KSQL - 在 WINDOW TUMBLING 子句中更改时区

Question

Here my KSQL using WINDOW TUMBLING clause:

SELECT 
    sale_date,
    region,
    SUM(total)
FROM orders
WINDOW TUMBLING (SIZE 24 HOURS)
GROUP BY sale_date, region;

Some result:

2018-09-29|+|zskx_fz : Window{start=1538179200000 end=-} | 2018-09-29 | zskx_fz | 16119.8
2018-09-30|+|zskx_fz : Window{start=1538179200000 end=-} | 2018-09-30 | zskx_fz | 2031.6
2018-09-30|+|zskx_fz : Window{start=1538265600000 end=-} | 2018-09-30 | zskx_fz | 894.7

And the epoch millis to date time is:

1538179200000 = 2018-09-29 08:00:00 (UTC+8)
1538265600000 = 2018-09-30 08:00:00 (UTC+8)

As we can see, I'm in UTC+8. But regardless the time zone, start date time should be 2018-09-29 00:00:00 not 8 hours earlier. So it's able to change the time zone?

PS: I tried out several window size at 2018-09-30 11:33:00 and I totally lost..

WINDOW TUMBLING (SIZE 1 minutes)    2018-09-30 11:32:00
WINDOW TUMBLING (SIZE 2 hours)      2018-09-30 10:00:00
WINDOW TUMBLING (SIZE 5 hours)      2018-09-30 07:00:00
WINDOW TUMBLING (SIZE 10 hours)     2018-09-30 02:00:00
WINDOW TUMBLING (SIZE 11 hours)     2018-09-30 07:00:00
WINDOW TUMBLING (SIZE 12 hours)     2018-09-30 08:00:00
WINDOW TUMBLING (SIZE 24 hours)     2018-09-30 08:00:00

Answer 1

Timestamp windows are always calculated relative to the epoch, which is UTC/GMT.

I can see the validity of wanting to aggregate by day based on your timezone. I've raised it as an issue on the KSQL github project , and suggest you track it there.

Answer 2

If you are only using tumbling window you can consider time as just another dimension and perform aggregations over this dimension and not use any windowing at all. Here is an example. Let's consider the input stream schema is as the following:

<sale_date BIGINT, region VARCHAR, total DOUBLE>

Assuming the sale_date is the timestamp of the sale and our local time is PST, we can use the TIMESTAMPTOSTRING function to extract different time granularities for each sale for a given timezone as the following:

CREATE STREAM foo AS SELECT TIMESTAMPTOSTRING(sale_date, 'yyyy-MM-dd HH', 'PST') AS sale_hour, TIMESTAMPTOSTRING(sale_date, 'yyyy-MM-dd', 'PST') AS sale_day, TIMESTAMPTOSTRING(sale_date, 'yyyy-MM', 'PST') AS sale_month, region, total FROM orders; Now, you should be able to write your aggregate queries over this stream. For instance for daily sales for each region you can write the following query:

CRAETE TABLE daily_sale AS SELECT sale_day, region, sum(total) FROM foo GROUP BY sale_day, region;

Note that you don't need to specify a window for the above query.