问题
I use Spark 2.2.0
I am reading a csv file as follows:
val dataFrame = spark.read.option("inferSchema", "true")
.option("header", true)
.option("dateFormat", "yyyyMMdd")
.csv(pathToCSVFile)
There is one date column in this file, and all records has a value equal to 20171001 for this particular column.
The issue is that spark is inferring that that the type of this column is integer rather than date. When I remove the "inferSchema" option, the type of that column is string.
There is no null values, nor any wrongly formatted line in this file.
What is the reason/solution for this issue?
回答1:
If my understanding is correct, the code implies the following order of type inference (with the first types being checked against first):
NullTypeIntegerTypeLongTypeDecimalTypeDoubleTypeTimestampTypeBooleanTypeStringType
With that, I think the issue is that 20171001 matches IntegerType before even considering TimestampType (which uses timestampFormat not dateFormat option).
One solution would be to define the schema and use it with schema operator (of DataFrameReader) or let Spark SQL infer the schema and use cast operator.
I'd choose the former if the number of fields is not high.
回答2:
In this case you simply cannot depend on the schema inference due to format ambiguity.
Since input can be parsed both as IntegerType (or any higher precision numeric format) as well as TimestamType and the former one has higher precedence (internally Spark tries IntegerType -> LongType -> DecimaType -> DoubleType -> TimestampType) inference mechanism will never reach TimestampType case.
To be specific, with schema inference enabled, Spark will call tryParseInteger, which will correctly parse the input and stop. Subsequent call will match the second case and finish at the same tryParseInteger call.
来源：https://stackoverflow.com/questions/46529404/how-to-force-inferschema-for-csv-to-consider-integers-as-dates-with-dateformat