Convert lines of JSON in RDD to dataframe in Apache Spark

问题

I have some 17,000 files in S3 that look like this:

{"hour": "00", "month": "07", "second": "00", "year": "1970", "timezone": "-00:00", "day": "12", "minute": "00"}
{"hour": "00", "month": "07", "second": "01", "year": "1970", "timezone": "-00:00", "day": "12", "minute": "00"}
{"hour": "00", "month": "07", "second": "02", "year": "1970", "timezone": "-00:00", "day": "12", "minute": "00"}
{"hour": "00", "month": "07", "second": "03", "year": "1970", "timezone": "-00:00", "day": "12", "minute": "00"}
{"hour": "00", "month": "07", "second": "04", "year": "1970", "timezone": "-00:00", "day": "12", "minute": "00"}

I have one file per day. Each file contains a record for each second. ∴ 86,000 records in a file. Each file has a file name like "YYYY-MM-DD".

Using boto3 I generate a list of the files in the bucket. Here I am selecting only 10 files using the prefix.

import boto3
s3_list = []
s3 = boto3.resource('s3')
my_bucket = s3.Bucket('time-waits-for-no-man')
for object in my_bucket.objects.filter(Prefix='1972-05-1):
    s3_list.append(object.key)

This function returns a list of files(S3 keys). I then define a function to fetch a file and return the rows:

def FileRead(s3Key):
    s3obj = boto3.resource('s3').Object(bucket_name='bucket', key=s3Key)
    contents = s3obj.get()['Body'].read().decode('utf-8')
    yield Row(**contents)

I then distribute this function using flatMap:

job = sc.parallelize(s3_list)
foo = job.flatMap(FileRead)

Problem

I'm not able to work out how to properly pump these rows into a Dataframe however.

>>> foo.toDF().show()
+--------------------+                                                          
|                  _1|
+--------------------+
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
|{"hour": "00", "m...|
+--------------------+

>>> foo.toDF().count()
10

Please could someone show me how to do this?

回答1:

You should probably use json reader directly (spark.read.json / sqlContext.read.json) but if you know the schema you can try parsing JSON string manually:

from pyspark.sql.types import StructField, StructType, StringType
from pyspark.sql import Row
import json

fields = ['day', 'hour', 'minute', 'month', 'second', 'timezone', 'year']
schema =  StructType([
  StructField(field, StringType(), True) for field in fields
])

def parse(s, fields):
    try:
        d = json.loads(s[0])
        return [tuple(d.get(field) for field in fields)]
    except:
        return []

spark.createDataFrame(foo.flatMap(lambda s: parse(s, fields)), schema)

You can also use get_json_object:

from pyspark.sql.functions import get_json_object

df.select([
    get_json_object("value", "$.{0}".format(field)).alias(field)
    for field in fields
])

回答2:

In the end I got it working with:

def FileRead(s3Key):
    s3obj = boto3.resource('s3').Object(bucket_name='bucket', key=s3Key)
    contents = s3obj.get()['Body'].read().decode()
    result = []
    meow = contents.split('\n')
    index = 0
    limit = 10
    for item in meow:
        index += 1
        result.append(json.loads(item))
        if index == limit:
            return result

job = sc.parallelize(s3_list)
foo = job.flatMap(distributedJsonRead)
df = foo.toDF()

Thanks @user6910411 for the inspiration.

来源：https://stackoverflow.com/questions/39818368/convert-lines-of-json-in-rdd-to-dataframe-in-apache-spark

标签

apache-spark

pyspark

apache-spark-sql

rdd