Speeding up SQL Server cross apply to get aggregated data - sql

Question

In SQL Server, I am trying to put together a single query which grabs a row and includes the aggregated data from a two hour window before that row as well as aggregated data from one hour window after. How can I make this run faster?
The rows have time stamps to a millisecond precision, and are not evenly spaced. I have over 50 million rows in this table, and the query does not seem to be completing. There are indexes in many places, but they don't seem to help. I was also thinking about using a window function, but I am not sure that its possible to have a sliding window with unevenly distributed rows. Also, for the future one hour window, I am not sure how that would be done with a SQL window.
Box is a string and has 10 unique values.
Process is a string and has 30 unique values.
The average duration_ms is 200 ms.
Errors account for less than 0.1% of the data.
The 50 million rows describes a years worth of data.
select
c1.start_time,
c1.end_time,
c1.box,
c1.process,
datediff(ms,c1.start_time,c1.end_time) as duration_ms,
datepart(dw,c1.start_time) as day_of_week,
datepart(hour,c1.start_time) as hour_of_day,
c3.*,
c5.*
from metrics_table c1
cross apply
(select
avg(cast(datediff(ms,c2.start_time,c2.end_time) as numeric)) as avg_ms,
count(1) as num_process_total,
count(distinct process) as num_process_unique,
count(distinct box) as num_box_unique
from metrics_table c2
where datediff(minute,c2.start_time,c1.start_time) <= 120
and c1.start_time> c2.start_time
and c2.error_code = 0
) c3
cross apply
(select
avg(case when datediff(ms,c4.start_time,c4.end_time)>1000 then 1.0 else 0.0 end) as percent_over_thresh
from metrics_table c4
where datediff(hour,c1.start_time,c4.start_time) <= 1
and c4.start_time> c1.start_time
and c4.error_code= 0
) c5
where
c1.error_code= 0
Edit
Version: SQL Azure 12.0
Adding execution plan:

Answer 1

The following should be a step in the right direction...
Note: c2.start_time & c4.start_time are no longer wrappen in DATEDIFF functions making them SARGable...
SELECT
c1.start_time,
c1.end_time,
c1.box,
c1.process,
DATEDIFF(ms, c1.start_time, c1.end_time) AS duration_ms,
DATEPART(dw, c1.start_time) AS day_of_week,
DATEPART(HOUR, c1.start_time) AS hour_of_day,
--c3.*,
avg_ms = CASE WHEN
c5.*
FROM
dbo.metrics_table c1
CROSS APPLY (
SELECT
AVG(CAST(DATEDIFF(ms, c2.start_time, c2.end_time) AS NUMERIC)) AS avg_ms,
COUNT(1) AS num_process_total,
COUNT(DISTINCT process) AS num_process_unique,
COUNT(DISTINCT box) AS num_box_unique
FROM
dbo.metrics_table c2
WHERE
--DATEDIFF(minute,c2.start_time,c1.start_time) <= 120
c2.start_time <= DATEADD(MINUTE, -120, c1.start_time)
--and c1.start_time> c2.start_time
AND c2.error_code = 0
) c3
CROSS APPLY (
SELECT
AVG(CASE WHEN DATEDIFF(ms, c4.start_time, c4.end_time) > 1000 THEN 1.0 ELSE 0.0 END
) AS percent_over_thresh
FROM
dbo.metrics_table c4
WHERE
--DATEDIFF(HOUR, c1.start_time, c4.start_time) <= 1
c4.start_time >= DATEADD(HOUR, 1, c1.start_time)
--and c4.start_time> c1.start_time
AND c4.error_code = 0
) c5
WHERE
c1.error_code = 0;
Of course, making a query SARGable doesn't do any good unless there's an appropriate index available. The following should be good for all 3 metrics_table references... (see what indexes are currently available, there's a chance that you may not need to create a new index)
CREATE NONCLUSTERED INDEX ixf_metricstable_errorcode_starttime ON dbo.metrics_table (
error_code,
start_time
)
INCLUDE (
end_time,
box,
process
)
WHERE
error_code = 0;

Answer 2

I used Between and got good performance in my simple test rig. I've also used columnstore as 50 million records is DW volumes:
CREATE TABLE dbo.metrics_table (
rowId INT IDENTITY,
start_time DATETIME NOT NULL,
end_time DATETIME NOT NULL,
box VARCHAR(10) NOT NULL,
process VARCHAR(10) NOT NULL,
error_code INT NOT NULL
);
-- Add records
;WITH cte AS (
SELECT TOP 3334 ROW_NUMBER() OVER ( ORDER BY ( SELECT 1 ) ) rn
FROM sys.columns c1
CROSS JOIN sys.columns c2
CROSS JOIN sys.columns c3
)
INSERT INTO dbo.metrics_table ( start_time, end_time, box, process, error_code )
SELECT
DATEADD( ms, rn, DATEADD( day, rn % 365, '1 Jan 2017' ) ) AS start_time,
DATEADD( ms, rn % 409, DATEADD( ms, rn, DATEADD( day, rn % 365, '1 Jan 2017' ) ) ) AS end_time,
'box' + CAST( boxes.box AS VARCHAR(10) ) box,
'process' + CAST( boxes.box AS VARCHAR(10) ) process,
ABS( CAST( rn % 3000 AS BIT ) -1 ) error_code
FROM cte c
CROSS JOIN ( SELECT TOP 10 rn FROM cte ) AS boxes(box)
CROSS JOIN ( SELECT TOP 30 rn FROM cte ) AS processes(process);
-- Create normal clustered index to order the data
CREATE CLUSTERED INDEX cci_metrics_table ON dbo.metrics_table ( start_time, end_time, box, process );
--CREATE CLUSTERED INDEX cci_metrics_table ON dbo.metrics_table ( box, process, start_time, end_time );
-- Convert to columnstore
CREATE CLUSTERED COLUMNSTORE INDEX cci_metrics_table ON dbo.metrics_table WITH ( MAXDOP = 1, DROP_EXISTING = ON );
IF OBJECT_ID('tempdb..#tmp1' ) IS NOT NULL DROP TABLE #tmp1
-- two hour window before, 1 hour window after
SELECT
c1.start_time,
c1.end_time,
c1.box,
c1.process,
DATEDIFF( ms, c1.start_time, c1.end_time ) AS duration_ms,
DATEPART( dw, c1.start_time ) AS day_of_week,
DATEPART( hour, c1.start_time ) AS hour_of_day,
c2.xavg,
c2.num_process_total,
c2.num_process_unique,
c2.num_box_unique,
c3.percent_over_thresh
INTO #tmp1
FROM dbo.metrics_table c1
CROSS APPLY
(
SELECT
COUNT(1) AS num_process_total,
AVG( CAST( DATEDIFF( ms, start_time, end_time ) AS NUMERIC ) ) xavg,
COUNT( DISTINCT process ) num_process_unique,
COUNT( DISTINCT box ) num_box_unique
FROM dbo.metrics_table c2
WHERE c2.error_code = 0
AND c2.start_time Between DATEADD( minute, -120, c1.start_time ) And c1.start_time
AND c1.start_time > c2.start_time
) c2
CROSS APPLY
(
SELECT
AVG( CASE WHEN DATEDIFF( ms, c4.start_time, c4.end_time ) > 1000 THEN 1.0 ELSE 0.0 END ) percent_over_thresh
FROM dbo.metrics_table c4
WHERE c4.error_code = 0
AND c4.start_time Between c1.start_time And DATEADD( minute, 60, c1.start_time )
AND c4.start_time > c1.start_time
) c3
WHERE error_code = 0