[Qgis-developer] [Server] performance questions

Sat Feb 27 01:32:22 PST 2016

Hi all,

I have made some comparison on requests done when QGIS Server is called for
a simple GetMap, on the first initialization and aftewards when the cache
is found.

Here are the logs

When QGIS builds the maplayer object, ie when there is no cache for this
layer yet (first call, for example with a getcapabilities). This is only
done if no cache is found for this layer in QGIS Server

1/ Log when estimated metadata is used (and stats are available in PostGIS)
https://paste.sh/yguqyRWL#EPcHHVH3uuijGgEtzH0pAlHb

2/ Log when QGIS does not use the metadata for a layer . Which means this
is done for any view at present, since no estimated metadata is available,
or also for a table if the user forgot to check the option "use metadata"
when he first created the Postgis connection (which can be often the case
because this option is not checked by default in PostgreSQL connection
dialog )
https://paste.sh/frCglZ9A#A1nQ6BXtdBeEiZWQmTJP8v9l

3/ Log when a layer is already in QGIS Server cache, for example for a
GetMap request. This means for table and views, since it does not depend on
metadata existence or not, but only on cache
https://paste.sh/Rkvqf757#K1MCRitovIxMblBVils1VF7r

My conclusions here :

In the first example 1/, we see all the queries QGIS does to get all the
needed information for building the maplayer object
* postgis information : postgis_version() ,  postgis_geos_version()
* search for topology features -> not really usefull server side I think (
? )
* feature count . In 1, estimated metadata is used, so this query is fast
* layer extent calculation . In 1/ QGIS uses st_estimatedextent, so it also
should be fast

In the second example 2/, without available stats (no use of estimated
metadata for views or when the option is not checked in layer ), we can see
the performance killing queries (for complex or big datasets) done by QGIS
to get the needed information :

* check if the field used as primary key is unique :
SELECT count(distinct ""nb"")=count(""nb"") FROM (SELECT row_number()
OVER() AS nb, * FROM ""2013"".v_geo_parcelle) AS ""subQuery_0"""

* Get the extent :
SELECT st_extent(""geom"") FROM (SELECT row_number() OVER() AS nb, * FROM
""2013"".v_geo_parcelle) AS ""subQuery_0""

In 3/, QGIS has already built the cache for the maplayer, and only fetches
the features, which is very fast.

My thoughts on how to improve QGIS server performance with PostgreSQL
provider (and possibly for other providers as well )

* check that layer cache in QGIS server is well used.
- I do not know for example if the 100 default number is for the whole
server (which is then reached very fast and will become useless if there
are many projects served by QGIS), or for each projects.
- there was a bug in QGIS Server which did not use cache whenever a GetMap
request was called with an empty STYLE parameter. It has been fixed but
check your version is not older to this commit :
https://github.com/qgis/QGIS/commit/a5450a78ffcbba9a75afa8be43460ad741fea050

* Add an option in QGIS ( in OWS server properties ) to tell QGIS Server to
not get some information if the user who publish the project knows what he
does, meaning
  - use the extent stored in <maplayer> and do not ask for it server side
  - do NOT check that the primary key is unique -> the person who built the
project and publish it does know that his layer has a unique primary key.
Note that there is a limit in the number of layers QGIS Server will cache (
100 by default).

* Add a even more hard way of telling QGIS Server "Never delete the cache
for this layer, as the extent and the other properties wont change (even if
I allow editing inside my known layer extent )" . This means the cache will
be store permanently for this layer and would be lost only if the apache /
nginx server is reloaded . I do not know if it is feasible

* Store layers metadata "the QGIS way" in the database, as we do for
styles. This means the users building the project has enough rights on the
postgreSQL database, but could be a handy QGIS solution for production
ready projects. This way we could store the metadata for tables AND views,
in a format close to what QGIS needs to build the maplayer object. The user
would define some options, such as cache duration in seconds ( for example,
overwrite the cache if the stored information has more than 1 week, or keep
it as long as I do not manually remove it, via a -1 value for cache
duration ).

As a conclusion, I think we really need a "QGIS Server context true/false"
when using some of the core classes, such as maplayer, to be able to do
some performance improvements.

Régis, René-Luc and myself can have a look at it, if you need some work
done toward this goal.

Cheers,
Michaël

2016-02-23 21:52 GMT+01:00 Andreas Neumann <a.neumann at carto.net>:

> Hi Alessandro,
>
> On 23.02.2016 17:22, Alessandro Pasotti wrote:
>
>>
>>
>> Yes, but Andreas probably meant that plugin support is still compiled
>> into the server.
>>
>
> yes - if you compile yourself, you can disable it in the CMAKE settings.
>
>
>> Still, we never measured any noticeable performance issue with plugins
>> enabled (as long as thre aren't any of course, otherwise it will obviously
>> depend on what the plugins do).
>>
>
> yes - only master versions between 2.12 and 2.14 were affected. It has
> been fixed meanwhile.
>
>
>> There was indeed a performance issue with the new access control
>> implementation (now hopefullt fixed) but IIRC it's still only in master, to
>> be release soon.
>>
>> Also, we aren't aware of any security issue, if you (Andreas) know about
>> any security issue I would definitely like to know more about that.
>>
> I am not aware of any security issues around QGIS server Python plugins.
> It is just good to disable stuff if you don't need it and opt-in if you
> need it.
>
> Andreas
>
> _______________________________________________
> Qgis-developer mailing list
> Qgis-developer at lists.osgeo.org
> List info: http://lists.osgeo.org/mailman/listinfo/qgis-developer
> Unsubscribe: http://lists.osgeo.org/mailman/listinfo/qgis-developer
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/qgis-developer/attachments/20160227/f4d62533/attachment-0001.html>