One of those things you never think about when reading a map or driving in a car, but it's the explanation behind why you fly all the way up past nova scotia when on a plane to europe.
I've got this on a site I'm building that offers a sort by distance option. There's definitely a performance hit. I'm going to look into 'dumber' ways of filtering out the data prior to running this function. Maybe start by including state in the where clause, for example.
Any other ideas?
One solution I saw used more simple arithmetic to calculate a range of coordinates within levels of distance. That could be pre-cached, but it's a lot less accurate.
Hash the points into large tiles. Only calculate the nearby tiles, then find the items that are in the list of tiles (which is faster, due to being indexed.) Then use Haversine or whatever to filter.
(Edit: to be more specific, you can get a pretty good distance measurement using Geohash and comparing strings. Obviously, indexing strings is something databases do well. The exact distance a single character corresponds to depends on longitude & latitude, but there are lookup tables for that. There are also edge conditions to be aware of which may affect your application)
Thanks for the pointer! This looks interesting. The edge conditions seem like they might pose a problem. I'll have to check out how often it would occur. Maybe the geospatial indexes are a better bet. It looks like MongoDB supports them also. Good excuse to try that out.
If you are prepared to introduce new technology specifically to solve this problem, then you should take a look at LocalLucene, too: http://www.gissearch.com/locallucene
Using a B-Tree on a Geohash (like MongoDB does) is a bit more efficient that just indexing min/max values, but not by much. MySQL, PostgreSQL and even SQLite have R-Tree indices that perform 10x better.
true, the distance calculation must NOT be in the WHERE clause if you want to use indexes (and you want).
What I'm doing, given a max distance and a search point, is to calculate the bounding box in which I want to search in filter results with
WHERE lat BETWEEN lat_min AND lat_max AND lng BETWEEN lng_min AND lng_max
Calculating latitude min/max is trivial knowing that 1 latitude degree is 111.2KM. Longitude is a bit more convoluted because longitude degree size changes moving north/south.
Also, thing is filtering is not much of a problem as you can use the Haversine formula ... it gets to be a problem if you also want sorting from closest to furthest.
<rant>I've got a better idea: use postgis and get rid of mysql and its quirks.
Yep, I just reinstalled an old app yesterday, and for some unfathomable reason it loses communication with mysql at some point (it worked perfectly for what, 6 years?). Gosh, I hate mysql.</rant>
It's a perl program running on Debian stable. Apparently it fails to communicate properly with mysql 5, but works fine with mysql 4. I assume the Debian dbd-mysql package to be compatible with the Debian mysql 5 server, so of course it could be Debian fault, or some error in the program. But I prefer to unjustly incriminate mysql because postgres is so much better anyway :)
Anecdote on Haversine: in 1.6, MongoDB's built-in "geonear" was using non-spherical indexing, but I needed to get more accurate results in a project, so I queried 3 times more records and used Haversine afterward to sort records again, client-side. Here's the ruby code:
29 comments
[ 2.7 ms ] story [ 69.1 ms ] threadAny other ideas?
One solution I saw used more simple arithmetic to calculate a range of coordinates within levels of distance. That could be pre-cached, but it's a lot less accurate.
(Edit: to be more specific, you can get a pretty good distance measurement using Geohash and comparing strings. Obviously, indexing strings is something databases do well. The exact distance a single character corresponds to depends on longitude & latitude, but there are lookup tables for that. There are also edge conditions to be aware of which may affect your application)
Or, use Postgres which has geospatial indexes.
Using a B-Tree on a Geohash (like MongoDB does) is a bit more efficient that just indexing min/max values, but not by much. MySQL, PostgreSQL and even SQLite have R-Tree indices that perform 10x better.
If you're already using MongoDB, it's really dead-easy to setup (see the docs).
You can basically reduce the filtering done to something like: WHERE lat BETWEEN val1 AND val2 AND lon BETWEEN val3 AND val4.
So indexing will work.
What I'm doing, given a max distance and a search point, is to calculate the bounding box in which I want to search in filter results with
WHERE lat BETWEEN lat_min AND lat_max AND lng BETWEEN lng_min AND lng_max
Calculating latitude min/max is trivial knowing that 1 latitude degree is 111.2KM. Longitude is a bit more convoluted because longitude degree size changes moving north/south.
$lat_min = $lat - $range_km * (1 / 111.2); $lat_max = $lat + $range_km * (1 / 111.2);
$k = $range_km/6371.04; $lng_min = $lng - rad2deg($k/cos(deg2rad($lat))); $lng_max = $lng + rad2deg($k/cos(deg2rad($lat)));
I swear - next time I'm going to get a R-Tree novelty account.
Also, thing is filtering is not much of a problem as you can use the Haversine formula ... it gets to be a problem if you also want sorting from closest to furthest.
Example: http://www.god-object.com/2009/10/20/geospatial-search-using...
Yep, I just reinstalled an old app yesterday, and for some unfathomable reason it loses communication with mysql at some point (it worked perfectly for what, 6 years?). Gosh, I hate mysql.</rant>
http://gist.github.com/559482
Since then, MongoDB 1.7 has been released with spherical sort support (http://www.mongodb.org/display/DOCS/Geospatial+Indexing#Geos...)