Comparison to Spark¶
Apache Spark is a popular distributed computing tool for tabular datasets that is growing to become a dominant name in Big Data analysis today. Dask has several elements that appear to intersect this space and we are often asked, “How does Dask compare with Spark?”
Answering such comparison questions in an unbiased and informed way is hard, particularly when the differences can be somewhat technical. This document tries to do this; we welcome any corrections.
Generally Dask is smaller and lighter weight than Spark. This means that it has fewer features and, instead, is used in conjunction with other libraries, particularly those in the numeric Python ecosystem. It couples with libraries like Pandas or Scikit-Learn to achieve high-level functionality.
|Written in Python and only really supports Python. It interoperates well with C/C++/ Fortran/LLVM or other natively compiled code linked through Python.||Written in Scala with some support for Python and R. It interoperates well with other JVM code.|
A component of the larger Python ecosystem.
It couples with and enhances other libraries like NumPy, Pandas, and Scikit-Learn
An all-in-one project that has inspired its own ecosystem.
It integrates well with many other Apache projects.
Age and trust¶
|Dask is younger (since 2014) and is an extension of the well trusted NumPy/Pandas /Scikit-learn/Jupyter stack.||Spark is older (since 2010) and has become a dominant and well-trusted tool in the Big Data enterprise world.|
|Applied more generally both to business intelligence applications, as well as a number of scientific and custom situations.||More focused on traditional business intelligence operations like SQL and lightweight machine learning.|
Lower level, and so lacks high level optimizations, but is able to implement more sophisticated algorithms and build more complex bespoke systems.
It is fundamentally based on generic task scheduling.
Higher level, providing good high level optimizations on uniformly applied computations, but lacking flexibility for more complex algorithms or ad-hoc systems.
It is fundamentally an extension of the Map-Shuffle-Reduce paradigm.
|Scales from a single node to thousand-node clusters.||Scales from a single node to thousand-node clusters.|
|Reuses the Pandas API and memory model. It implements neither SQL nor a query optimizer. It is able to do random access, efficient time series operations, and other Pandas-style indexed operations.||Has its own API and memory model. It also implements a large subset of the SQL language. Spark includes a high-level query optimizer for complex queries.|
|Relies on and interoperates with existing libraries like Scikit-Learn and XGBoost. These can be more familiar or higher performance, but generally results in a less-cohesive whole. See the dask-ml project for integrations.||Spark MLLib is a cohesive project with support for common operations that are easy to implement with Spark’s Map-Shuffle-Reduce style system. People considering MLLib might also want to consider other JVM-based machine learning libraries like H2O, which may have better performance.|
|Fully supports the NumPy model for scalable multi-dimensional arrays.||Does not include support for multi-dimensional arrays natively (this would be challenging given their computation model), although some support for two-dimensional matrices may be found in MLLib. People may also want to look at the Thunder project, which combines Apache Spark with NumPy arrays.|
|Provides a real-time futures interface that is lower-level than Spark streaming. This enables more creative and complex use-cases, but requires more work than Spark streaming.||Support for streaming data is first-class and integrates well into their other APIs. It follows a mini-batch approach. This provides decent performance on large uniform streaming operations.|
Graphs / complex networks¶
|Provides no such library.||Provides GraphX, a library for graph processing.|
|Allows you to specify arbitrary task graphs for more complex and custom systems that are not part of the standard set of collections.||Generally expects users to compose computations out of their high-level primitives (map, reduce, groupby, join, …). It is also possible to extend Spark through subclassing RDDs, although this is rarely done.|
Reasons you might choose Spark¶
- You prefer Scala or the SQL language
- You have mostly JVM infrastructure and legacy systems
- You want an established and trusted solution for business
- You are mostly doing business analytics with some lightweight machine learning
- You want an all-in-one solution
Reasons you might choose Dask¶
- You prefer Python or native code, or have large legacy code bases that you do not want to entirely rewrite
- Your use case is complex or does not cleanly fit the Spark computing model
- You want a lighter-weight transition from local computing to cluster computing
- You want to interoperate with other technologies and don’t mind installing multiple packages
Reasons to choose both¶
It is easy to use both Dask and Spark on the same data and on the same cluster.
They can both read and write common formats, like CSV, JSON, ORC, and Parquet, making it easy to hand results off between Dask and Spark workflows.
They can both deploy on the same clusters. Most clusters are designed to support many different distributed systems at the same time, using resource managers like Kubernetes and YARN. If you already have a cluster on which you run Spark workloads, it’s likely easy to also run Dask workloads on your current infrastructure and vice versa.
In particular, for users coming from traditional Hadoop/Spark clusters (such as those sold by Cloudera/Hortonworks) you are using the Yarn resource manager. You can deploy Dask on these systems using the Dask Yarn project, as well as other projects, like JupyterHub on Hadoop.
Both Spark and Dask represent computations with directed acyclic graphs. These graphs however represent computations at very different granularities.
One operation on a Spark RDD might add a node like
the graph. These are high-level operations that convey meaning and will
eventually be turned into many little tasks to execute on individual workers.
This many-little-tasks state is only available internally to the Spark
Dask graphs skip this high-level representation and go directly to the
many-little-tasks stage. As such, one
map operation on a Dask collection
will immediately generate and add possibly thousands of tiny tasks to the Dask
This difference in the scale of the underlying graph has implications on the kinds of analysis and optimizations one can do and also on the generality that one exposes to users. Dask is unable to perform some optimizations that Spark can because Dask schedulers do not have a top-down picture of the computation they were asked to perform. However, Dask is able to easily represent far more complex algorithms and expose the creation of these algorithms to normal users.
Spark is mature and all-inclusive. If you want a single project that does everything and you’re already on Big Data hardware, then Spark is a safe bet, especially if your use cases are typical ETL + SQL and you’re already using Scala.
Dask is lighter weight and is easier to integrate into existing code and hardware. If your problems vary beyond typical ETL + SQL and you want to add flexible parallelism to existing solutions, then Dask may be a good fit, especially if you are already using Python and associated libraries like NumPy and Pandas.