SQL get n results from each set with defined deviation - sql

Question

Using Oracle 11g and having a table like:
USER | TIME
----- | --------
User1 | 08:15:50
User1 | 10:42:22
User1 | 10:42:24
User1 | 10:42:35
User1 | 10:50:01
User2 | 13:23:05
User2 | 13:23:34
User2 | 13:24:01
User2 | 13:24:02
For each user I need to get (if available) exactly 3 records with deviation between first and last less than a minute. If rows are more than 3 they won't match the criteria. Could you give me some clue?
The result should look like:
User1 | 10:42:22
User1 | 10:42:24
User1 | 10:42:35

Answer 1

Here's my stab at this. I don't have a live Oracle and SQLFiddle isn't working, so please advise how it turns out:
CREATE TABLE t (
u VARCHAR(5),
t DATETIME
);
INSERT INTO t
(u, t)
VALUES
('User1', '2001-01-01 08:15:50'),
('User1', '2001-01-01 10:42:22'),
('User1', '2001-01-01 10:42:24'),
('User1', '2001-01-01 10:42:35'),
('User1', '2001-01-01 10:50:01'),
('User2', '2001-01-01 13:23:05'),
('User2', '2001-01-01 13:23:34'),
('User2', '2001-01-01 13:24:01'),
('User2', '2001-01-01 13:24:02');
SELECT
z.u,
min(z.t) evt_start,
max(z.t) evt_end
FROM
(
SELECT y.*, SUM(prev_or_2prev_not_within) OVER(PARTITION BY u ORDER BY t ROWS UNBOUNDED PRECEDING) as ctr
FROM
(
SELECT
t.*,
CASE WHEN
t - LAG(t) OVER(PARTITION BY u ORDER BY t) < 1.0/1440.0 OR
t - LAG(t, 2) OVER(PARTITION BY u ORDER BY t) < 1.0/1440.0
THEN 0 ELSE 1
END as prev_or_2prev_not_within
FROM
t
) y
) z
GROUP BY
z.u,
z.ctr
HAVING COUNT(*) = 3
I believe it will establish an incrementing counter that doesn't increment when the previous or previousprevious row is within a minute of the current row. It does this by classing rows as 0 or 1, and when 0 occurs the sum-all-preceding-rows operation generates a counter that doesn't change. It then groups on this counter having exactly 3 occurrences. The partition makes the counter work per user
You can see it in action here: https://dbfiddle.uk/?rdbms=sqlserver_2017&fiddle=018125210ecd071f3d11e3d4b3d3e670
It's SQL Server (as noted, I don't have an oracle) but the terms used for sqlserver and the logic should be broadly similar for oracle - oracle supports lag, unbounded sums, having etc, and it does date math in terms of dateA - dateB -> a floating point number representative of whole or parts of a day (and 1440 minutes per day, 1/1440 should represent a float of one minute). The data types sqlserver uses might differ slightly to oracle, and this query does depend on TIME (I called it t - dislike column names that are reserved words/keywords) column being a datetime, not a string that looks like a time. If your data is a string, sort it out so it isn't (use an inner subquery to generate a datetime, or change your data storage so it's stored as a datetime type)
You said you wanted a result that tells the user and the event time - the simplest way to do that was to use min and max to give you the date range. If you're desperate to have all 3 rows on show, you can join the output of this query back to the table with date between evt_start and evt_end, or you can use some sort of string_aggregate type function to give you a list of times straight out of the outermost group operation

Answer 2

I would use analytical count() with range clause:
SQL Fiddle demo
select user_, to_char(time_, 'hh24:mi:ss') time_
from (
select user_, time_,
count(1) over (partition by user_ order by time_
range between interval '1' minute preceding
and interval '1' minute following) cnt
from (select user_, to_date(time_, 'hh24:mi:ss') time_ from tbl))
where cnt = 3
Result:
USER_ TIME_
----- --------
User1 10:42:22
User1 10:42:24
User1 10:42:35
Edit:
As #CaiusJard noticed first answer may show incorrect values when there are intervals like 10:52:01, 10:53:00, 10:53:59. There are some ways to correct this. First is to find min and max time in group and check condition numtodsinterval( max - min, 'day') <= interval '1' minute. Second is to number all rows, then assign flag to these row where prior, current and leading count = 3.
Finally show flagged rows joined with original table:
with t as (
select row_number() over (order by user_, time_) rn, tbl.*,
count(1) over (partition by user_ order by time_
range between interval '1' minute preceding
and interval '1' minute following) cnt
from (select user_, to_date(time_, 'hh24:mi:ss') time_ from tbl) tbl),
r as (select rn,
case when 3 = lag(cnt) over (partition by user_ order by time_)
and 3 = cnt
and 3 = lead(cnt) over (partition by user_ order by time_)
then 1
end flag
from t )
select * from t
join (select rn-1 r1, rn r2, rn+1 r3 from r where flag = 1) r
on rn in (r1, r2, r3)