Restore values in a time series table - sql

Question

I need to restore the most actual values from a time series table for the other values in specific time.
Let is say, that we have a table like that (I use SQL Server 2016), (this is pseudo-code, I did not check whether it works):
use sample
go
-- create time series table
drop table if exists dbo.PropertyHistory
go
create table dbo.PropertyHistory (
Id int
, Timestamp datetime
, Value int
)
go
-- fill dbo.PropertyHistory
insert into
dbo.PropertyHistory(Id, Timestamp, Value)
values
(1, '2019-01-01 12:00:00', 10)
, (1, '2019-01-01 13:00:00', 20)
, (2, '2019-01-01 13:00:00', 15)
, (3, '2019-01-01 14:00:00', 1)
, (4, '2019-01-01 15:00:00', 10)
, (1, '2019-01-01 16:00:00', 6)
, (4, '2019-01-01 17:00:00', 5)
, (2, '2019-01-01 17:00:00', 50)
, (2, '2019-01-01 19:00:00', 7)
, (1, '2019-01-01 19:00:00', 44)
go
I need to for example each row with the property id = 1 to have the last actual value (actual by datetime of course) of the property id = 2.
| Id | Timestamp | Value | Property2Value |
-------------------------------------------------------
| 1 | 2019-01-01 12:00:00 | 10 | NULL |
| 1 | 2019-01-01 13:00:00 | 20 | 15 |
| 1 | 2019-01-01 16:00:00 | 6 | 15 |
| 1 | 2019-01-01 19:00:00 | 44 | 7 |
-------------------------------------------------------
The ideas:
To create the function kinda of create function A (#propertyId int, #toDateTime datetime) which finds the latest row for the specified value restricted by the datetime. And then for each row with property id = 1 cross apply to this function. The performance is bad.
I think that it is possible to somehow use cumulative sum kinda of sum (case when PropertyId = 2 then Value else 0 end) over (order by Timestamp) but it will be cumulate more and more...
So, please help me to obtain the expected result.

Answer 1

If I understand correctly, this is a good use of apply:
select ph1.*, ph2.value as value2
from propertyhistory ph1 outer apply
(select top (1) ph2.*
from propertyhistory ph2
where ph2.id = 2 and ph2.timestamp <= ph1.timestamp
order by ph2.timestamp desc
) ph2
where ph1.id = 1;
Here is a db<>fiddle.
You can also do this with window functions, with the following logic:
For each row in the original data, get the most recent "2" timestamp.
Get the value for the "2" timestamp.
Filter down to just the "1"s
This looks like:
select ph.*
from (select ph.*,
max(case when ph.id = 2 then ph.value end) over (partition by timestamp_2) as value_2
from (select ph.*,
max(case when ph.id = 2 then ph.timestamp end) over (order by ph.timestamp) as timestamp_2
from propertyhistory ph
) ph
) ph
where id = 1;

Answer 2

We can handle this requirement by a judicious use of ROW_NUMBER, combined with some pivoting logic:
WITH cte AS (
SELECT *, ROW_NUMBER() OVER (PARTITION BY Id ORDER BY Timestamp DESC) rn
FROM dbo.PropertyHistory
)
SELECT
1 AS Id,
MAX(CASE WHEN Id = 1 THEN Timestamp END) AS Timestamp,
MAX(CASE WHEN Id = 1 THEN Value END) AS Value,
MAX(CASE WHEN Id = 2 THEN Value END) AS Property2Value
FROM cte
GROUP BY
rn
ORDER BY
MAX(CASE WHEN Id = 1 THEN Timestamp END);
Demo
The idea here is to compute a row number label for each record, numbered separately for each Id value. Then, we can aggregate by the row number, which brings the Id values from 1 and 2 into line, in a single record.