Re: Improve PostGIS performance with 62 million rows?

[Israel Brewster <israel@xxxxxxxxxxxxxx>]

So just for interests sake, to kick things up a notch (and out of sheer morbid curiosity), I loaded a higher-resolution dataset (Elevation data for the state of Alaska, 2 arc second resolution, as opposed to 100 meter resolution before). Same structure/indexes and everything, just higher resolution. So the new database has 1,642,700,002 rows, and is somewhere around 300GB in size (including index). Due to the larger data size, I moved the database to a different table space which resides on a mirrored 2TB spinning platter disk (i.e. slower both because of the RAID and lack of SSD). Friday evening I ran the following query:

EXPLAIN ANALYZE WITH segments AS (

    SELECT ST_MakeLine( lag((pt).geom , 1, NULL) OVER (ORDER BY (pt).path)

                          ,(pt).geom)::GEOGRAPHY AS short_line

    FROM ST_DumpPoints(

          ST_Segmentize(

            ST_GeographyFromText('SRID=4326;LINESTRING(-150.008056 61.179167,-156.77 71.285833)'),

            5000

        )::geometry

    ) as pt

)

SELECT elevation

FROM data ,segments

WHERE segments.short_line IS NOT NULL

  AND  ST_DWithin(location, segments.short_line, 100) = TRUE

ORDER BY elevation DESC

limit 1;

Which is the same query that took around 300 ms on the smaller dataset. The result was this (https://explain.depesz.com/s/mKFN):

                                                                                     QUERY PLAN                                                                                     
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Limit  (cost=354643835.82..354643835.83 rows=1 width=9) (actual time=225998.319..225998.320 rows=1 loops=1)
   CTE segments
     ->  WindowAgg  (cost=60.08..82.58 rows=1000 width=64) (actual time=0.488..4.032 rows=234 loops=1)
           ->  Sort  (cost=60.08..62.58 rows=1000 width=64) (actual time=0.460..0.875 rows=234 loops=1)
                 Sort Key: pt.path
                 Sort Method: quicksort  Memory: 57kB
                 ->  Function Scan on st_dumppoints pt  (cost=0.25..10.25 rows=1000 width=64) (actual time=0.354..0.387 rows=234 loops=1)
   ->  Sort  (cost=354643753.25..354645115.32 rows=544829 width=9) (actual time=225998.319..225998.319 rows=1 loops=1)
         Sort Key: data.elevation DESC
         Sort Method: top-N heapsort  Memory: 25kB
         ->  Nested Loop  (cost=0.68..354641029.10 rows=544829 width=9) (actual time=349.784..225883.557 rows=159654 loops=1)
               ->  CTE Scan on segments  (cost=0.00..20.00 rows=995 width=32) (actual time=0.500..4.823 rows=233 loops=1)
                     Filter: (short_line IS NOT NULL)
                     Rows Removed by Filter: 1
               ->  Index Scan using location_gist_idx on data  (cost=0.68..356423.07 rows=5 width=41) (actual time=71.416..969.196 rows=685 loops=233)
                     Index Cond: (location && _st_expand(segments.short_line, '100'::double precision))
                     Filter: ((segments.short_line && _st_expand(location, '100'::double precision)) AND _st_dwithin(location, segments.short_line, '100'::double precision, true))
                     Rows Removed by Filter: 8011
 Planning time: 4.554 ms
 Execution time: 225998.839 ms
(20 rows)

So a little less than four minutes. Not bad (given the size of the database), or so I thought.

This morning (so a couple of days later) I ran the query again without the explain analyze to check the results, and noticed that it didn't take anywhere near four minutes to execute. So I ran the explain analyze again, and got this:

                                                                                     QUERY PLAN                                                                                     
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Limit  (cost=354643835.82..354643835.83 rows=1 width=9) (actual time=9636.165..9636.166 rows=1 loops=1)
   CTE segments
     ->  WindowAgg  (cost=60.08..82.58 rows=1000 width=64) (actual time=0.345..1.137 rows=234 loops=1)
           ->  Sort  (cost=60.08..62.58 rows=1000 width=64) (actual time=0.335..0.428 rows=234 loops=1)
                 Sort Key: pt.path
                 Sort Method: quicksort  Memory: 57kB
                 ->  Function Scan on st_dumppoints pt  (cost=0.25..10.25 rows=1000 width=64) (actual time=0.198..0.230 rows=234 loops=1)
   ->  Sort  (cost=354643753.25..354645115.32 rows=544829 width=9) (actual time=9636.165..9636.165 rows=1 loops=1)
         Sort Key: data.elevation DESC
         Sort Method: top-N heapsort  Memory: 25kB
         ->  Nested Loop  (cost=0.68..354641029.10 rows=544829 width=9) (actual time=1.190..9602.606 rows=159654 loops=1)
               ->  CTE Scan on segments  (cost=0.00..20.00 rows=995 width=32) (actual time=0.361..1.318 rows=233 loops=1)
                     Filter: (short_line IS NOT NULL)
                     Rows Removed by Filter: 1
               ->  Index Scan using location_gist_idx on data  (cost=0.68..356423.07 rows=5 width=41) (actual time=0.372..41.126 rows=685 loops=233)
                     Index Cond: (location && _st_expand(segments.short_line, '100'::double precision))
                     Filter: ((segments.short_line && _st_expand(location, '100'::double precision)) AND _st_dwithin(location, segments.short_line, '100'::double precision, true))
                     Rows Removed by Filter: 8011
 Planning time: 0.941 ms
 Execution time: 9636.285 ms
(20 rows)

So from four minutes on the first run to around 9 1/2 seconds on the second. Presumably this difference is due to caching? I would have expected any caches to have expired by the time I made the second run, but the data *is* static, so I guess not. Otherwise, I don't know how to explain the improvement on the second run - the query plans appear identical (at least to me). *IS* there something else (for example, auto vacuum running over the weekend) that could explain the performance difference?

Assuming this performance difference *is* due to caching, that brings up a couple of questions for me:

1) Is there any way to "force" PostgreSQL to cache the data? Keep in mind that the database is close to a couple of hundred Gigs of data, so there is no way it can all fit in RAM.

2) In lieu of forcing a cache (which is probably not going to work well, even if possible), what could I do to help ensure that performance is closer to the 9 second mark than the 4 minute mark in general? For example, would it be likely to make a significant difference if I was to add a couple of larger SSD's to hold this data and put them in a stripe RAID (rather than the mirrored 7200 RPM platter drives it is on now)? Since the data is static, loosing the data due to drive failure is of little concern to me. Or would adding more RAM (and tweaking PostgreSQL settings) to be able to increase the cache size help more, even though there would still not be enough to cache everything?

In the end, the low resolution data is probably good enough, and I may be able to come up with some sort of method to use them both - i.e. return a result quickly from the low resolution dataset, while simultaneously firing off the same request to the high resolution dataset, and returning that result when ready, or only using the high-resolution data set when explicitly requested. So having to wait four minutes on occasion for a result from the high-resolution set may not be an issue. That said, it would be nice to know all the options I can present to my boss :-)

-----------------------------------------------

Israel Brewster

Systems Analyst II

Ravn Alaska

5245 Airport Industrial Rd

Fairbanks, AK 99709

(907) 450-7293

-----------------------------------------------

BEGIN:VCARD
VERSION:3.0
N:Brewster;Israel;;;
FN:Israel Brewster
ORG:Frontier Flying Service;MIS
TITLE:PC Support Tech II
EMAIL;type=INTERNET;type=WORK;type=pref:israel@xxxxxxxxxxxxxxxxxx
TEL;type=WORK;type=pref:907-450-7293
item1.ADR;type=WORK;type=pref:;;5245 Airport Industrial Wy;Fairbanks;AK;99701;
item1.X-ABADR:us
CATEGORIES:General
X-ABUID:36305438-95EA-4410-91AB-45D16CABCDDC\:ABPerson
END:VCARD

On Jan 5, 2017, at 1:55 PM, Israel Brewster <israel@xxxxxxxxxxxxxx> wrote:

On Jan 5, 2017, at 1:38 PM, Rémi Cura <remi.cura@xxxxxxxxx> wrote:

Hey,

1 sec seems really good in this case,
and I'm assuming you tuned postgres so the main index fits into ram (work_mem and all other stuff).

You could avoid a CTE by mixing both cte.

WITH pts AS (
    SELECT (pt).geom, (pt).path[1] as vert

    FROM 

    ST_DumpPoints(

        ST_Segmentize(

            ST_GeographyFromText('SRID=4326;LINESTRING(-150.008056 61.179167,-156.77 71.285833)'),

            600

        )::geometry

    ) as pt

) 

SELECT elevation 

FROM data 

INNER JOIN (SELECT 

    ST_MakeLine(ARRAY[a.geom, b.geom]) as short_line

    FROM pts a 

    INNER JOIN pts b 

    ON a.vert=b.vert-1 AND b.vert>1) segments

ON  ST_DWithin(location, segments.short_line, 600)

ORDER BY elevation DESC limit 1;

Then you could remove the useless and (potentially explosive if you have large number of dump points) inner join on points : 
"FROM pts a 

    INNER JOIN pts b " 

You could simply use a window function to generate the segments, like in here.

The idea is to dump points, order them by path, and then link each point with the previous one (function lag).

Assuming you don't want to use the available function,

this would be something like : 

 

WITH segments AS (
    SELECT ST_MakeLine( lag((pt).geom , 1, NULL) OVER (ORDER BY (pt).path)
                          ,(pt).geom) AS short_line

    FROM ST_DumpPoints(

          ST_Segmentize(

            ST_GeographyFromText('SRID=4326;LINESTRING(-150.008056 61.179167,-156.77 71.285833)'),

            600

        )::geometry

    ) as pt

) 

SELECT elevation 

FROM data ,segments

WHERE segments.short_line IS NOT NULL --the first segment is null by design (lag function)

  AND  ST_DWithin(location, segments.short_line, 600) = TRUE

ORDER BY elevation DESC 
limit 1;

I don't know if you can further improve this query after that,

but I'll guess it would reduce your time and be more secure regarding scaling.

if you want to further improve your result, 

you'll have to reduce the number of row in your index, 

that is partition your table into several tables !

This is not easy to do with current postgres partitionning methods as far as I know

(partitionning is easy, automatic efficient query is hard).

Another way would be to reduce you requirement, and consider that in some case you may want less details in the altimetry, which would allow you to use a Level Of Detail approach.

Congrats for the well explained query/problem anyway !

Cheers,

Rémi-C

Ooooh, nice use of a window function - that change right there cut the execution time in half! I was able to shave off a few hundreds of a second more but tweaking the ST_Segmentize length parameter up to 5,000 (still have to play with that number some), so execution time is now down to the sub-300ms range. If I reduce the radius I am looking around the line, I can additionally improve the time to around 200 ms, but I'm not sure that will be an option. Regardless, 300ms is rather impressive, I think. Thanks!

2017-01-05 23:09 GMT+01:00 Paul Ramsey <pramsey@xxxxxxxxxxxxxxxxx>:

Varying the segment length upwards might have a salutary effect for a while, as the efficiency improvement of fewer inner loops battles with the inefficiency of having more points selected by the index filter. Worth an experiment.

P

On Thu, Jan 5, 2017 at 1:00 PM, Israel Brewster <israel@xxxxxxxxxxxxxx> wrote:

On Jan 5, 2017, at 10:38 AM, Paul Ramsey <pramsey@xxxxxxxxxxxxxxxxx> wrote:

Yes, you did. You want a query that spits out a tupleset of goemetries (one each for each wee segment), and then you can join that set to your main table using st_dwithin() as the join clause.

So start by ditching the main table and just work on a query that generates a pile of wee segments.

Ahhh, I see you've done this sort of thing before (http://blog.cleverelephant.ca/2015/02/breaking-linestring-into-segments.html) :-)

So following that advice I came up with the following query:

WITH dump AS (SELECT

    ST_DumpPoints(

        ST_Segmentize(

            ST_GeographyFromText('SRID=4326;LINESTRING(-150.008056 61.179167,-156.77 71.285833)'),

            600

        )::geometry

    ) as pt

),

pts AS (

    SELECT (pt).geom, (pt).path[1] as vert FROM dump

)

SELECT elevation 

FROM data 

INNER JOIN (SELECT 

    ST_MakeLine(ARRAY[a.geom, b.geom]) as short_line

    FROM pts a 

    INNER JOIN pts b 

    ON a.vert=b.vert-1 AND b.vert>1) segments

ON  ST_DWithin(location, segments.short_line, 600)

ORDER BY elevation DESC limit 1;

Which yields the following EXPLAIN ANALYZE (https://explain.depesz.com/s/RsTD):

                                                                                                                 QUERY PLAN                                                                                                                 
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Limit  (cost=11611706.90..11611706.91 rows=1 width=4) (actual time=1171.814..1171.814 rows=1 loops=1)
   CTE dump
     ->  Result  (cost=0.00..5.25 rows=1000 width=32) (actual time=0.024..1.989 rows=1939 loops=1)
   CTE pts
     ->  CTE Scan on dump  (cost=0.00..20.00 rows=1000 width=36) (actual time=0.032..4.071 rows=1939 loops=1)
   ->  Sort  (cost=11611681.65..11611768.65 rows=34800 width=4) (actual time=1171.813..1171.813 rows=1 loops=1)
         Sort Key: data.elevation DESC
         Sort Method: top-N heapsort  Memory: 25kB
         ->  Nested Loop  (cost=0.55..11611507.65 rows=34800 width=4) (actual time=0.590..1167.615 rows=28408 loops=1)
               ->  Nested Loop  (cost=0.00..8357.50 rows=1665 width=64) (actual time=0.046..663.475 rows=1938 loops=1)
                     Join Filter: (a.vert = (b.vert - 1))
                     Rows Removed by Join Filter: 3755844
                     ->  CTE Scan on pts b  (cost=0.00..22.50 rows=333 width=36) (actual time=0.042..0.433 rows=1938 loops=1)
                           Filter: (vert > 1)
                           Rows Removed by Filter: 1
                     ->  CTE Scan on pts a  (cost=0.00..20.00 rows=1000 width=36) (actual time=0.000..0.149 rows=1939 loops=1938)
               ->  Index Scan using location_gix on data  (cost=0.55..6968.85 rows=1 width=36) (actual time=0.085..0.256 rows=15 loops=1938)
                     Index Cond: (location && _st_expand((st_makeline(ARRAY[a.geom, b.geom]))::geography, '600'::double precision))
                     Filter: (((st_makeline(ARRAY[a.geom, b.geom]))::geography && _st_expand(location, '600'::double precision)) AND _st_dwithin(location, (st_makeline(ARRAY[a.geom, b.geom]))::geography, '600'::double precision, true))
                     Rows Removed by Filter: 7
 Planning time: 4.318 ms
 Execution time: 1171.994 ms
(22 rows)

So not bad. Went from 20+ seconds to a little over 1 second. Still noticeable for a end user, but defiantly usable - and like mentioned, that's a worst-case scenario query. Thanks!

Of course, if you have any suggestions for further improvement, I'm all ears :-)

-----------------------------------------------

Israel Brewster

Systems Analyst II

Ravn Alaska

5245 Airport Industrial Rd

Fairbanks, AK 99709

(907) 450-7293

-----------------------------------------------

On Thu, Jan 5, 2017 at 11:36 AM, Israel Brewster <israel@xxxxxxxxxxxxxx> wrote:

On Jan 5, 2017, at 8:50 AM, Paul Ramsey <pramsey@xxxxxxxxxxxxxxxxx> wrote:

The index filters using bounding boxes.  A long, diagonal route will have a large bounding box, relative to the area you actually care about (within a narrow strip of the route). Use ST_Segmentize() to add points to your route, ST_DumpPoints() to dump those out as point and ST_MakeLine to generate new lines from those points, each line very short. The maximum index effectiveness will come when your line length is close to your buffer width.

P

Ok, I think I understand the concept. So attempting to follow your advice, I modified the query to be:

SELECT elevation

FROM data

WHERE

    ST_DWithin(

        location,

        (SELECT ST_MakeLine(geom)::geography as split_line

         FROM (SELECT

        (ST_DumpPoints(

            ST_Segmentize(

                ST_GeographyFromText('SRID=4326;LINESTRING(-150.008056 61.179167,-156.77 71.285833)'),

                600

            )::geometry

        )).geom

    ) s1),

        600

    )

ORDER BY elevation DESC limit 1;

It took some fiddling to find a syntax that Postgresql would accept, but eventually that's what I came up with. Unfortunately, far from improving performance, it killed it - in running the query, it went from 22 seconds to several minutes (EXPLAIn ANALYZE has yet to return a result). Looking at the query execution plan shows, at least partially, why:

                                  QUERY PLAN                                  
------------------------------------------------------------------------------
 Limit  (cost=17119748.98..17119748.98 rows=1 width=4)
   InitPlan 1 (returns $0)
     ->  Aggregate  (cost=17.76..17.77 rows=1 width=32)
           ->  Result  (cost=0.00..5.25 rows=1000 width=32)
   ->  Sort  (cost=17119731.21..17171983.43 rows=20900890 width=4)
         Sort Key: data.elevation DESC
         ->  Seq Scan on data  (cost=0.00..17015226.76 rows=20900890 width=4)
               Filter: st_dwithin(location, $0, '600'::double precision)
(8 rows)

So apparently it is now doing a sequential scan on data rather than using the index. And, of course, sorting 20 million rows is not trivial either. Did I do something wrong with forming the query?

-----------------------------------------------

Israel Brewster

Systems Analyst II

Ravn Alaska

5245 Airport Industrial Rd

Fairbanks, AK 99709

(907) 450-7293

-----------------------------------------------

On Thu, Jan 5, 2017 at 9:45 AM, Israel Brewster <israel@xxxxxxxxxxxxxx> wrote:

I have a database (PostgreSQL 9.6.1) containing 62,702,675 rows of latitude (numeric), longitude(numeric), elevation(integer) data, along with a PostGIS (2.3.0) geometry column (location), running on a CentOS 6.8 box with 64GB RAM and a RAID10 SSD data drive. I'm trying to get the maximum elevation along a path, for which purpose I've come up with the following query (for one particular path example):

SELECT elevation FROM data                                                                                                                                                                                                                                                                                                                                                                                WHERE ST_DWithin(location, ST_GeographyFromText('SRID=4326;LINESTRING(-150.008056 61.179167,-156.77 71.285833)'), 600)                                                                                                                                                                                                                                                                              ORDER BY elevation LIMIT 1;

The EXPLAIN ANALYZE output of this particular query (https://explain.depesz.com/s/heZ) shows:

                                                                                                                                                                      QUERY PLAN                                                                                                                                                                      
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Limit  (cost=4.83..4.83 rows=1 width=4) (actual time=22653.840..22653.842 rows=1 loops=1)
   ->  Sort  (cost=4.83..4.83 rows=1 width=4) (actual time=22653.837..22653.837 rows=1 loops=1)
         Sort Key: elevation DESC
         Sort Method: top-N heapsort  Memory: 25kB
         ->  Index Scan using location_gix on data  (cost=0.42..4.82 rows=1 width=4) (actual time=15.786..22652.041 rows=11081 loops=1)
               Index Cond: (location && '0102000020E6100000020000002C11A8FE41C062C0DFC2BAF1EE964E40713D0AD7A39863C086C77E164BD25140'::geography)
               Filter: (('0102000020E6100000020000002C11A8FE41C062C0DFC2BAF1EE964E40713D0AD7A39863C086C77E164BD25140'::geography && _st_expand(location, '600'::double precision)) AND _st_dwithin(location, '0102000020E6100000020000002C11A8FE41C062C0DFC2BAF1EE964E40713D0AD7A39863C086C77E164BD25140'::geography, '600'::double precision, true))
               Rows Removed by Filter: 4934534
 Planning time: 0.741 ms
 Execution time: 22653.906 ms
(10 rows)

So it is using the index properly, but still takes a good 22 seconds to run, most of which appears to be in the Index Scan.

Is there any way to improve this, or is this going to be about as good as it gets with the number of rows being dealt with? I was planning to use this for a real-time display - punch in a couple of points, get some information about the route between, including maximum elevation - but with it taking 22 seconds for the longer routes at least, that doesn't make for the best user experience.

It's perhaps worth noting that the example above is most likely a worst case scenario. I would expect the vast majority of routes to be significantly shorter, and I want to say the shorter routes query much faster [testing needed]. That said, the faster the better, even for short routes :-)

-----------------------------------------------

Israel Brewster

Systems Analyst II

Ravn Alaska

5245 Airport Industrial Rd

Fairbanks, AK 99709

(907) 450-7293

-----------------------------------------------