spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Eskilson,Aleksander" <Alek.Eskil...@Cerner.com>
Subject Re: SparkR DataFrame Column Casts esp. from CSV Files
Date Wed, 03 Jun 2015 17:49:42 GMT
Hi Shivaram,

As far as databricks’ spark-csv API shows, it seems there’s currently only support for
explicit definition of column types. In JSON we have nice typed fields, but in CSVs, all bets
are off. In the SQL version of the API, it appears you specify the column types when you create
the table you’re populating with CSV data.

Thanks for the clarification on individual column casting, I was missing the more obvious
syntax.

I’ll file a JIRA for resetting the schema after loading a DF.

Thanks,
Alek


From: Shivaram Venkataraman <shivaram@eecs.berkeley.edu<mailto:shivaram@eecs.berkeley.edu>>
Reply-To: "shivaram@eecs.berkeley.edu<mailto:shivaram@eecs.berkeley.edu>" <shivaram@eecs.berkeley.edu<mailto:shivaram@eecs.berkeley.edu>>
Date: Wednesday, June 3, 2015 at 12:29 PM
To: Aleksander Eskilson <Alek.Eskilson@cerner.com<mailto:Alek.Eskilson@cerner.com>>
Cc: "dev@spark.apache.org<mailto:dev@spark.apache.org>" <dev@spark.apache.org<mailto:dev@spark.apache.org>>,
"hossein@databricks.com<mailto:hossein@databricks.com>" <hossein@databricks.com<mailto:hossein@databricks.com>>
Subject: Re: SparkR DataFrame Column Casts esp. from CSV Files

cc Hossein who knows more about the spark-csv options

You are right that the default CSV reader options end up creating all columns as string. I
know that the JSON reader infers the schema [1] but I don't know if the CSV reader has any
options to do that.  Regarding the SparkR syntax to cast columns, I think there is a simpler
way to do it by just assigning to the same column name. For example I have a flights DataFrame
with the `year` column typed as string. To cast it to int I just use

flights$year <- cast(flights$year, "int")

Now the dataframe has the same number of columns as before and you don't need a selection.

However this still doesn't address the part about casting multiple columns -- Could you file
a new JIRA to track the need for casting multiple columns or rather being able to set the
schema after loading a DF ?

Thanks
Shivaram

[1] http://spark.apache.org/docs/latest/sql-programming-guide.html#json-datasets<https://urldefense.proofpoint.com/v2/url?u=http-3A__spark.apache.org_docs_latest_sql-2Dprogramming-2Dguide.html-23json-2Ddatasets&d=AwMFaQ&c=NRtzTzKNaCCmhN_9N2YJR-XrNU1huIgYP99yDsEzaJo&r=0vZw1rBdgaYvDJYLyKglbrax9kvQfRPdzxLUyWSyxPM&m=aCZhOxAn5Iu762hWogwQK__JsZigsbLZFMaz44UcKQw&s=BX3MuobG748zhfm7hc_SnZA4MnFbwgFreNVEjkzkENc&e=>

On Wed, Jun 3, 2015 at 7:51 AM, Eskilson,Aleksander <Alek.Eskilson@cerner.com<mailto:Alek.Eskilson@cerner.com>>
wrote:
It appears that casting columns remains a bit of a trick in Spark’s DataFrames. This is
an issue because tools like spark-csv will set column types to String by default and will
not attempt to infer types. Although spark-csv supports specifying  types for columns in its
options, it’s not clear how that might be integrated into SparkR (when loading the spark-csv
package into the R session).

Looking at the column.R spec we can cast a column to a different data type with the cast function
[1], but it’s notable that this is not a mutator, and it returns a column object as opposed
to a DataFrame. It appears the column cast can only be ‘applied’ by using the withColumn()
or mutate() (an alias for withColumn).

The other way to cast with Spark DataFrames is to write UDFs that operate on a column value
and return a coerced value. It looks like SparkR doesn’t have UDFs just yet [2], but it
seems like they’d be necessary to do a natural one-off column cast in R, something like

df.col1toInt <- withColumn(df, “intCol1”, udf(df$col1, function(x) as.numeric(x)))

(where col1 was originally ‘character’ type)

Currently it seems one has to
df.col1cast <- cast(df$col1, “int”)
df.col1toInt <- withColumn(df, df.col1cast)

If we wanted just our casted columns and not the original column from the data frame, we’d
still have to do a select. There was a conversation about CSV files just yesterday. Types
are already problematic, but they’re a very common data source in R, even at scale.

But only being able to coerce one column at a time is really unwieldy. Can the current spark-csv
SQL API for specifying types [3] be extended SparkR? And are there any thoughts on implementing
some kind of type inferencing perhaps based on a sampling of some number of rows (an implementation
I’ve seen before)? R’s read.csv() and read.delim() get types by inferring from the whole
file. Getting something that can achieve that functionality via explicit definition of types
or sampling will probably be necessary to work with CSV files that have enough columns to
merit R at Spark’s scale.

Regards,
Alek Eskilson

[1] - https://github.com/apache/spark/blob/master/R/pkg/R/column.R#L190<https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_apache_spark_blob_master_R_pkg_R_column.R-23L190&d=AwMFaQ&c=NRtzTzKNaCCmhN_9N2YJR-XrNU1huIgYP99yDsEzaJo&r=0vZw1rBdgaYvDJYLyKglbrax9kvQfRPdzxLUyWSyxPM&m=aCZhOxAn5Iu762hWogwQK__JsZigsbLZFMaz44UcKQw&s=pETagpDWepAmeaxucEKv1BgoCjqqpIejSjZhXZFF_y8&e=>
[2] - https://issues.apache.org/jira/browse/SPARK-6817<https://urldefense.proofpoint.com/v2/url?u=https-3A__issues.apache.org_jira_browse_SPARK-2D6817&d=AwMFaQ&c=NRtzTzKNaCCmhN_9N2YJR-XrNU1huIgYP99yDsEzaJo&r=0vZw1rBdgaYvDJYLyKglbrax9kvQfRPdzxLUyWSyxPM&m=aCZhOxAn5Iu762hWogwQK__JsZigsbLZFMaz44UcKQw&s=tiLELUAU2Sgk680gUGLr9fR9YxEU6lJEs2e0gWenWhs&e=>
[3] - https://github.com/databricks/spark-csv#sql-api<https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_databricks_spark-2Dcsv-23sql-2Dapi&d=AwMFaQ&c=NRtzTzKNaCCmhN_9N2YJR-XrNU1huIgYP99yDsEzaJo&r=0vZw1rBdgaYvDJYLyKglbrax9kvQfRPdzxLUyWSyxPM&m=aCZhOxAn5Iu762hWogwQK__JsZigsbLZFMaz44UcKQw&s=89QC5nymwl5GjjpMwUD--828WaTvjqik9glbCHR7T-8&e=>

CONFIDENTIALITY NOTICE This message and any included attachments are from Cerner Corporation
and are intended only for the addressee. The information contained in this message is confidential
and may constitute inside or non-public information under international, federal, or state
securities laws. Unauthorized forwarding, printing, copying, distribution, or use of such
information is strictly prohibited and may be unlawful. If you are not the addressee, please
promptly delete this message and notify the sender of the delivery error by e-mail or you
may call Cerner's corporate offices in Kansas City, Missouri, U.S.A at (+1) (816)221-1024<tel:%28%2B1%29%20%28816%29221-1024>.


Mime
View raw message