Ordered count of consecutive repeats / duplicates

Question

I highly doubt I'm doing this in the most efficient manner, which is why I tagged plpgsql on here. I need to run this on 2 billion rows for a thousand measurement systems.

You have measurement systems that often report the previous value when they lose connectivity, and they lose connectivity for spurts often but sometimes for a long time. You need to aggregate but when you do so, you need to look at how long it was repeating and make various filters based on that information. Say you are measuring mpg on a car but it's stuck at 20 mpg for an hour than moves around to 20.1 and so on. You'll want to evaluate the accuracy when it's stuck. You could also place some alternative rules that look for when the car is on the highway, and with window functions you can generate the 'state' of the car and have something to group on. Without further ado:

--here's my data, you have different systems, the time of measurement, and the actual measurement
--as well, the raw data has whether or not it's a repeat (hense the included window function
select * into temporary table cumulative_repeat_calculator_data
FROM
    (
    select 
    system_measured, time_of_measurement, measurement, 
    case when 
     measurement = lag(measurement,1) over (partition by system_measured order by time_of_measurement asc) 
     then 1 else 0 end as repeat
    FROM
    (
    SELECT 5 as measurement, 1 as time_of_measurement, 1 as system_measured
    UNION
    SELECT 150 as measurement, 2 as time_of_measurement, 1 as system_measured
    UNION
    SELECT 5 as measurement, 3 as time_of_measurement, 1 as system_measured
    UNION
    SELECT 5 as measurement, 4 as time_of_measurement, 1 as system_measured
    UNION
    SELECT 5 as measurement, 1 as time_of_measurement, 2 as system_measured
    UNION
    SELECT 5 as measurement, 2 as time_of_measurement, 2 as system_measured
    UNION
    SELECT 5 as measurement, 3 as time_of_measurement, 2 as system_measured
    UNION
    SELECT 5 as measurement, 4 as time_of_measurement, 2 as system_measured
    UNION
    SELECT 150 as measurement, 5 as time_of_measurement, 2 as system_measured
    UNION
    SELECT 5 as measurement, 6 as time_of_measurement, 2 as system_measured
    UNION
    SELECT 5 as measurement, 7 as time_of_measurement, 2 as system_measured
    UNION
    SELECT 5 as measurement, 8 as time_of_measurement, 2 as system_measured
    ) as data
) as data;

--unfortunately you can't have window functions within window functions, so I had to break it down into subquery
--what we need is something to partion on, the 'state' of the system if you will, so I ran a running total of the nonrepeats
--this creates a row that stays the same when your data is repeating - aka something you can partition/group on
select * into temporary table cumulative_repeat_calculator_step_1
FROM
    (
    select 
    *,
    sum(case when repeat = 0 then 1 else 0 end) over (partition by system_measured order by time_of_measurement asc) as cumlative_sum_of_nonrepeats_by_system
    from cumulative_repeat_calculator_data
    order by system_measured, time_of_measurement
) as data;

--finally, the query. I didn't bother showing my desired output, because this (finally) got it
--I wanted a sequential count of repeats that restarts when it stops repeating, and starts with the first repeat
--what you can do now is take the average measurement under some condition based on how long it was repeating, for example  
select *, 
case when repeat = 0 then 0
else
row_number() over (partition by cumlative_sum_of_nonrepeats_by_system, system_measured order by time_of_measurement) - 1
end as ordered_repeat
from cumulative_repeat_calculator_step_1
order by system_measured, time_of_measurement

So, what would you do differently in order to run this on a huge table, or what alternative tools would you use? I'm thinking plpgsql because I suspect this needs to done in-database, or during the data insertion process, although I generally work with the data after it's loaded. Is there any way to get this in one sweep without resorting to sub-queries?

I have tested one alternative method, but it still relies on a sub-query and I think this is faster. For that method you create a "starts and stops" table with start_timestamp, end_timestamp, system. Then you join to the larger table and if the timestamp is between those, you classify it as being in that state, which is essentially an alternative to cumlative_sum_of_nonrepeats_by_system. But when you do this, you join on 1=1 for thousands of devices and thousands or millions of 'events'. Do you think that's a better way to go?

Answer 1

Test case

First, a more useful way to present your data - or even better, in an sqlfiddle, ready to play with:

CREATE TEMP TABLE data(
   system_measured int
 , time_of_measurement int
 , measurement int
);

INSERT INTO data VALUES
 (1, 1, 5)
,(1, 2, 150)
,(1, 3, 5)
,(1, 4, 5)
,(2, 1, 5)
,(2, 2, 5)
,(2, 3, 5)
,(2, 4, 5)
,(2, 5, 150)
,(2, 6, 5)
,(2, 7, 5)
,(2, 8, 5);

Simplified query

Since it remains unclear, I am assuming only the above as given.
Next, I simplified your query to arrive at:

WITH x AS (
   SELECT *, CASE WHEN lag(measurement) OVER (PARTITION BY system_measured
                               ORDER BY time_of_measurement) = measurement
                  THEN 0 ELSE 1 END AS step
   FROM   data
   )
   , y AS (
   SELECT *, sum(step) OVER(PARTITION BY system_measured
                            ORDER BY time_of_measurement) AS grp
   FROM   x
   )
SELECT * ,row_number() OVER (PARTITION BY system_measured, grp
                             ORDER BY time_of_measurement) - 1 AS repeat_ct
FROM   y
ORDER  BY system_measured, time_of_measurement;

Now, while it is all nice and shiny to use pure SQL, this will be much faster with a plpgsql function, because it can do it in a single table scan where this query needs at least three scans.

Faster with plpgsql function:

CREATE OR REPLACE FUNCTION x.f_repeat_ct()
  RETURNS TABLE (
    system_measured int
  , time_of_measurement int
  , measurement int, repeat_ct int
  )  LANGUAGE plpgsql AS
$func$
DECLARE
   r    data;     -- table name serves as record type
   r0   data;
BEGIN

-- SET LOCAL work_mem = '1000 MB';  -- uncomment an adapt if needed, see below!

repeat_ct := 0;   -- init

FOR r IN
   SELECT * FROM data d ORDER BY d.system_measured, d.time_of_measurement
LOOP
   IF  r.system_measured = r0.system_measured
       AND r.measurement = r0.measurement THEN
      repeat_ct := repeat_ct + 1;   -- start new array
   ELSE
      repeat_ct := 0;               -- start new count
   END IF;

   RETURN QUERY SELECT r.*, repeat_ct;

   r0 := r;                         -- remember last row
END LOOP;

END
$func$;

Call:

SELECT * FROM x.f_repeat_ct();

Be sure to table-qualify your column names at all times in this kind of plpgsql function, because we use the same names as output parameters which would take precedence if not qualified.

Billions of rows

If you have billions of rows, you may want to split this operation up. I quote the manual here:

Note: The current implementation of RETURN NEXT and RETURN QUERY stores the entire result set before returning from the function, as discussed above. That means that if a PL/pgSQL function produces a very large result set, performance might be poor: data will be written to disk to avoid memory exhaustion, but the function itself will not return until the entire result set has been generated. A future version of PL/pgSQL might allow users to define set-returning functions that do not have this limitation. Currently, the point at which data begins being written to disk is controlled by the work_mem configuration variable. Administrators who have sufficient memory to store larger result sets in memory should consider increasing this parameter.

Consider computing rows for one system at a time or set a high enough value for work_mem to cope with the load. Follow the link provided in the quote on more about work_mem.

One way would be to set a very high value for work_mem with SET LOCAL in your function, which is only effective for for the current transaction. I added a commented line in the function. Do not set it very high globally, as this could nuke your server. Read the manual.