[
https://issues.apache.org/jira/browse/SPARK-42635?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Chenhao Li updated SPARK-42635:
-------------------------------
Description:
# When the time is close to daylight saving time transition, the result may be
discontinuous and not monotonic.
As pointed out by me and @Utkarsh Agarwal in
[https://github.com/databricks/runtime/pull/54936/files#r1118047445], the
result is conter-intuitive when the time is close to daylight saving time
transition and the added amount is close to the multiple of days.
We currently have:
{{scala> spark.conf.set("spark.sql.session.timeZone", "America/Los_Angeles")
scala> spark.sql("select timestampadd(second, 24 * 3600 - 1,
timestamp'2011-03-12 03:00:00')").show
+------------------------------------------------------------------------+
|timestampadd(second, ((24 * 3600) - 1), TIMESTAMP '2011-03-12 03:00:00')|
+------------------------------------------------------------------------+
| 2011-03-13 03:59:59|
+------------------------------------------------------------------------+
scala> spark.sql("select timestampadd(second, 24 * 3600, timestamp'2011-03-12
03:00:00')").show
+------------------------------------------------------------------+
|timestampadd(second, (24 * 3600), TIMESTAMP '2011-03-12 03:00:00')|
+------------------------------------------------------------------+
| 2011-03-13 03:00:00|
+------------------------------------------------------------------+}}
In the second query, adding one more second will set the time back one hour
instead. Plus, there is only 23 * 3600seconds from 2011-03-12 03:00:00 to
2011-03-13 03:00:00, instead of 24 * 3600 seconds, due to the daylight saving
time transition.
The root cause of the problem is the Spark code at
[https://src.dev.databricks.com/databricks/runtime/-/blob/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/util/DateTimeUtils.scala?L797]
wrongly assumes every day has MICROS_PER_DAYseconds, and does the day and
time-in-day split before looking at the timezone.
2. Adding month, quarter, and year silently ignores Int overflow during unit
conversion.
The root cause is here
[https://src.dev.databricks.com/databricks/runtime/-/blob/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/util/DateTimeUtils.scala?L1254].
quantity is multiplied by 3 or MONTHS_PER_YEARwithout checking overflow. Note
that we do have overflow checking in adding the amount to the timestamp, so the
behavior is inconsistent.
This can cause counter-intuitive results like this:
{{scala> spark.sql("select timestampadd(quarter, 1431655764,
timestamp'1970-01-01')").show
+------------------------------------------------------------------+
|timestampadd(quarter, 1431655764, TIMESTAMP '1970-01-01 00:00:00')|
+------------------------------------------------------------------+
| 1969-09-01 00:00:00|
+------------------------------------------------------------------+}}
3. Adding sub-month units (week, day, hour, minute, second, millisecond,
microsecond)silently ignores Long overflow during unit conversion.
This is similar to the previous problem:
{{scala> spark.sql("select timestampadd(day, 106751992,
timestamp'1970-01-01')").show(false)
+-------------------------------------------------------------+
|timestampadd(day, 106751992, TIMESTAMP '1970-01-01 00:00:00')|
+-------------------------------------------------------------+
|-290308-12-22 15:58:10.448384 |
+-------------------------------------------------------------+}}
> Several counter-intuitive behaviours in the TimestampAdd expression
> -------------------------------------------------------------------
>
> Key: SPARK-42635
> URL: https://issues.apache.org/jira/browse/SPARK-42635
> Project: Spark
> Issue Type: Bug
> Components: Spark Core, SQL
> Affects Versions: 3.3.0, 3.3.1, 3.3.2
> Reporter: Chenhao Li
> Priority: Major
>
> # When the time is close to daylight saving time transition, the result may
> be discontinuous and not monotonic.
> As pointed out by me and @Utkarsh Agarwal in
> [https://github.com/databricks/runtime/pull/54936/files#r1118047445], the
> result is conter-intuitive when the time is close to daylight saving time
> transition and the added amount is close to the multiple of days.
> We currently have:
>
> {{scala> spark.conf.set("spark.sql.session.timeZone", "America/Los_Angeles")
> scala> spark.sql("select timestampadd(second, 24 * 3600 - 1,
> timestamp'2011-03-12 03:00:00')").show
> +------------------------------------------------------------------------+
> |timestampadd(second, ((24 * 3600) - 1), TIMESTAMP '2011-03-12 03:00:00')|
> +------------------------------------------------------------------------+
> | 2011-03-13 03:59:59|
> +------------------------------------------------------------------------+
> scala> spark.sql("select timestampadd(second, 24 * 3600, timestamp'2011-03-12
> 03:00:00')").show
> +------------------------------------------------------------------+
> |timestampadd(second, (24 * 3600), TIMESTAMP '2011-03-12 03:00:00')|
> +------------------------------------------------------------------+
> | 2011-03-13 03:00:00|
> +------------------------------------------------------------------+}}
> In the second query, adding one more second will set the time back one hour
> instead. Plus, there is only 23 * 3600seconds from 2011-03-12 03:00:00 to
> 2011-03-13 03:00:00, instead of 24 * 3600 seconds, due to the daylight saving
> time transition.
> The root cause of the problem is the Spark code at
> [https://src.dev.databricks.com/databricks/runtime/-/blob/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/util/DateTimeUtils.scala?L797]
> wrongly assumes every day has MICROS_PER_DAYseconds, and does the day and
> time-in-day split before looking at the timezone.
> 2. Adding month, quarter, and year silently ignores Int overflow during unit
> conversion.
> The root cause is here
> [https://src.dev.databricks.com/databricks/runtime/-/blob/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/util/DateTimeUtils.scala?L1254].
> quantity is multiplied by 3 or MONTHS_PER_YEARwithout checking overflow.
> Note that we do have overflow checking in adding the amount to the timestamp,
> so the behavior is inconsistent.
> This can cause counter-intuitive results like this:
>
> {{scala> spark.sql("select timestampadd(quarter, 1431655764,
> timestamp'1970-01-01')").show
> +------------------------------------------------------------------+
> |timestampadd(quarter, 1431655764, TIMESTAMP '1970-01-01 00:00:00')|
> +------------------------------------------------------------------+
> | 1969-09-01 00:00:00|
> +------------------------------------------------------------------+}}
> 3. Adding sub-month units (week, day, hour, minute, second, millisecond,
> microsecond)silently ignores Long overflow during unit conversion.
> This is similar to the previous problem:
>
> {{scala> spark.sql("select timestampadd(day, 106751992,
> timestamp'1970-01-01')").show(false)
> +-------------------------------------------------------------+
> |timestampadd(day, 106751992, TIMESTAMP '1970-01-01 00:00:00')|
> +-------------------------------------------------------------+
> |-290308-12-22 15:58:10.448384 |
> +-------------------------------------------------------------+}}
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
