[DRILL-5360] Timestamp type documented as UTC, implemented as local time

• Type: Improvement
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Affects Version/s: 1.10.0
• Assignee: Unassigned
• Reporter: Paul Rogers
• Created: 16/Mar/17 21:47
• Updated: 15/Feb/23 06:32

Description

The Drill documentation implies that the Timestamp type is in UTC:

JDBC timestamp in year, month, date hour, minute, second, and optional milliseconds format: yyyy-MM-dd HH:mm:ss.SSS. ... TIMESTAMP literals: Drill stores values in Coordinated Universal Time (UTC). Drill supports time functions in the range 1971 to 2037. ... Drill does not support TIMESTAMP with time zone.

The above is ambiguous. The first part talks about JDBC timestamps. From the JDK Javadoc:

Timestamp: A thin wrapper around java.util.Date. ... Date class is intended to reflect coordinated universal time (UTC)...

So, a JDBC timestamp is intended to represent time in UTC. (The "indented to reflect" statement leaves open the possibility of misusing Date to represent times in other time zones. This was common practice in early Java development and was the reason for the eventual development of the Joda, then Java 8 date/time classes.)

The Drill documentation implies that timestamp literals are in UTC, but a careful read of the documentation does allow an interpretation that the internal representation can be other than UTC. If this is true, then we would also rely on a liberal reading of the Java `Timestamp` class to also not be UTC. (Or, we rely on the Drill JDBC driver to convert from the (unknown) server time zone to a UTC value returned by the Drill JDBC client.)

Still, a superficial reading (and common practice) would suggest that a Drill Timestamp should be in UTC.

However, a test on a Mac, with an embedded Drillbit (run in the Pacific time zone, with Daylight Savings Time in effect) shows that the Timestamp binary value is actual local time:

      long before = System.currentTimeMillis();
      long value = getDateValue(client, "SELECT NOW() FROM (VALUES(1))" );
      double hrsDiff = (value - before) / (1000.00 * 60 * 60);
      System.out.println("Hours: " + hrsDiff);

The above gets the actual UTC time from Java. Then, it runs a query that gets Drill's idea of the current time using the NOW() function. (The getDateValue function uses the new test framework to access the actual long value from the returned value vector.) Finally, we compute the difference between the two times, converted to hours. Output:

Hours: -6.9999975

As it turns out, this is the difference between UTC and PDT. So, the time is in local time, not UTC.

Since the documentation and implementation are both ambiguous, it is hard to know the intent of the Drill Timestamp. Clearly, common practice is to use UTC. But, there is wiggle-room.

If the Timestamp value is supposed to be local time, then Drill should provide a function to return the server's time zone offset (in ms) from UTC so that the client can to the needed local-to-UTC conversion to get a true timestamp.

On the other hand, if the Timestamp is supposed to be UTC (per common practice), then NOW() should not report local time, it should return UTC.

Further, if NOW() returns local time, but Timestamp literals are UTC, then it is hard to see how any query can be rationally written if one timestamp value is local, but a literal is UTC.

So, job #1 is to define the Timestamp semantics. Then, use that to figure out where the bug lies to make implementation consistent with documentation (or visa-versa.)

Issue Links

• depends upon: DRILL-6242 Output format for nested date, time, timestamp values in an object hierarchy
• links to: Impala Proposal