PySpark，通过 JSON 文件导入模式

Question

tbschema.json looks like this:

[{"TICKET":"integer","TRANFERRED":"string","ACCOUNT":"STRING"}]

I load it using following code

>>> df2 = sqlContext.jsonFile("tbschema.json")
>>> f2.schema
StructType(List(StructField(ACCOUNT,StringType,true),
    StructField(TICKET,StringType,true),StructField(TRANFERRED,StringType,true)))
>>> df2.printSchema()
root
 |-- ACCOUNT: string (nullable = true)
 |-- TICKET: string (nullable = true)
 |-- TRANFERRED: string (nullable = true)

1. Why does the schema elements gets sorted, when I want the elements in the same order as they appear in the JSON.

2. The data type integer has been converted into StringType after the JSON has been derived, how do I retain the datatype.

Answer 1

Why does the schema elements gets sorted, when i want the elemets in the same order as they appear in the json.

Because order of fields is not guaranteed. While it is not explicitly stated it becomes obvious when you take a look a the examples provided in the JSON reader doctstring. If you need specific ordering you can provide schema manually:

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

schema = StructType([
    StructField("TICKET", StringType(), True),
    StructField("TRANFERRED", StringType(), True),
    StructField("ACCOUNT", StringType(), True),
])
df2 = sqlContext.read.json("tbschema.json", schema)
df2.printSchema()

root
 |-- TICKET: string (nullable = true)
 |-- TRANFERRED: string (nullable = true)
 |-- ACCOUNT: string (nullable = true)

The data type integer has been converted into StringType after the json has been derived, how do i retain the datatype.

Data type of JSON field TICKET is string hence JSON reader returns string. It is JSON reader not some-kind-of-schema reader.

Generally speaking you should consider some proper format which comes with schema support out-of-the-box, for example Parquet , Avro or Protocol Buffers . But if you really want to play with JSON you can define poor man's "schema" parser like this:

from collections import OrderedDict 
import json

with open("./tbschema.json") as fr:
    ds = fr.read()

items = (json
  .JSONDecoder(object_pairs_hook=OrderedDict)
  .decode(ds)[0].items())

mapping = {"string": StringType, "integer": IntegerType, ...}

schema = StructType([
    StructField(k, mapping.get(v.lower())(), True) for (k, v) in items])

Problem with JSON is that there is really no guarantee regarding fields ordering whatsoever, not to mention handling missing fields, inconsistent types and so on. So using solution as above really depends on how much you trust your data.

Alternatively you can use built-in schema import / export utilities .