SQL Pattern Length in New Column - sql

Question

I have a (very very large) table of similar format to the following:
+--------+-------+
| id | value |
+--------+-------+
| 1 | 5 |
| 2 | 6 |
| 3 | 6 |
| 4 | 4 |
| 5 | 3 |
| 6 | 2 |
| 7 | 4 |
| 8 | 5 |
+--------+-------+
What I'd like to be able to do is return the pattern length of the value column increasing or decreasing in a third column (with pattern being negative for decreasing and positive for increasing), while ignoring IDs where there is no change. The pattern should reset to 1 or -1 when the pattern is broken.
I've not explained that well at all, so with the table above, ideally the result would be:
+--------+-------+---------+
| id | value | pattern |
+--------+-------+---------+
| 1 | 5 | 0/NULL |
| 2 | 6 | 1 |
| 3 | 6 | 1 |
| 4 | 4 | -1 |
| 5 | 3 | -2 |
| 6 | 2 | -3 |
| 7 | 4 | 1 |
| 8 | 5 | 2 |
+--------+-------+---------+
I did some research and came across pattern matching, but it turns out either the version of SQL I'm using (it's the version used by/on Amazon Redshift , which according to them is 'based on' PostgreSQL 8.0.2 http://docs.aws.amazon.com/redshift/latest/dg/c_redshift-and-postgres-sql.html)) doesn't support it, or I'm being very silly.
So, is this something that is even possible with SQL, and if so how should I go about it? Many thanks.

Answer 1

In SQL Server 2012, you can do this with lead() and lag() and cumulative sum.
Something that comes quite close is this:
select t.*, sum(nextinc) over (order by id) as pattern
from (select t.*,
(case when lead(t.value) > t.value then 1
when lead(t.value) = t.value then 0
else -1 end) as nextinc,
(case when lag(t.value) > t.value then 1 else 0 end) as previnc
from table t
) t;
However, the pattern goes up and down in increments of 1 instead of starting over. So, we need to find the pattern breaks. The following defines the breaks in the pattern and then increments pattern for for sequences of increasing/decreasing values:
select t.*,
sum(nextinc) over (partition by grp order by id) as pattern
from (select t.*,
sum(case when (prev_value <= value and value <= next_value) or
(prev_value >= value and value >= next_value)
then 0 else 1
end) over (order by id) as grp
from (select t.*, lead(t.value) over (order by id) as next_value,
lag(t.value) over (order by id) as prev_value,
(case when lead(t.value) over (order by id) > t.value then 1
when lead(t.value) over (order by id) = t.value then 0
else -1 end) as nextinc
from table t
) t
) t

Answer 2

For the given example, the following seems to do the job:
SELECT
S3.id
, S3.value
, S3.pattern
, SUM(minusNullPlus) OVER (PARTITION BY sequenceID ORDER BY id) calculated
FROM
(SELECT
S2.*
, SUM(newSequence) OVER (ORDER BY id) sequenceID
FROM
(SELECT
S1.*
, CASE
WHEN minusNullPlus = LAG(minusNullPlus, 1, NULL) OVER (ORDER BY id)
OR
minusNullPlus = 0
OR
(minusNullPlus = 1
AND
value - LAG(value, 1, NULL) OVER (ORDER BY id) = 1
)
OR
(minusNullPlus = -1
AND
value - LAG(value, 1, NULL) OVER (ORDER BY id) = -1
)
THEN 0
ELSE 1
END newSequence
FROM
(SELECT
id
, value
, CASE
WHEN value > LAG(value, 1, NULL) OVER (ORDER BY id) THEN 1
WHEN value < LAG(value, 1, NULL) OVER (ORDER BY id) THEN -1
WHEN value = LAG(value, 1, NULL) OVER (ORDER BY id) THEN 0
ELSE 0
END minusNullPlus
, CASE
WHEN value - LAG(value, 1, NULL) OVER (ORDER BY id) = 0 THEN 0
ELSE 1
END change
, pattern
FROM SomeTable
) S1
) S2
) S3
ORDER BY id
;
See it in action: SQL Fiddle
It uses some additional data to check against - please verify the respective patterns to be actually in line with your expectations/requirements.
NB: The suggested solution relies on some of the particularities of the provided sample data (and its expansion in above SQL Fiddle).
Please comment, if and as adjustment / further detail is required.