Apache Spark - how to set timezone to UTC? currently defaulted to Zulu

Now you can use:

spark.conf.set("spark.sql.session.timeZone", "UTC")

Since https://issues.apache.org/jira/browse/SPARK-18936 in 2.2.0

EDIT:

Additionally I set my default TimeZone to UTC to avoid implicit conversions

TimeZone.setDefault(TimeZone.getTimeZone("UTC"))

Otherwise you will get implicit conversions from your default Timezone to UTC when no Timezone information is present in the Timestamp you're converting

Example:

val rawJson = """ {"some_date_field": "2018-09-14 16:05:37"} """val dsRaw = sparkJob.spark.createDataset(Seq(rawJson))val output =  dsRaw    .select(      from_json(        col("value"),        new StructType(          Array(            StructField("some_date_field", DataTypes.TimestampType)          )        )      ).as("parsed")    ).select("parsed.*")

If my default TimeZone is Europe/Dublin which is GMT+1 and Spark sql session timezone is set to UTC, Spark will assume that "2018-09-14 16:05:37" is in Europe/Dublin TimeZone and do a conversion (result will be "2018-09-14 15:05:37")

In some cases you will also want to set the JVM timezone. For example, when loading data into a TimestampType column, it will interpret the string in the local JVM timezone. To set the JVM timezone you will need to add extra JVM options for the driver and executor:

spark = pyspark.sql.SparkSession \    .Builder()\    .appName('test') \    .master('local') \    .config('spark.driver.extraJavaOptions', '-Duser.timezone=GMT') \    .config('spark.executor.extraJavaOptions', '-Duser.timezone=GMT') \    .config('spark.sql.session.timeZone', 'UTC') \    .getOrCreate()

We do this in our local unit test environment, since our local time is not GMT.

Useful reference:https://en.wikipedia.org/wiki/List_of_tz_database_time_zones

As described in these SPARK bug reports (link, link), the most current SPARK versions (3.0.0 and 2.4.6 at time of writing) do not fully/correctly support setting the timezone for all operations, despite the answers by @Moemars and @Daniel.

I suggest avoiding time operations in SPARK as much as possible, and either perform them yourself after extraction from SPARK or by using UDFs, as used in this question.