Using predicates to filter rows from pyarrow.parquet.ParquetDataset

Note: I’ve expanded this into a comprehensive guide to Python and Parquet in this post

Parquet Format Partitions

In order to use filters you need to store your data in Parquet format using partitions. Loading a few Parquet columns and partitions out of many can result in massive improvements in I/O performance with Parquet versus CSV. Parquet can partition files based on values of one or more fields and it creates a directory tree for the unique combinations of the nested values, or just one set of directories for one partition column. The PySpark Parquet documentation explains how Parquet works fairly well.

A partition on gender and country would look like this:

path└── to    └── table        ├── gender=male        │   ├── ...        │   │        │   ├── country=US        │   │   └── data.parquet        │   ├── country=CN        │   │   └── data.parquet        │   └── ...

There is also row group partitioning if you need to further partition your data, but most tools only support specifying row group size and you have to do the key-->row group lookup yourself, which is ugly (happy to answer about that in another question).

Writing Partitions with Pandas

You need to partition your data using Parquet and then you can load it using filters. You can write the data in partitions using PyArrow, pandas or Dask or PySpark for large datasets.

For example, to write partitions in pandas:

df.to_parquet(    path='analytics.xxx',     engine='pyarrow',    compression='snappy',    columns=['col1', 'col5'],    partition_cols=['event_name', 'event_category'])

This lays the files out like:

analytics.xxx/event_name=SomeEvent/event_category=SomeCategory/part-0001.c000.snappy.parquetanalytics.xxx/event_name=SomeEvent/event_category=OtherCategory/part-0001.c000.snappy.parquetanalytics.xxx/event_name=OtherEvent/event_category=SomeCategory/part-0001.c000.snappy.parquetanalytics.xxx/event_name=OtherEvent/event_category=OtherCategory/part-0001.c000.snappy.parquet

Loading Parquet Partitions in PyArrow

To grab events by one property using the partition columns, you put a tuple filter in a list:

import pyarrow.parquet as pqimport s3fsfs = s3fs.S3FileSystem()dataset = pq.ParquetDataset(    's3://analytics.xxx',     filesystem=fs,     validate_schema=False,     filters=[('event_name', '=', 'SomeEvent')])df = dataset.to_table(    columns=['col1', 'col5']).to_pandas()

Filtering with Logical ANDs

To grab an event with two or more properties using AND you just create a list of filter tuples:

import pyarrow.parquet as pqimport s3fsfs = s3fs.S3FileSystem()dataset = pq.ParquetDataset(    's3://analytics.xxx',     filesystem=fs,     validate_schema=False,     filters=[        ('event_name',     '=', 'SomeEvent'),        ('event_category', '=', 'SomeCategory')    ])df = dataset.to_table(    columns=['col1', 'col5']).to_pandas()

Filtering with Logical ORs

To grab two events using OR you need to nest the filter tuples in their own lists:

import pyarrow.parquet as pqimport s3fsfs = s3fs.S3FileSystem()dataset = pq.ParquetDataset(    's3://analytics.xxx',     filesystem=fs,     validate_schema=False,     filters=[        [('event_name', '=', 'SomeEvent')],        [('event_name', '=', 'OtherEvent')]    ])df = dataset.to_table(    columns=['col1', 'col5']).to_pandas()

Loading Parquet Partitions with AWS Data Wrangler

As another answer mentioned, the easiest way to load data filtering to just certain columns in certain partitions wherever the data is located (locally or in the cloud) is to use the awswrangler module. If you're using S3, check out the documentation for awswrangler.s3.read_parquet() and awswrangler.s3.to_parquet(). The filtering works the same as with the examples above.

import awswrangler as wrdf = wr.s3.read_parquet(    path="analytics.xxx",    columns=["event_name"],     filters=[('event_name', '=', 'SomeEvent')])

Loading Parquet Partitions with pyarrow.parquet.read_table()

If you're using PyArrow, you can also use pyarrow.parquet.read_table():

import pyarrow.parquet as pqfp = pq.read_table(    source='analytics.xxx',    use_threads=True,    columns=['some_event', 'some_category'],    filters=[('event_name', '=', 'SomeEvent')])df = fp.to_pandas()

Loading Parquet Partitions with PySpark

Finally, in PySpark you can use pyspark.sql.DataFrameReader.read_parquet()

import pyspark.sql.functions as Ffrom pyspark.sql import SparkSessionspark = SparkSession.builder.master("local[1]") \                    .appName('Stack Overflow Example Parquet Column Load') \                    .getOrCreate()# I automagically employ Parquet structure to load the selected columns and partitionsdf = spark.read.parquet('s3://analytics.xxx') \          .select('event_name', 'event_category') \          .filter(F.col('event_name') == 'SomeEvent')

Hopefully this helps you work with Parquet :)

For anyone getting here from Google, you can now filter on rows in PyArrow when reading a Parquet file. Regardless if you read it via pandas or pyarrow.parquet.

From the documentation:

filters (List[Tuple] or List[List[Tuple]] or None (default)) – Rows which do not match the filter predicate will be removed from scanned data. Partition keys embedded in a nested directory structure will be exploited to avoid loading files at all if they contain no matching rows. If use_legacy_dataset is True, filters can only reference partition keys and only a hive-style directory structure is supported. When setting use_legacy_dataset to False, also within-file level filtering and different partitioning schemes are supported.

Predicates are expressed in disjunctive normal form (DNF), like [[('x', '=', 0), ...], ...]. DNF allows arbitrary boolean logical combinations of single column predicates. The innermost tuples each describe a single column predicate. The list of inner predicates is interpreted as a conjunction (AND), forming a more selective and multiple column predicate. Finally, the most outer list combines these filters as a disjunction (OR).

Predicates may also be passed as List[Tuple]. This form is interpreted as a single conjunction. To express OR in predicates, one must use the (preferred) List[List[Tuple]] notation.

Currently, the filters functionality is only implemented at the file level, not yet at the row level.

So if you have a dataset as a collection of multiple, partitioned parquet files in a nested hierarchy (the type of partitioned datasets described here: https://arrow.apache.org/docs/python/parquet.html#partitioned-datasets-multiple-files), you can use the filters argument to only read a subset of the files.
But, you can't yet use it for reading only a subset of the row groups of a single file (see https://issues.apache.org/jira/browse/ARROW-1796).

But, it would be nice that you get an error message of specifying such an invalid filter. I opened an issue for that: https://issues.apache.org/jira/browse/ARROW-5572