Loading Multiple CSV files across all subfolder levels with Wildcard file name - csv

Question

I want to Load Multiple CSV files matching certain names into a dataframe. Currently i am looping through the whole folder and creating a list of filenames and then loading those csv's into the dataframe list and then concatenating that dataframe.
The approach i want to use (if possible) is to bypass all the code and read all files in a one liner kind of approach.
I know this can be done easily for single level of subfolders, but my subfolder structure is as follows
Root Folder
|
Subfolder1
|
Subfolder 2
|
X01.csv
Y01.csv
Z01.csv
|
Subfolder3
|
Subfolder4
|
X01.csv
Y01.csv
|
Subfolder5
|
X01.csv
Y01.csv
I want to read all "X01.csv" files while reading from Root Folder.
Is there a way i can read all the required files in code something like the below
filepath = "rootpath" + "/**/X*.csv"
df = spark.read.format("com.databricks.spark.csv").option("recursiveFilelookup","true").option("header","true").load(filepath)
This code works fine for single level of subfolders, is there any equivalent of this for multi level folders ? i thought the "recursiveFilelookup" option would look across all levels of subfolders, but apparently this is not the way it works.
Currently i am getting a
Path not found ... filepath
exception
any help please

Answer 1

Have you tried using the glob.glob function?
You can use it to search for files that match certain criteria inside a root path, and pass the list of files it finds to spark.read.csv function.
For example, I've recreated the folder structure from your example inside a Google Colab environment:
To get a list of all CSV files matching the criteria you've specified, you can use the following code:
import glob
rootpath = './Root Folder/'
# The following line of code looks through all files
# inside the rootpath recursively, trying to match the
# pattern specified. In this case, it tries to find any
# CSV file that starts with the letters X, Y, or Z,
# and ends with 2 numbers (ranging from 0 to 9).
glob.glob(rootpath + "**/[X|Y|Z][0-9][0-9].csv", recursive=True)
# Returns:
# ['./Root Folder/Subfolder5/Y01.csv',
# './Root Folder/Subfolder5/X01.csv',
# './Root Folder/Subfolder1/Subfolder 2/Y01.csv',
# './Root Folder/Subfolder1/Subfolder 2/Z01.csv',
# './Root Folder/Subfolder1/Subfolder 2/X01.csv']
Now you can combine this with spark.read.csv capability of reading a list of files to get the answer you're looking for:
import glob
from pyspark.sql import SparkSession
spark = SparkSession.builder.getOrCreate()
rootpath = './Root Folder/'
spark.read.csv(glob.glob(rootpath + "**/[X|Y|Z][0-9][0-9].csv", recursive=True), inferSchema=True, header=True)
Note
You can specify more general patterns like:
glob.glob(rootpath + "**/*.csv", recursive=True)
To return a list of all csv files inside any subdirectory of rootpath.
Additionally, to consider only the immediate subdirectories files, you could use something like:
glob.glob(rootpath + "*.csv", recursive=True)
Edit
Based on your comments to this answer, does something like this works on Databricks?
from notebookutils import mssparkutils as ms
# databricks has a module called dbutils.fs.ls
# that works similarly to mssparkutils.fs, based on
# the following page of its documentation:
# https://docs.databricks.com/dev-tools/databricks-utils.html#ls-command-dbutilsfsls
def scan_dir(
initial_path: str,
search_str: str,
account_name: str,
):
"""Scan a directory and subdirectories for a string.
Parameters
----------
initial_path : str
The path to start the search. Accepts either a valid container name,
or the entire connection string.
search_str : str
The string to search.
account_name : str
The name of the account to access the container folders.
This value is only used, when the `initial_path`, doesn't
conform with the format: "abfss://<initial_path>#<account_name>.dfs.core.windows.net/"
Raises
------
FileNotFoundError
If the `initial_path` informed doesn't exist.
ValueError
If `initial_path` is not a string.
"""
if not isinstance(initial_path, str):
raise ValueError(
f'`initial_path` needs to be of type string, not {type(initial_path)}'
)
elif not initial_path.startswith('abfss'):
initial_path = f'abfss://{initial_path}#{account_name}.dfs.core.windows.net/'
try:
fdirs = ms.fs.ls(initial_path)
except Py4JJavaError as exc:
raise FileNotFoundError(
f'The path you informed \"{initial_path}\" doesn\'t exist'
) from exc
found = []
for path in fdirs:
p = path.path
if path.isDir:
found = [*found, *scan_dir(p, search_str)]
if search_str.lower() in path.name.lower():
# print(p.split('.net')[-1])
found = [*found, p.replace(path.name, "")]
return list(set(found))
Example:
# Change .parquet to .csv
spark.read.parquet(*scan_dir("abfss://CONTAINER_NAME#ACCOUNTNAME.dfs.core.windows.net/ROOT/FOLDER/", ".parquet"))
This method above worked for on Azure Synapse: