SQL select max of a consecutive run of data - sql

Question

Given a table with a consecutive run of data: a number that always increases while a task is in progress and resets back to zero when the next task starts, how do you select the maximum of each run of data?
Each consecutive run can have any number of rows, and the runs of data are marked by a a "start" and "end" row, eg the data might look like
user_id, action, qty, datetime
1, start, 0, 2017-01-01 00:00:01
1, record, 0, 2017-01-01 00:00:01
1, record, 4, 2017-01-01 00:00:02
1, record, 5, 2017-01-01 00:00:03
1, record, 6, 2017-01-01 00:00:04
1, end, 0, 2017-01-01 00:00:04
1, start, 0, 2017-01-01 00:00:05
1, record, 0, 2017-01-01 00:00:05
1, record, 2, 2017-01-01 00:00:06
1, record, 3, 2017-01-01 00:00:07
1, end, 0, 2017-01-01 00:00:07
2, start, 0, 2017-01-01 00:00:08
2, record, 0, 2017-01-01 00:00:08
2, record, 3, 2017-01-01 00:00:09
2, record, 8, 2017-01-01 00:00:10
2, end, 0, 2017-01-01 00:00:10
And the results would be the maximum value of each run:
user_id, action, qty, datetime
1, record, 6, 2017-01-01 00:00:04
1, record, 3, 2017-01-01 00:00:07
2, record, 8, 2017-01-01 00:00:10
Using any postgres sql syntax (9.3)? Its some kind of grouping then selecting max from each group, but I don't see how to do the grouping part.

Answer 1

If theres no overlapping for a single user and the next run always starts at a later time, then you can use LAG() window function.
with the_table(user_id, action, qty, datetime) as (
select 1,'start', 0, '2017-01-01 00:00:01'::timestamp union all
select 1,'record', 0, '2017-01-01 00:00:01'::timestamp union all
select 1,'record', 4, '2017-01-01 00:00:02'::timestamp union all
select 1,'record', 5, '2017-01-01 00:00:03'::timestamp union all
select 1,'record', 6, '2017-01-01 00:00:04'::timestamp union all
select 1,'end', 0, '2017-01-01 00:00:04'::timestamp union all
select 1,'start', 0, '2017-01-01 00:00:05'::timestamp union all
select 1,'record', 0, '2017-01-01 00:00:05'::timestamp union all
select 1,'record', 2, '2017-01-01 00:00:06'::timestamp union all
select 1,'record', 3, '2017-01-01 00:00:07'::timestamp union all
select 1,'end', 0, '2017-01-01 00:00:07'::timestamp union all
select 2,'start', 0, '2017-01-01 00:00:08'::timestamp union all
select 2,'record', 0, '2017-01-01 00:00:08'::timestamp union all
select 2,'record', 3, '2017-01-01 00:00:09'::timestamp union all
select 2,'record', 8, '2017-01-01 00:00:10'::timestamp union all
select 2,'end', 0, '2017-01-01 00:00:10'::timestamp
)
select n_user_id, n_action, n_qty, n_datetime from (
select action,
lag(user_id) over(partition by user_id order by datetime, case when action = 'start' then 0 when action = 'record' then 1 else 2 end, qty) as n_user_id,
lag(action) over(partition by user_id order by datetime, case when action = 'start' then 0 when action = 'record' then 1 else 2 end, qty) as n_action,
lag(qty) over(partition by user_id order by datetime, case when action = 'start' then 0 when action = 'record' then 1 else 2 end, qty) as n_qty,
lag(datetime) over(partition by user_id order by datetime, case when action = 'start' then 0 when action = 'record' then 1 else 2 end, qty) as n_datetime
from the_table
)t
where action = 'end'
Because some action = record rows have same datetime as start and end rows, I use CASE in ORDER BY, to be clear that start is first, then is record and then end.

Answer 2

Quick and dirty, assuming runs do not overlap
with bounds as (select starts.rn, starts.datetime as s, ends.datetime as e from
(select datetime,ROW_NUMBER() OVER () as rn from runs where action = 'start' order by datetime) as starts
join
(select datetime,ROW_NUMBER() OVER () as rn from runs where action = 'end' order by datetime) as ends
on starts.rn = ends.rn)
,with_run as (SELECT *, (select rn from bounds where s <= r.datetime and e >= r.datetime) as run
from runs as r)
,max_qty as (
SELECT run,max(qty) as qty
from with_run
GROUP BY run)
SELECT s.user_id,s.action,s.qty,s.datetime from with_run as s join max_qty as f on s.run = f.run AND s.qty = f.qty;
-- TEST DATA --
create table runs (user_id int, action text, qty int, datetime TIMESTAMP);
insert INTO runs VALUES
(1, 'start', 0, '2017-01-01 00:00:01')
,(1, 'record', 0, '2017-01-01 00:00:01')
,(1, 'record', 4, '2017-01-01 00:00:02')
,(1, 'record', 5, '2017-01-01 00:00:03')
,(1, 'record', 6, '2017-01-01 00:00:04')
,(1, 'end', 0, '2017-01-01 00:00:04')
,(1, 'start', 0, '2017-01-01 00:00:05')
,(1, 'record', 0, '2017-01-01 00:00:05')
,(1, 'record', 2, '2017-01-01 00:00:06')
,(1, 'record', 3, '2017-01-01 00:00:07')
,(1, 'end', 0, '2017-01-01 00:00:07')
,(2, 'start', 0, '2017-01-01 00:00:08')
,(2, 'record', 0, '2017-01-01 00:00:08')
,(2, 'record', 3, '2017-01-01 00:00:09')
,(2, 'record', 8, '2017-01-01 00:00:10')
,(2, 'end', 0, '2017-01-01 00:00:10');
UPDATE
#Oto Shavadze answer can be shortened
with lookup as (select action,lag(t.*) over(order by datetime, case when action = 'start' then 0 when action = 'record' then 1 else 2 end) as r from runs t)
select (r::runs).user_id
,(r::runs).action
,(r::runs).qty
,(r::runs).datetime
from lookup where action = 'end';
I think OP unclear about what considers maximum, last record before end or highest qty in run.