If you add helper fields to the coordinates table, you can improve response time of the query.
Like this:
CREATE TABLE `Coordinates` (
`id` INT(10) UNSIGNED NOT NULL COMMENT 'id for the object',
`type` TINYINT(4) UNSIGNED NOT NULL DEFAULT '0' COMMENT 'type',
`sin_lat` FLOAT NOT NULL COMMENT 'sin(lat) in radians',
`cos_cos` FLOAT NOT NULL COMMENT 'cos(lat)*cos(lon) in radians',
`cos_sin` FLOAT NOT NULL COMMENT 'cos(lat)*sin(lon) in radians',
`lat` FLOAT NOT NULL COMMENT 'latitude in degrees',
`lon` FLOAT NOT NULL COMMENT 'longitude in degrees',
INDEX `lat_lon_idx` (`lat`, `lon`)
)
If you're using TokuDB, you'll get even better performance if you add clustering
indexes on either of the predicates, for example, like this:
alter table Coordinates add clustering index c_lat(lat);
alter table Coordinates add clustering index c_lon(lon);
You'll need the basic lat and lon in degrees as well as sin(lat) in radians, cos(lat)*cos(lon) in radians and cos(lat)*sin(lon) in radians for each point.
Then you create a mysql function, smth like this:
CREATE FUNCTION `geodistance`(`sin_lat1` FLOAT,
`cos_cos1` FLOAT, `cos_sin1` FLOAT,
`sin_lat2` FLOAT,
`cos_cos2` FLOAT, `cos_sin2` FLOAT)
RETURNS float
LANGUAGE SQL
DETERMINISTIC
CONTAINS SQL
SQL SECURITY INVOKER
BEGIN
RETURN acos(sin_lat1*sin_lat2 + cos_cos1*cos_cos2 + cos_sin1*cos_sin2);
END
This gives you the distance.
Don't forget to add an index on lat/lon so the bounding boxing can help the search instead of slowing it down (the index is already added in the CREATE TABLE query above).
INDEX `lat_lon_idx` (`lat`, `lon`)
Given an old table with only lat/lon coordinates, you can set up a script to update it like this: (php using meekrodb)
$users = DB::query('SELECT id,lat,lon FROM Old_Coordinates');
foreach ($users as $user)
{
$lat_rad = deg2rad($user['lat']);
$lon_rad = deg2rad($user['lon']);
DB::replace('Coordinates', array(
'object_id' => $user['id'],
'object_type' => 0,
'sin_lat' => sin($lat_rad),
'cos_cos' => cos($lat_rad)*cos($lon_rad),
'cos_sin' => cos($lat_rad)*sin($lon_rad),
'lat' => $user['lat'],
'lon' => $user['lon']
));
}
Then you optimize the actual query to only do the distance calculation when really needed, for example by bounding the circle (well, oval) from inside and outside.
For that, you'll need to precalculate several metrics for the query itself:
// assuming the search center coordinates are $lat and $lon in degrees
// and radius in km is given in $distance
$lat_rad = deg2rad($lat);
$lon_rad = deg2rad($lon);
$R = 6371; // earth's radius, km
$distance_rad = $distance/$R;
$distance_rad_plus = $distance_rad * 1.06; // ovality error for outer bounding box
$dist_deg_lat = rad2deg($distance_rad_plus); //outer bounding box
$dist_deg_lon = rad2deg($distance_rad_plus/cos(deg2rad($lat)));
$dist_deg_lat_small = rad2deg($distance_rad/sqrt(2)); //inner bounding box
$dist_deg_lon_small = rad2deg($distance_rad/cos(deg2rad($lat))/sqrt(2));
Given those preparations, the query goes something like this (php):
$neighbors = DB::query("SELECT id, type, lat, lon,
geodistance(sin_lat,cos_cos,cos_sin,%d,%d,%d) as distance
FROM Coordinates WHERE
lat BETWEEN %d AND %d AND lon BETWEEN %d AND %d
HAVING (lat BETWEEN %d AND %d AND lon BETWEEN %d AND %d) OR distance <= %d",
// center radian values: sin_lat, cos_cos, cos_sin
sin($lat_rad),cos($lat_rad)*cos($lon_rad),cos($lat_rad)*sin($lon_rad),
// min_lat, max_lat, min_lon, max_lon for the outside box
$lat-$dist_deg_lat,$lat+$dist_deg_lat,
$lon-$dist_deg_lon,$lon+$dist_deg_lon,
// min_lat, max_lat, min_lon, max_lon for the inside box
$lat-$dist_deg_lat_small,$lat+$dist_deg_lat_small,
$lon-$dist_deg_lon_small,$lon+$dist_deg_lon_small,
// distance in radians
$distance_rad);
EXPLAIN on the above query might say that it's not using index unless there's enough results to trigger such. The index will be used when there's enough data in the coordinates table.
You can add
FORCE INDEX (lat_lon_idx)
to the SELECT to make it use the index with no regards to the table size, so you can verify with EXPLAIN that it is working correctly.
With the above code samples you should have a working and scalable implementation of object search by distance with minimal error.
