets module depends on Acero. Both scanning and writing datasets uses Acero. The scan node and the write node are part of the datasets module. This helps to keep the complexity of file formats and filesystems out of the core Acero logic.

Substrait#

Substrait is a project establishing standards for query plans. Acero executes query plans and generates data. This makes Acero a Substrait consumer. There are more details on the Substrait capabilities in Using Acero with Substrait.

Datafusion / DuckDb / Velox / Etc.#

There are many columnar data engines emerging. We view this as a good thing and encourage projects like Substrait to help allow switching between engines as needed. We generally discourage comparative benchmarks as they are almost inevitably going to be workload-driven and rarely manage to capture an apples-vs-apples comparison. Discussions of the pros and cons of each is beyond the scope of this guide.

Relation to Arrow C++#

The Acero module is part of the Arrow C++ implementation. It is built as a separate module but it depends on core Arrow modules and does not stand alone. Acero uses and extends the capabilities from the core Arrow module and the Arrow compute kernels.

A diagram of layers with core on the left, compute in the middle, and acero on the right

The core Arrow library provides containers for buffers and arrays that are laid out according to the Arrow columnar format. With few exceptions the core Arrow library does not examine or modify the contents of buffers. For example, converting a string array from lowercase strings to uppercase strings would not be a part of the core Arrow library because that would require examining the contents of the array.

The compute module expands on the core library and provides functions which analyze and transform data. The compute module’s capabilites are all exposed via a function registry. An Arrow “function” accepts zero or more arrays, batches, or tables, and produces an array, batch, or table. In addition, function calls can be combined, along with field references and literals, to form an expression (a tree of function calls) which the compute module can evaluate. For example, calculating x + (y * 3) given a table with columns x and y.

A sample expression tree

Acero expands on these capabilities by adding compute operations for streams of data. For example, a project node can apply a compute expression on a stream of batches. This will create a new stream of batches with the result of the expression added as a new column. These nodes can be combined into a graph to form a more complex execution plan. This is very similar to the way functions are combined into a tree to form a complex expression.

A simple plan that uses compute expressions

Note

Acero does not use the arrow::Table or arrow::ChunkedArray containers from the core Arrow library. This is because Acero operates on streams of batches and so there is no need for a multi-batch container of data. This helps to reduce the complexity of Acero and avoids tricky situations that can arise from tables whose columns have different chunk sizes. Acero will often use arrow::Datum which is a variant from the core module that can hold many different types. Within Acero, a datum will always hold either an arrow::Array or a arrow::Scalar.

Core Concepts#

ExecNode#

The most basic concept in Acero is the ExecNode. An ExecNode has zero or more inputs and zero or one outputs. If an ExecNode has zero inputs we call it a source and if an ExecNode does not have an output then we call it a sink. There are many different kinds of nodes and each one transforms its inputs in different ways. For example:

  • A scan node is a source node that reads data from files

  • An aggregate node accumulates batches of data to compute summary statistics

  • A filter node removes rows from the data according to a filter expression

  • A table sink node accumulates data into a table

Note

A full list of the available compute modules is included in the user’s guide

ExecBatch#

Batches of data are represented by the ExecBatch class. An ExecBatch is a 2D structure that is very similar to a RecordBatch. It can have zero or more columns and all of the columns must have the same length. There are a few key differences from ExecBatch:

../../_images/rb_vs_eb.svg

Both the record batch and the exec batch have strong ownership of the arrays & buffers#

  • An ExecBatch does not have a schema. This is because an ExecBatch is assumed to be part of a stream of batches and the stream is assumed to have a consistent schema. So the schema for an ExecBatch is typically stored in the ExecNode.

  • Columns in an ExecBatch are either an Array or a Scalar. When a column is a Scalar this means that the column has a single value for every row in the batch. An ExecBatch also has a length property which describes how many rows are in a batch. So another way to view a Scalar is a constant array with length elements.

  • An ExecBatch contains additional information used by the exec plan. For example, an index can be used to describe a batch’s position in an ordered stream. We expect that ExecBatch will also evolve to contain additional fields such as a selection vector.

../../_images/scalar_vs_array.svg

There are four different ways to represent the given batch of data using different combinations of arrays and scalars. All four exec batches should be considered semantically equivalent.#

Converting from a record batch to an exec batch is always zero copy. Both RecordBatch and ExecBatch refer to the exact same underlying arrays. Converting from an exec batch to a record batch is only zero copy if there are no scalars in the exec batch.

Note

Both Acero and the compute module have “lightweight” versions of batches and arrays. In the compute module these are called BatchSpan, ArraySpan, and BufferSpan. In Acero the concept is called KeyColumnArray. These types were developed concurrently and serve the same purpose. They aim to provide an array container that can be completely stack allocated (provided the data type is non-nested) in order to avoid heap allocation overhead. Ideally these two concepts will be merged someday.

ExecPlan#

An ExecPlan represents a graph of ExecNode objects. A valid ExecPlan must always have at least one source node but it does not technically need to have a sink node. The ExecPlan contains resources shared by all of the nodes and has utility functions to control starting and stopping execution of the nodes. Both ExecPlan and ExecNode are tied to the lifecycle of a single execution. They have state and are not expected to be restartable.

Warning

The structures within Acero, including ExecBatch, are still experimental. The ExecBatch class should not be used outside of Acero. Instead, an ExecBatch should be converted to a more standard structure such as a RecordBatch.

Similarly, an ExecPlan is an internal concept. Users creating plans should be using Declaration objects. APIs for consuming and executing plans should abstract away the details of the underlying plan and not expose the object itself.

Declaration#

A Declaration is a blueprint for an ExecNode. Declarations can be combined into a graph to form the blueprint for an ExecPlan. A Declaration describes the computation that needs to be done but is not actually responsible for carrying out the computation. In this way, a Declaration is analogous to an expression. It is expected that Declarations will need to be converted to and from various query representations (e.g. Substrait). The Declaration objects are the public API, combined with the DeclarationToXyz methods, are the current public API for Acero.

../../_images/decl_vs_ep.svg

A declaration is a blueprint that is used to instantiate exec plan instances#