Skip to content

Instantly share code, notes, and snippets.

What would you like to do?
Writing Spark DataFrame to file
# Write a spark DataFrame into a single CSV files (to open with Excel/other tools easily)
# Save the file to S3
import s3fs
import pyspark.sql.functions as F # noqa: N812
def spark_to_csv(spark_df, out_path):
Save the file in part files with spark and then append them together
:param spark_df: The spark dataframe to write
:param out_path: The S3 location where the CSV should be kept
s3 = s3fs.S3FileSystem(anon=False)
out_path = out_path.rstrip('/')
drop_names = []
for name, dtype in spark_df.dtypes:
if dtype.startswith('map'):
print('Skipping column:', name, 'as it has dtype:', dtype)
if dtype.startswith('array'):
spark_df = spark_df.withColumn(name, F.udf(lambda x: str(x))(spark_df[name]))
spark_df = spark_df.drop(*drop_names)
print('Writing to {} CSV part files ...'.format(spark_df.rdd.getNumPartitions()))
# Spark needs s3a:// to write S3 files (in my clusters atleast)
spark_df.write.format('csv') \
.option('header', 'false') \
.option('escape', '"') \
.save(out_path.replace('s3://', 's3a://') + '_parts', mode='overwrite')
# Merge the part files into the output
print('Writing part-files to merged CSV file ...')
partfiles = sorted(s3.walk(out_path + '_parts' + '/'))
with, 'wb') as fhandler:
# Add the header as the first row
fhandler.write((','.join(list(spark_df.columns)) + '\n').encode('utf-8'))
# Add the rest of the data from the part files
for ipfile, pfile in partfiles:
print('Merging part {} of {}'.format(ipfile, len(partfiles)))
with, 'rb') as pfilehandler:
# Cleanup the temp part file location that was created
s3.rm(out_path + '_parts', recursive=True)
print('Final file saved at: {}'.format(out_path))
spark_to_csv(df, 's3://datapath/out.csv')
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment