Earlier, development for the Feather file format moved to the Apache Arrow codebase. I will explain how this has already affected Feather and what to expect from the project going forward.
Feather: How's it related to Arrow?
Shortly after we announced the formation of Apache Arrow in February 2016, Hadley Wickham and I met up and discussed how we could foster more collaboration in the Python and R community around shared infrastructure for data science. Hadley suggested (as I recall) the idea of a binary file format for data frames that uses Arrow as the unifying technology.
Since it seemed likely that the broader build out of the Arrow project would take some time, we decided to create a minimalistic implementation of Arrow with just enough file metadata to support the major Python and R data types. We released this as the Feather project. The implementation was overall incredibly simple and has been very useful to data scientists that use both Python and R.
Some time after, in the Arrow project we designed more general streaming and batch file formats that allow fast, zero-copy access of arbitrarily large tabular data sets.
Compared with Arrow streams and files, Feather has some limitations:
- Supports limited scalar value types, adequate only for representing typical data found in R and pandas
- Supports only a single batch of rows, where general Arrow streams support an arbitrary number
- Only non-nested data types and categorical (dictionary-encoded) types are supported
So, Feather files are Arrow memory on disk (and thus support zero-copy access), but have more limited metadata. There are some obvious other things we'd like to add to the Feather format:
- Column-wise compression (e.g. using LZ4 or ZSTD codecs)
- Chunked writes
- Ability to append to existing files
- Support for nested data
One easy way to achieve these goals is to replace the Feather internals with the more general Arrow stream formats. This isn't as much as it might seem, but will require solving some logistical hurdles...
Moving the Feather format to the Arrow codebase
Earlier this year, I ported the Feather file implementation to fit in with the
rest of the more general Arrow C++ in-memory data structures and memory
model. Reading a Feather file column is a zero-copy operation that returns an
arrow::Column C++ object. These C++ containers can then be converted to
data.frame, or whichever desired consumer of the data.
Moving Feather development to the Arrow codebase has made the implementation substantially simpler, and Feather users immediately benefit from any performance improvements or general Arrow-related development work. So Feather has gotten faster and faster in Python this year as we've worked on enhancing the bridge between pandas and Arrow this year.
One problem with moving Feather to the Arrow codebase is that there are not yet R bindings available for the Arrow C++ libraries. As soon as we have R bindings available for Arrow, I will be quite keen to get to work adding new Feather features and transitioning Feather internally to reuse Arrow's stream format and metadata, making the implementation even simpler.
Performance improvements in Feather for Python users
I released Feather 0.4.0 on May 24 as a simple wrapper around
pyarrow.feather. So this means that when
pyarrow gets faster, so does
feather. One of the primary benefits has been multithreaded
conversions. We have had multithreaded conversions from Arrow to pandas
since Arrow 0.2.0. I recently implemented multithreaded conversions from pandas
to Arrow, and these will ship in Arrow 0.8.0.
Let's take a look at read speeds in Feather 0.3.1 (which had a self-contained implementation) compared with Feather 0.4.0 using Arrow 0.7.1. To do this, I generated a 512 MB DataFrame with floating point columns, write it to Feather, then measure the performance for reading it completely back into memory. I varied the number of columns in the DataFrame while keeping the total data size the same. For some reason, Feather 0.3.1 is faster reading this dataset with 50 columns than with 10 columns. You can look at the complete notebook for the benchmark.
Perhaps amazingly, Feather 0.4.0 based on pyarrow 0.7.1 is much faster than Feather 0.3.1:
To put this performance concretely: on my laptop, Feather 0.3.1 can read a 50-column DataFrame at about 1.6 GB/s while Feather 0.4.0 (based on pyarrow 0.7.1) can read at about 2.9 GB/s, a reduction in runtime of over 40%.
This improved performance is for multiple reasons:
- Multithreaded conversion from Arrow's bitmap-based null encoding to pandas's
NaN-based null encoding for doubles
- More precise and efficient creation of the pandas DataFrame internal data structure
What to expect in the future
Until R bindings for Apache Arrow ship, it will be difficult to innovate on
Feather's feature set, since the existing R
feather library in CRAN is based
on the Feather 0.3.1 codebase from the wesm/feather repository. Once that
happens, we should be able to make R's Feather read performance faster and more
consistent as we have with the Python implementation above.
I'm excited to engage with the R community on this work and look forward to delivering more Apache Arrow-based IO and computation tools to both the R and Python communities.