[
https://issues.apache.org/jira/browse/DRILL-5360?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15929136#comment-15929136
]
Paul Rogers commented on DRILL-5360:
------------------------------------
Talking with other engineers, it seems that it never was Drill's intent for
Timestamp to be UTC. Instead, it tried to follow the SQL standard which is to
have relative times (Date, Time) with no implied TZ, and the newer "TZ" types
that have a date/time with a TZ offset.
The {{Timestamp}} type avoids the need for a timezone by declaring that all
values are in the same timezone. On Unix (and the internet and other modern
systems), timestamps are always UTC and are often relative to the Unix epoch.
Drill, however, did not implement the TZ-based date/time type. So, the
assumption is that {{Timestamp}} can't have a timezone. Since it has no
timezone, it should be local time.
This misunderstands how Unix timezones work, but seems to be closer to how
databases worked (before they added time zones and absolute time.) So, the
Drill {{Timestamp}} is the worst of worlds: it tries to be absolute, but is
actually local, and is converted wrong in JDBC.
Perhaps the best solution is to implement either the standard SQL TZ-based
types *or* implement a non-standard {{TimestampUTC}} value that is always in
UTC. (When times are in UTC, they have a time zone, but that zone is fixed and
so need not appear in the value vector itself. This is distinct from relative
dates where the time zone is not specified but is also not predefined.)
> Timestamp type documented as UTC, implemented as local time
> -----------------------------------------------------------
>
> Key: DRILL-5360
> URL: https://issues.apache.org/jira/browse/DRILL-5360
> Project: Apache Drill
> Issue Type: Bug
> Affects Versions: 1.10.0
> Reporter: Paul Rogers
>
> The Drill documentation implies that the {{Timestamp}} type is in UTC:
> bq. 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:
> bq. 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:
> {code}
> 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);
> {code}
> 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:
> {code}
> Hours: -6.9999975
> {code}
> 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.)
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
