RecordIO

RecordIO is a file format created for PaddlePaddle Elastic Deep Learning. It is generally useful for distributed computing.

Motivations

In distributed computing, we often need to dispatch tasks to worker processes. Usually, a task is defined as a partition of the input data, like what MapReduce and distributed machine learning do.

Most distributed filesystems, including HDFS, Google FS, and CephFS, prefer a small number of big files. Therefore, it is impractical to create each task as a small file; instead, we need a format for big files that is

1. appendable, so that applications can append records to the file without updating the meta-data, thus fault tolerable,
2. partitionable, so that applications can quickly scan over the file to count the total number of records, and create tasks each corresponds to a sequence of records.

RecordIO is such a file format.

Write

import recordio

# write
with recordio.File('demo.recordio', 'w') as rdio_w: 
    rdio_w.write('abc')
    rdio_w.write('def')

Read

with recordio.File('demo.recordio', 'r') as rdio_r:
    # Random access
    for i in range(rdio_r.count()):
        print(rdio_r.get(i))

    # Range reading
    for record in rdio_r.get_reader(2, 10):
        print(record)

    # Direct iteration
    for record in rdio_r:
        print(record)

Unittest

In this directory:

python -m unittest recordio/recordio/*_test.py

Packaging

The package process largely follows the example of Tensorflow custom op

First build RecordIO devel Docker image:

docker build -t recordio:dev -f Dockerfile .

Start Docker container and map git and bazel .cache directories:

docker run --rm -it \
    -v $HOME/git:/git \
    -v $HOME/.cache:/.cache \
    -w /git/pyrecordio \
    recordio:dev

Inside container, build the pip package:

bazel build build_pip_pkg 
bazel-bin/build_pip_pkg artifacts

After building the package, force install it to replace the existing version.

pip install -I artifacts/recordio-<version>.whl

To test the installed TensorFlow RecordIO Dataset op:

python recordio/tensorflow_op/python/tf_recordio_dataset_test.py

There is also a prepackded version checked into the repo for convenience.