Re: Processor usage/tuning question - Mailing list pgsql-general

Thank you all for the advice. It looks like the load is due to a query that is taking around 1300ms to complete - a query that is run by every client connected (probably half a dozen or so, although I don't have specific numbers), every fifteen seconds or so. As you can imagine, that keeps the server rather busy :-) Specifically, it looks like the time is due to a sort (PARTITION BY tail ORDER BY pointtime DESC) that operates on around 100,000 rows.

The lovely details:

The query in question is the following:

SELECT *
FROM (SELECT tail, to_char(pointtime,'MM/DD/YYYY HH24:MI:SS'), lat,lng,altitude,heading,speed,source,pointtime,
      ROW_NUMBER() OVER (PARTITION BY tail ORDER BY pointtime DESC) as row
FROM data
WHERE tail in (<list of about 55 values or so>) and pointtime>='<timestamp of 24 hours prior to current UTC time>'::timestamp) s1
WHERE s1.row<=5
ORDER BY tail, pointtime DESC

In english, it boils down to get the five most recent data points for each listed tail number. I look at the last 24 hours of data because it is quite possible that a tail number may have no recent data points.

One obvious optimization is to look at a smaller time range. This will definitely speed up the query, but at the risk of not getting any data points for one or more of the requested tail numbers (there is already this risk, but looking back 24 hours keeps it fairly small for us).

The table description:
tracking=# \d data
                                     Table "public.data"
  Column   |            Type             |                     Modifiers
-----------+-----------------------------+---------------------------------------------------
 id        | bigint                      | not null default nextval('data_id_seq'::regclass)
 tail      | character varying(16)       | not null
 timerecp  | timestamp without time zone | not null default now()
 altitude  | integer                     |
 pointtime | timestamp without time zone |
 lat       | numeric(7,5)                | not null
 lng       | numeric(8,5)                | not null
 speed     | integer                     |
 heading   | integer                     |
 source    | character varying(64)       |
 syncd     | boolean                     | default false
Indexes:
    "data_pkey" PRIMARY KEY, btree (id)
    "pointtime_idx" btree (pointtime)
    "syncd_idx" btree (syncd)
    "tail_idx" btree (tail)
    "tailtime_idx" btree (tail, pointtime DESC)
    "timerecp_idx" btree (timerecp)

tracking=#

Adding the two-column sorted index didn't seem to affect the query time much.

The table current contains 1303951 rows, and any given 24 hour period has around 110,000 rows.

The results of the explain analyze command can be seen here: http://explain.depesz.com/s/H5w (nice site, btw. I'll have to be sure to bookmark it), where it clearly shows the the sort on data.tail,data.pointtime is the largest timesink (if I am reading it right).