Skip to content

Latest commit

 

History

History
50 lines (31 loc) · 2.24 KB

README.md

File metadata and controls

50 lines (31 loc) · 2.24 KB
Raw

schemadiff

schemadiff is a niche package designed for situations where a — large — number of files on a filesystem are expected to have identical schemas, but they don't. This can present a challenge when working with distributed computing systems like Apache Spark or Google BigQuery, as unexpected schema differences can disrupt data loading and processing.

Consider a scenario where you are processing thousands of files, and a subset of them have schemas that are almost identical but not completely matching. This can lead to errors such as:

  • BigQuery: Error while reading data, error message: Parquet column '<COLUMN_NAME>' has type INT32 which does not match the target cpp_type DOUBLE File: gs://bucket/file.parquet
  • Spark: Error: java.lang.UnsupportedOperationException: org.apache.parquet.column.values.dictionary.PlainValuesDictionary$PlainDoubleDictionary

schemadiff addresses these issues by efficiently identifying the files with schema inconsistencies through reading file metadata.

Installation

Install the package with pip:

pip install schemadiffed # schemadiff taken :p

Usage

The package can be used as a Python library or as a command-line tool.

Python Library

Here's an example of using schemadiff to group files by their schema:

import os
from schemadiff import compare_schemas

os.environ['GOOGLE_CLOUD_CREDENTIALS'] = 'key.json'
grouped_files = compare_schemas('path/to/parquet_files', report_path='/desired/path/to/report.json')

In this example, compare_schemas groups the Parquet files in the directory path/to/parquet_files by their schema. It saves the results to report.json and also returns the grouped files as a list for potential downstream use.

Command-Line Interface

schemadiff can also be used as a command-line tool. After installation, the command compare-schemas is available in your shell:

python schemadiff  --dir_path 'gs://<bucket>/yellow/*_2020*.parquet' --fs_type 'gcs' --report_path 'report.json' --return_type 'as_list'

Features

  • Efficient processing by reading the metadata of Parquet files.
  • Supports local, GCS, S3 filesystems (you must be authenticated to your cloud service first).
  • Supports wildcard characters for flexible file selection.