Snowflake Time-Series Weighted Values - Stack Overflow

admin管理员组

文章数量:1388067

I have data from a PLC coming in ONLY ON CHANGE into table format (TIMESTAMP, TAG, VALUE). I have a visualisation tool (seeq) that queries this base table in snowflake and shows the data on a time series chart. If a user selects a long time-range, then this data will need to be aggregated (max 2000 points per time series plot). I want this aggregation (average) to be weighted for how long a tag has been on that value before its change. For example if I have a tag = 'cheese' and t=0 -> t=5 has a value of 100, then t=6 -> t=100 has a value of 500. If the user in seeq selects this tag, on a long period window (i.e. spans from t=0 to t=100000), the data registered from this tag has to be aggregated to (5100+95500)/100 for t=50 (mid point) on the plot. How to generate a query for this in snowflake using this base table format of TIMESTAMP, TAG, VALUE.

Tried doing a cross join to a time dimension table created from a tag dimension table then using a lead function to get every single second output from the raw data spread across every time and then weight it accordingly. It was not very performant in terms of speed.

I have data from a PLC coming in ONLY ON CHANGE into table format (TIMESTAMP, TAG, VALUE). I have a visualisation tool (seeq) that queries this base table in snowflake and shows the data on a time series chart. If a user selects a long time-range, then this data will need to be aggregated (max 2000 points per time series plot). I want this aggregation (average) to be weighted for how long a tag has been on that value before its change. For example if I have a tag = 'cheese' and t=0 -> t=5 has a value of 100, then t=6 -> t=100 has a value of 500. If the user in seeq selects this tag, on a long period window (i.e. spans from t=0 to t=100000), the data registered from this tag has to be aggregated to (5100+95500)/100 for t=50 (mid point) on the plot. How to generate a query for this in snowflake using this base table format of TIMESTAMP, TAG, VALUE.

Tried doing a cross join to a time dimension table created from a tag dimension table then using a lead function to get every single second output from the raw data spread across every time and then weight it accordingly. It was not very performant in terms of speed.

• time-series
• snowflake-cloud-data-platform

Share Improve this question Follow asked Mar 17 at 22:15 AustinAustin 1 1

• something like NTILE docs.snowflake/en/sql-reference/functions/ntile might be a way to chunk the data, and then do some average/weighted operation. – Simeon Pilgrim Commented Mar 17 at 22:15

Add a comment  | 

1 Answer 1

Sorted by: Reset to default 0

so I am not really sure what you are trying todo, but an explosive way of doing something like what you describe can be done like so:

with d0 as (
    select * from values
     ('cheese', 0, 5, 100),
     ('cheese', 6, 100, 500)
     t(tag, _s, _e, val)
 ), d1 as (
     select 
        tag,
        value::number rn,
        val,
        ntile(10) over (partition by tag order by rn) as tile,
     from d0, 
        table(flatten(array_generate_range(_s, _e+1)))
)
select
    tag,
    tile,
    avg(rn) as mid,
    avg(val) as val
from d1
group by 1,2
order by 1,2;

which gives:

TAG	TILE	MID	VAL
cheese	1	5.000000	281.818182
cheese	2	15.500000	500.000000
cheese	3	25.500000	500.000000
cheese	4	35.500000	500.000000
cheese	5	45.500000	500.000000
cheese	6	55.500000	500.000000
cheese	7	65.500000	500.000000
cheese	8	75.500000	500.000000
cheese	9	85.500000	500.000000
cheese	10	95.500000	500.000000

those rows do not really need expanding, and the interpolation can be driven against a d0 like table, if that is how your data is sourced..

本文标签： Snowflake TimeSeries Weighted ValuesStack Overflow