Read delta table in spark with NOT NULL constraint-CodePudding

I am reading a delta table like this

df = (spark
  .read
  .format("delta")
  .load("/path/to/foo/table")
  .select("Foo_ID", "Bar")
)

now when I look at the DataFrame schema, spark has somehow incorrectly inferred nullability:

>>> df.schema
...
StructType([StructField('Foo_ID', LongType(), True), StructField('Bar', TimestampType(), True)])

I know the values should not be null and I want the schema to reflect that by imposing a not null constraint on the columns. In case I try to load a table that actually does contain a null I want to see a big fat error message, rather than silently switching. Is there a clean way to do this? I came up with the following workaround, but it doesn't feel very nice:

my_schema = StructType([StructField('Foo_ID', LongType(), False), StructField('Bar', TimestampType(), False)])
df2 = spark.createDataFrame(df.rdd, schema=my_schema)

Is there a better way?

CodePudding user response：

If the table was created with the constraint, then Spark will infer nullability:

CREATE OR REPLACE TABLE delta.`/tmp/delta/table` (
  foo_id LONG NOT NULL,
  bar TIMESTAMP
) USING DELTA;

val df = spark.read.load("/tmp/delta/table")
df.printSchema

root
 |-- foo_id: long (nullable = false)
 |-- bar: timestamp (nullable = true)

Apparently it wasn't. But you can always change it:

CREATE OR REPLACE TABLE delta.`/tmp/delta/table` (
  foo_id LONG,
  bar TIMESTAMP
) USING DELTA;

ALTER TABLE delta.`/tmp/delta/table` CHANGE COLUMN foo_id foo_id LONG NOT NULL;
ALTER TABLE delta.`/tmp/delta/table` CHANGE COLUMN bar bar TIMESTAMP NOT NULL;

val df = spark.read.load("/tmp/delta/table")
df.printSchema

root
 |-- foo_id: long (nullable = false)
 |-- bar: timestamp (nullable = false)

Or, if you don't want or cannot change it, then what you came up with is not a workaround, it is a solution. You just explicitly specify a schema you want to read data with.