[postgis-devel] Geog/Geom Hack
Paragon Corporation
lr at pcorp.us
Sat Oct 31 08:02:27 PDT 2009
Nicklas and Paul,
Yap that's the main point. To add
I'm not really in disagreement with Paul. I see his point too. I'm just
prodding him to think about all his use cases a little more because I don't
feel he has.
My feelings to sum up
1) We have not thought about the complete ramifications of this hack and I'm
really concerned about the novice that transitions to an expert rather than
just getting them hooked on PostGIS. Perhaps I'm being overly silly with
that and even said, Paul's approach might be an easier to transition
solution.
2) My concern is the penalty of putting it in and having to take it out
later might be very great (both from a code, testing, as well as a mindset
perspective). I just feel it needs more thought and testing and really if
we want to make our December deadline, I don't want it rushed in so lightly.
Unless of course Paul -- you want to wait till January or February to
release 1.5?
Now the ST_Max_Distance is a separate issue. I would put in the
ST_ConvexHull hack in place. The reason being is that it just improves
performance any way I can think you slice it and an experienced user would
do exactly the same thing always and when you finally incorporate it into
the core function, there is no change in existing code just a speed
improvement. So to me its basically our ST_DWithin hack -- a very tried and
trued obvious answer. Its an implementation detail with no clear leaky
effects.
With that said, there are some clear functions in geometry that are safe to
put a geography cover over. Those ones where there is clearly only one
answer and don't involve transformation.
like ST_X, ST_Y etc. That don't require transformation so no screw up in
data. Also observe that even the geometry(geography .... in these there is
no penalty becuase the geometry/geography isn't changing so the planner can
cache the geometry to geography conversion. The ST_Transformation ones
however, the geometry/geography is changing slightly at each step since
ST_Tranformation is a lossy operation so you are not only incurring overhead
(because these answers can't be cached), but also adding in extra errors .
To me this is a bleeding abstraction and that is the main reason I don't
like it.
So getting back to you Paul,
What functions exactly are you planning to put a veil over? ST_Buffer well
that one is used so much and is not as exact anyway that I suppose I can
grudgingly accept that as okay.
Thanks,
Regina
_____
From: postgis-devel-bounces at postgis.refractions.net
[mailto:postgis-devel-bounces at postgis.refractions.net] On Behalf Of
nicklas.aven at jordogskog.no
Sent: Saturday, October 31, 2009 10:04 AM
To: PostGIS Development Discussion
Subject: Re: [postgis-devel] Geog/Geom Hack
Hallo
I can see the points from both of you. I think the most important argument
from you Regina is that it is not transparent enough. A skillfull user
trying postgis might be disapointed when realizing that the nested functions
caused an unnecessary big rounding-error.
But if it is obvious that it is a "special" function at least for the
experienced user knowing about common function-names I don't think it is a
big problem and might work as an easy shortcut at least during the process
of learning.
As I understand it this could be a solution for using many functions against
geography so, why not note it in the function name like:
ST_tBuffer for transformed buffer. Then when time is to introduse a "real"
variant of the function they can coexist and it will not change the bahavior
inside an application without someone consciously changes the function name
and remove the t.
the t would be independent of geography-geometry in semantics and just
indicate that it is a lower-precision variant. I fit was commonly used it
would work as a warning to experienced users.
I have a similar question about st_max_distance. The function gets very much
more effective when ran together with convexhull. I saw the trick in
ST_MinimumBoundingCircle and id makes a big difference to do :
st_max_distance(st_convexhull(the_geom)) instead of just
st_max_distance(the_geom).
The question is: Should that be put in the sql-function?
My opinion now is that we just tell about it in the documentation and aims
at doing that trick internally in C in the future. Maybe together with
moving the whole convexhull to postgis-native from geos. It didn't look that
impossible fromthe JTS-code.
/Nicklas
2009-10-31 Paul Ramsey wrote:
I still think I'm right :) Honestly, I've got to a lot of trouble to
>make this stuff for newbies, and I don't think "learn how it works" is
>the right answer for them. They had that option before, but taking GIS
>101 is not an option for these people, they need something that "just
>works". It's easier to teach the experienced people the pitfalls than
>the inexperienced people the basics.
>
>BTW, I just upgraded distance_sphere and distance_spheroid to be as
>powerful (handling point/line/polygon) as the geography variants,
>removing excuses for transforming geometries into geographies for
>processing purposes.
>
>P.
>
>On Fri, Oct 30, 2009 at 9:15 PM, Paragon Corporation wrote:
>> Paul,
>> For what its worth, here is another reason why I don't like this idea and
I
>> think we should at least think about its ramifications more so should put
it
>> off for consideration until 2.0.
>>
>> In geometry processing, its common practice to apply a lot of functions
in
>> succession
>>
>> process1(process2(process3(geometry/geography)
>>
>> With your hackish approach -- the unsuspecting novice user will be
incurring
>> a lot of transformation rounding errors with each process
>>
>> The advanced user, won't know if this is okay or not -- because they
can't
>> tell by looking at the function call the hidden transformations going on.
>>
>> If these did not exist, they would transform once before the processes
and
>> once after) and incurr much less penalty
>>
>> But if they both exist, they will treat them as being on equal footing
>>
>> ST_Buffer(geometry) and ST_Buffer(geography)
>>
>>
>> So your approach while well-meaning gives a questionable benefit to
novices
>> and is putting experienced users at a disadvantage.
>>
>> Thanks,
>> Regina
>>
>> -----Original Message-----
>> From: postgis-devel-bounces at postgis.refractions.net
>> [mailto:postgis-devel-bounces at postgis.refractions.net] On Behalf Of
Paragon
>> Corporation
>> Sent: Friday, October 30, 2009 11:54 PM
>> To: 'PostGIS Development Discussion'
>> Cc: 'PostGIS Users Discussion'
>> Subject: Re: [postgis-devel] Geog/Geom Hack
>>
>> Paul,
>> I suppose we can't just put this decision off till 2.0. Isn't this a bit
of
>> scope creep? I'm not absolutely sure which way is better, but I know the
>> cost of rolling back the change is more.
>>
>> If you are going to do this, how many functions are you planning to do
this
>> for?
>>
>> I'm cc'ing the postgis users group too to get more of an opinion on this
>> topic.
>>
>> So the question is it it a good idea to introduce a hack that transforms
a
>> geography into what we call BestSRID to perform geometry operations on
and
>> then transform back. My concern is that this is a silent operation that
>> gives the impression that these functions are natively done in spheroid
>> space just for the benefit of catering to less technical users.
>>
>>
http://trac.osgeo.org/postgis/browser/trunk/postgis/geography.sql.in.c#L541
>>
>> So you can't really tell by looking the penalty
>>
>> Main examples of this as shown for ST_Buffer
>>
>> CREATE OR REPLACE FUNCTION _ST_BestSRID(geography, geography)
>> 530 RETURNS integer
>> 531 AS 'MODULE_PATHNAME','geography_bestsrid'
>> 532 LANGUAGE 'C' IMMUTABLE STRICT;
>> 533
>> 534 -- Availability: 1.5.0
>> 535 CREATE OR REPLACE FUNCTION _ST_BestSRID(geography)
>> 536 RETURNS integer
>> 537 AS 'SELECT _ST_BestSRID($1,$1)'
>> 538 LANGUAGE 'SQL' IMMUTABLE STRICT;
>> 539
>> 540 -- Availability: 1.5.0
>> 541 CREATE OR REPLACE FUNCTION ST_Buffer(geography, float8)
>> 542 RETURNS geography
>> 543 AS 'SELECT
>> geography(ST_Transform(ST_Buffer(ST_Transform(geometry($1),
>> _ST_BestSRID($1)), $2), 4326))'
>> 544 LANGUAGE 'SQL' IMMUTABLE STRICT;
>> 545
>>
>>
>>
>> Thanks,
>> Regina
>>
>> -----Original Message-----
>> From: postgis-devel-bounces at postgis.refractions.net
>> [mailto:postgis-devel-bounces at postgis.refractions.net] On Behalf Of Paul
>> Ramsey
>> Sent: Friday, October 30, 2009 11:07 PM
>> To: PostGIS Development Discussion
>> Subject: Re: [postgis-devel] Geog/Geom Hack
>>
>> We're going to have to agree to disagree on this one, Regina. Catering to
>> the less technical users is what this exercise is all about, to my mind,
and
>> that includes allowing easy flipping into geometry for calculations that
>> aren't supported in geography yet. Oracle does this too.
>>
>> What do other folks think?
>>
>> P.
>>
>> On Fri, Oct 30, 2009 at 7:06 PM, Paragon Corporation wrote:
>>> Paul,
>>> Hmm when I am comparing distance of two geometries in different
>>> spatial refs which I do a lot.
>>>
>>> I still don't like the hack even if you disregard the above or if you
>>> must hack -- don't give it the same name as the non-hacked functions.
>>>
>>> the whole idea of picking BestSRID for a person to cater to less
>>> technical users I find extremely annoying as I can think of 20
>>> "BestSRID" depending on what I am doing. If they get to that level of
>>> sophistication, I would rather have them think a little more and
>>> understand the implications of those decisions.
>>>
>>> We must learn to crawl before we can learn to walk,because walking
>>> without understanding will just get you into trouble in the long run.
>>>
>>>
>>> Thanks,
>>> Regina
>>>
>>> -----Original Message-----
>>> From: postgis-devel-bounces at postgis.refractions.net
>>> [mailto:postgis-devel-bounces at postgis.refractions.net] On Behalf Of
>>> Paul Ramsey
>>> Sent: Friday, October 30, 2009 9:47 PM
>>> To: PostGIS Development Discussion
>>> Subject: Re: [postgis-devel] Geog/Geom Hack
>>>
>>> You *can*, but I strongly doubt you *will*. Because there's nothing in
>>> geography that isn't already in geometry. So you as a primary geometry
>>> user are going to have no working need to cast things to geography.
>>>
>>> On the other hand, the very first question from users of geography
>>> will be "how can I access ?" So having a
>>> relatively full set of functions already available in geography makes
>>> sense to me, even if they are hacked in with a planar trick.
>>>
>>> P.
>>>
>>> On Fri, Oct 30, 2009 at 5:46 PM, Paragon Corporation wrote:
>>>> Paul,
>>>> I can put a functional geography index on can't I and take advantage
>>>> of geography index bindings?
>>>>
>>>> Lets say I have a large network of tables broken out by region so I
>>>> know a specific table has one srid.
>>>>
>>>> For many queries, I may go to that table directly or if I'm doing
>>>> single geometry processing, really don't care what srid as long as
>>>> its in utm or whatever - so I can use the full power of GEOS.
>>>>
>>>> For my across the board distance checks and so forth, I would want to
>>>> use geography and I could use a geography index if I put a functional
>>>> geography index on my geometry correct? Though that needs some more
>>> testing.
>>>>
>>>>
>>>> So in short if 90% of my workload involves geometry processing, I
>>>> will want to keep my data in geometry
>>>>
>>>> But the 10% I would want to convert to geography on the fly.
>>>>
>>>> Thanks,
>>>> Regina
>>>>
>>>> -----Original Message-----
>>>> From: postgis-devel-bounces at postgis.refractions.net
>>>> [mailto:postgis-devel-bounces at postgis.refractions.net] On Behalf Of
>>>> Paul Ramsey
>>>> Sent: Friday, October 30, 2009 8:34 PM
>>>> To: PostGIS Development Discussion
>>>> Subject: Re: [postgis-devel] Geog/Geom Hack
>>>>
>>>> I'm not sure I understand why you would ever convert a geometry to a
>>>> geography as part of a query on a geometry table. I fully expect
>>>> geography to be used as a storage type, because of the utility of
>>>> having the correct spherical indexes, which are not available when
>>>> you're just converting in via a cast. Since there's no functions
>>>> available on geography that are not already available on geometry,
>>>> why would you ever do a geometry->geography cast unless you are (a)
>>>> testing geography or (b) bulk converting a table into geography for
>>> storage in that type.
>>>>
>>>> P.
>>>>
>>>>
>>>>
>>>> On Fri, Oct 30, 2009 at 5:24 PM, Paragon Corporation wrote:
>>>>> Paul,
>>>>> I would rather you didn't for 2 reasons
>>>>>
>>>>> 1) I'm lazy and for each of these things we'd have to apply the text
>>>>> additional function proto hack to prevent from it breaking geometry.
>>>>> which we will probably end up taking out anyway.
>>>>>
>>>>> 2) I don't like the hiddenness of it since it becomes especially
>>>>> annoying if you have your native in geometry and you are converting
>>>>> to geography for a special usecase, then you end up with a slower
>>>>> implementation
>>>>>
>>>>> as you would really end up doing accidentally
>>>>>
>>>>> geometry -> geography -> geometry ->operation (and why do I want my
>>>>> calcs done in UTM?)
>>>>>
>>>>> Instead of the more efficient
>>>>>
>>>>> geometry -> operation
>>>>>
>>>>> Thanks,
>>>>> Regina
>>>>>
>>>>> -----Original Message-----
>>>>> From: postgis-devel-bounces at postgis.refractions.net
>>>>> [mailto:postgis-devel-bounces at postgis.refractions.net] On Behalf Of
>>>>> Paul Ramsey
>>>>> Sent: Friday, October 30, 2009 8:16 PM
>>>>> To: PostGIS Development Discussion
>>>>> Subject: [postgis-devel] Geog/Geom Hack
>>>>>
>>>>> I'm interested to know what the general opinion is of the trick I've
>>>>> used on
>>>>> ST_Buffer(geography):
>>>>>
>>>>> http://trac.osgeo.org/postgis/browser/trunk/postgis/geography.sql.in.
>>>>> c
>>>>> #L541
>>>>>
>>>>> I ask because I could apply the same idea to the larger suite of OGC
>>>>> SFSQL predicates before release. Is half-a-loaf better than no loaf
>>>>> in
>>>> this case?
>>>>> (Note that there will be failure cases for really large geometry,
>>>>> like a polygon of "Asia" or "Russia" that have polygons over the
>>>>> dateline.)
>>>>>
>>>>> P.
>>>>> _______________________________________________
>>>>> postgis-devel mailing list
>>>>> postgis-devel at postgis.refractions.net
>>>>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> postgis-devel mailing list
>>>>> postgis-devel at postgis.refractions.net
>>>>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>>>>
>>>> _______________________________________________
>>>> postgis-devel mailing list
>>>> postgis-devel at postgis.refractions.net
>>>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> postgis-devel mailing list
>>>> postgis-devel at postgis.refractions.net
>>>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>>>
>>> _______________________________________________
>>> postgis-devel mailing list
>>> postgis-devel at postgis.refractions.net
>>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>>
>>>
>>> _______________________________________________
>>> postgis-devel mailing list
>>> postgis-devel at postgis.refractions.net
>>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>>
>> _______________________________________________
>> postgis-devel mailing list
>> postgis-devel at postgis.refractions.net
>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>
>>
>> _______________________________________________
>> postgis-devel mailing list
>> postgis-devel at postgis.refractions.net
>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>
>>
>> _______________________________________________
>> postgis-devel mailing list
>> postgis-devel at postgis.refractions.net
>> http://postgis.refractions.net/mailman/listinfo/postgis-devel
>>
>_______________________________________________
>postgis-devel mailing list
>postgis-devel at postgis.refractions.net
>http://postgis.refractions.net/mailman/listinfo/postgis-devel
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/postgis-devel/attachments/20091031/3282467b/attachment.html>
More information about the postgis-devel
mailing list