[
https://issues.apache.org/jira/browse/LUCENE-4978?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13786519#comment-13786519
]
David Smiley commented on LUCENE-4978:
--------------------------------------
Hi Oli,
This issue is indeed the root cause of the one you refer to in ES. I spent a
little time on fixing the problem months ago but held off because I wanted to
better understand the performance trade-off, and I hadn't yet developed a
benchmark -- through I have one now in LUCENE-2844.
Correct me if I'm wrong but I heard ES has a point-query optimization. At
least I thought I saw something like that when I looked through ES's docs a
couple months ago. I would like to add such an optimization within
Lucene-spatial which would effectively avoid this particular issue you hit
because it would end up being a simple Lucene term query. This underlying
issue would still exist though, it just wouldn't show up with a point query.
If you want a quick solution that only addresses intersection with a Point
query, then you could modify the code I reference in the comment above to not
use cell.getCenter() when queryShape is an instance of Point. Make sense?
To be clear though, the quick solution or a solution optimizing a point query
doesn't actually address the underlying problem; it just fixes it for point
queries only. It's still possible to index a point that fits inside a query
rectangle extremely close to the edge, and depending on which side of the grid
line the rectangle border is, you might not match the point.
> Spatial search with point query won't find identical indexed point
> ------------------------------------------------------------------
>
> Key: LUCENE-4978
> URL: https://issues.apache.org/jira/browse/LUCENE-4978
> Project: Lucene - Core
> Issue Type: Bug
> Components: modules/spatial
> Affects Versions: 4.1
> Reporter: David Smiley
> Assignee: David Smiley
> Priority: Minor
> Fix For: 4.6
>
>
> Given a document with indexed POINT (10 20), when a search for INTERSECTS(
> POINT (10 20)) is issued, no results are returned.
> The work-around is to not search with a point shape, use a very small-radius
> circle or rectangle. (I'm marking this issue as "minor" because it's easy to
> do this).
> An unstated objective of the PrefixTree/grid approximation is that no matter
> what precision you use, an intersects query will find all true-positives.
> Due to approximations, it may also find some close false-positives. But in
> the case above, that unstated promise is violated. But it can also happen
> for query shapes other than points which do in fact barely enclose the point
> given at index time yet the indexed point is in-effect shifted to the center
> point of a cell which could be outside the query shape, and ultimately
> leading to a false-negative.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
