Hi Kant, How about doing something like this?
import org.apache.spark.sql.functions._ // val df2 = df.select(df("body").cast(StringType).as("body")) val df2 = Seq("""{"a": 1}""").toDF("body") val schema = spark.read.json(df2.as[String].rdd).schema df2.select(from_json(col("body"), schema)).show() 2016-12-05 19:51 GMT+09:00 kant kodali <kanth...@gmail.com>: > Hi Michael, > > " Personally, I usually take a small sample of data and use schema > inference on that. I then hardcode that schema into my program. This > makes your spark jobs much faster and removes the possibility of the schema > changing underneath the covers." > > This may or may not work for us. Not all rows have the same schema. The > number of distinct schemas we have now may be small but going forward this > can go to any number moreover a distinct call can lead to a table scan > which can be billions of rows for us. > > I also would agree to keep the API consistent than making an exception > however I wonder if it make sense to provide an action call to infer the > schema which would return a new dataframe after the action call finishes > (after schema inference)? For example, something like below ? > > val inferedDF = df.inferSchema(col1); > > Thanks, > > > > > On Mon, Nov 28, 2016 at 6:12 PM, Michael Armbrust <mich...@databricks.com> > wrote: > >> You could open up a JIRA to add a version of from_json that supports >> schema inference, but unfortunately that would not be super easy to >> implement. In particular, it would introduce a weird case where only this >> specific function would block for a long time while we infer the schema >> (instead of waiting for an action). This blocking would be kind of odd for >> a call like df.select(...). If there is enough interest, though, we >> should still do it. >> >> To give a little more detail, your version of the code is actually doing >> two passes over the data: one to infer the schema and a second for whatever >> processing you are asking it to do. We have to know the schema at each >> step of DataFrame construction, so we'd have to do this even before you >> called an action. >> >> Personally, I usually take a small sample of data and use schema >> inference on that. I then hardcode that schema into my program. This >> makes your spark jobs much faster and removes the possibility of the schema >> changing underneath the covers. >> >> Here's some code I use to build the static schema code automatically >> <https://databricks-prod-cloudfront.cloud.databricks.com/public/4027ec902e239c93eaaa8714f173bcfc/1023043053387187/1128172975083446/2840265927289860/latest.html> >> . >> >> Would that work for you? If not, why not? >> >> On Wed, Nov 23, 2016 at 2:48 AM, kant kodali <kanth...@gmail.com> wrote: >> >>> Hi Michael, >>> >>> Looks like all from_json functions will require me to pass schema and >>> that can be little tricky for us but the code below doesn't require me to >>> pass schema at all. >>> >>> import org.apache.spark.sql._ >>> val rdd = df2.rdd.map { case Row(j: String) => j } >>> spark.read.json(rdd).show() >>> >>> >>> On Tue, Nov 22, 2016 at 2:42 PM, Michael Armbrust < >>> mich...@databricks.com> wrote: >>> >>>> The first release candidate should be coming out this week. You can >>>> subscribe to the dev list if you want to follow the release schedule. >>>> >>>> On Mon, Nov 21, 2016 at 9:34 PM, kant kodali <kanth...@gmail.com> >>>> wrote: >>>> >>>>> Hi Michael, >>>>> >>>>> I only see spark 2.0.2 which is what I am using currently. Any idea on >>>>> when 2.1 will be released? >>>>> >>>>> Thanks, >>>>> kant >>>>> >>>>> On Mon, Nov 21, 2016 at 5:12 PM, Michael Armbrust < >>>>> mich...@databricks.com> wrote: >>>>> >>>>>> In Spark 2.1 we've added a from_json >>>>>> <https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/functions.scala#L2902> >>>>>> function that I think will do what you want. >>>>>> >>>>>> On Fri, Nov 18, 2016 at 2:29 AM, kant kodali <kanth...@gmail.com> >>>>>> wrote: >>>>>> >>>>>>> This seem to work >>>>>>> >>>>>>> import org.apache.spark.sql._ >>>>>>> val rdd = df2.rdd.map { case Row(j: String) => j } >>>>>>> spark.read.json(rdd).show() >>>>>>> >>>>>>> However I wonder if this any inefficiency here ? since I have to >>>>>>> apply this function for billion rows. >>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >> >