Save schema of dataframe in S3 location

user_7546442 0
Viewed 4580

I read in a parquet file from S3 in databricks using the following command

df = sqlContext.read.parquet('s3://path/to/parquet/file')

I want to read the schema of the dataframe, which I can do using the following command:

df_schema = df.schema.json()

But I am not able to write the df_schama object to a file on S3. Note: I am open to not creating a json file. I just want to save the schema of the dataframe to any file type (possibly a text file) in AWS S3.

I have tried writing the json schema as follows,

df_schema.write.csv("s3://path/to/file")

or

a.write.format('json').save('s3://path/to/file')

Both of them give me the following errors:

AttributeError: 'str' object has no attribute 'write'

pyspark databricks apache-spark amazon-s3
2 Answers

Here is a working example of saving a schema and applying it to new csv data:

# funcs
from pyspark.sql.functions import *
from pyspark.sql.types import *

# example old df schema w/ long datatype
df = spark.range(10)
df.printSchema()
df.write.mode("overwrite").csv("old_schema")

root
 |-- id: long (nullable = false)

# example new df schema we will save w/ int datatype
df = df.select(col("id").cast("int"))
df.printSchema()

root
 |-- id: integer (nullable = false)

# get schema as json object
schema = df.schema.json()

# write/read schema to s3 as .txt
import json

with open('s3:/path/to/schema.txt', 'w') as F:  
    json.dump(schema, F)

with open('s3:/path/to/schema.txt', 'r') as F:  
    saved_schema = json.load(F)

# saved schema
saved_schema
'{"fields":[{"metadata":{},"name":"id","nullable":false,"type":"integer"}],"type":"struct"}'

# construct saved schema object
new_schema = StructType.fromJson(json.loads(saved_schema))

new_schema
StructType(List(StructField(id,IntegerType,false)))

# use saved schema to read csv files ... new df has int datatype and not long
new_df = spark.read.csv("old_schema", schema=new_schema)
new_df.printSchema()
root
 |-- id: integer (nullable = true)
user_112865880

df.schema.json() results string object and string objects won't have .write method.

In RDD Api:

df_schema = df.schema.json()

parallelize df_schema variable to create rdd and then use .saveAsTextFile method to write the schema to s3.

sc.parallelize([df_schema]).saveAsTextFile("s3://path/to/file")

(or)

In Dataframe Api:

from pyspark.sql import Row
df_schema = df.schema.json()
df_sch=sc.parallelize([Row(schema=df1)]).toDF()
df_sch.write.csv("s3://path/to/file")
df_sch.write.text("s3://path/to/file") //write as textfile
user_76326950
Related