lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chris Male (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (LUCENE-4157) Improve Spatial Testing
Date Wed, 27 Jun 2012 08:51:44 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-4157?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13402063#comment-13402063
] 

Chris Male commented on LUCENE-4157:
------------------------------------

{quote}RE QuadPrefixTree, I'll see if I can reproduce your test errors. I'm not surprised
if the QuadPrefixTree.MAX_LEVELS_POSSIBLE is perhaps too big (notice the comment at it's declaration
"not really sure how big this should be". Assuming the default 12 levels pass, I think we
can find a safer max number to use for the time being that is less than 50, and maybe one
day when we have time we can confidently determine exactly what it can support. I venture
to guess it might be similar to the mantissa of a double which is 53, but perhaps not or maybe
it's half that or something. FYI about 26 is needed for ~1meter accuracy. If a non-geo scenario
is needed, then who knows what your requirements might be.{quote}

Thanks for that explanation.  I tried with the default of 12 and the tests still failed but
no error this time.  That could be just related to the fact quad trees are less precise than
geohashes or maybe some problems with the tests.  I think we should just try to come up with
some tests for the trees themselves to verify that they work as expected.  I see SpatialPrefixTreeTest
does some testing of GeohashPrefixTree currently, but we should really spin that off into
its own test class and take QuadTree separately.

{quote}RE Testing of TermQueryPrefixGridStrategy, I agree that its tests are too minimal,
in Lucene spatial. FWIW, I'm about to update a patch to SOLR-3304 that tests a variety of
strategies against the same test code (based on test code from Solr 3 spatial filter tests).
TermQueryPrefixGridStrategy passes fine.{quote}

Good to know.  I have confidence in TermQueryPrefixGridStrategy since it is extremely simple
but I think we need to come up with tests to ensure that any changes we make to the indexing
process is compatible with the querying.

{quote}I definitely welcome any input on making the tests better overall. It's a bit of a
challenge because there are a variety of strategies, and some like TwoDoublesStrategy are
known to not yet support certain geo cases like the poles (if I recall). I'm not sure if the
idea of a test file of query cases was your idea or Ryan's (e.g. cities-IsWithin-BBox), but
instead or in addition, I like the idea of automatically generating random data and queries,
and then double checking search results against a simple brute force algorithm.{quote}

I don't really like the test file idea at all.  Having them for benchmarking is good but we
aren't at that stage yet.  Instead I think we should construct simple unit tests, indexing
a few Shapes and querying for them.  We should do that for each Strategy, obviously only indexing
Points for TwoDoublesStrategy.  Having random data and query generation can come later, once
we have enough crafted tests to be sure that this works.

We should then randomize the use of QuadTree vs GeohashTree or actually repeat the tests for
both.

We have a big question mark around testing with polygons.  My concern is that users will rightly
start using JTS Geometrys and our Strategies will fail.  We really need to think about how
to handle this.

{quote}If you don't feel any better about these two classes, then I like your suggestion of
not releasing them in 4.0 and leave in trunk.{quote}

QuadTree is my main concern since I don't know whether it is working correctly and is just
less precise than geohashes or has a bug.  If we can't quickly come up with a couple of tests
and fix any broken behavior then we should remove it from 4.0.  

We should also take this opportunity to remove any unused code / code that doesn't actually
test anything.  For this I see TruncateFilter, the current TestTermQueryPrefixGridStrategy
and TestSpatialPrefixField.

I'll try to help out here especially with cleaning out the dead code, but any help with testing
QuadTree would be great.



                
> Improve Spatial Testing
> -----------------------
>
>                 Key: LUCENE-4157
>                 URL: https://issues.apache.org/jira/browse/LUCENE-4157
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: modules/spatial
>            Reporter: David Smiley
>            Assignee: David Smiley
>            Priority: Critical
>             Fix For: 4.0
>
>         Attachments: LUCENE-4157_Improve_Lucene_Spatial_testing_p1.patch
>
>
> Looking back at the tests for the Lucene Spatial Module, they seem half-baked.  (At least
Spatial4j is well tested).  I've started working on some improvements:
> * Some tests are in an abstract base class which have a subclass that provides a SpatialContext.
The idea was that the same tests could test other contexts (such as geo vs not or different
distance calculators (haversine vs vincenty) but this can be done using RandomizedTesting's
nifty parameterized test feature, once there is a need to do this.
> * Port the complex geohash recursive prefix tree test that was developed on the Solr
side to the Lucene side where it belongs.
> And some things are not tested or aren't well tested:
> * Distance order as the query score
> * Indexing shapes other than points (i.e. shapes with area / regions)

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org


Mime
View raw message