Large numbers of short GPolylines run massively slower than small numbers of long GPolylines.
The speed difference between Google Maps v2 and Google Maps v3 is not going to be significant, because most of the CPU time will be taken up by the actual graphics system of the browser. Google Maps uses the VML, SVG or Canvas graphics systems, depending on the browser. Of these, VML is by far the slowest, and that gets used whenever the browser is MSIE.
Before embarking on tackling 250,000 line segments, I suggest you take a look at this quick speed test of 200 random polylines. Try zooming and paning that map in MSIE.
Then, also consider the amount of data that needs to be sent from the server to the client to specify 250,000 line segments. The amount of data will vary depending on whether you choose KML or JSON or GeoRSS, but if you end up with 20 bytes per line segment that would take 50 seconds to fetch on a 1 megabit broadband connection. Consider whether your users would be prepared to sit around for 50 seconds.
The only solution that really makes sense is to do what Google do for their traffic overlay, and draw the lines onto tiles in the server, and have those tiles be displayed as a GTileLayerOverlay in the client.
What you need is a spatially aware database, and a server-side graphics library like gd or ImageMagik. The client asks for a tile from the server. If the zoom is above a certain level the server scans the database for line segments that have bounding boxes that overlap the bounding box of the requested tile and use the graphics library to draw them.
The zoom level limit is there to limit the amount of work that your database and server needs to do. You don't want to end up drawing 250,000 line segments onto a single zoomed out tile because that's an awful lot of hard work for the server, and isn't going to mean very much to the user.
Regarding click handling:
The easy thing to do is to listen for clicks on the map, rather than on the objects, and send the click details to a server. The server then uses the click location to search the spatially aware database and returns the details of the clicked object if there is one. The client code does this:
GEvent.addListener(map,"click",function(overlay,point) {
var url="clickserver.php?lat=" + point.lat() + "&lng=" +point.lng();
GDownloadUrl(url, function(html) {
if (html.length) {
map.openInfoWindow(html)
}
});
});
The harder thing to do is to handle the changing of the cursor when the pointer is over the polylines. There's a known technique for doing cursor changes for small markers, which works like this:
Whenever a tile is fetched, the .getTileUrl() also makes a call to a server that returns a list of hotspot boxes for that tile. As the mouse moves, the client constantly calculates which tile the mouse is over, and then scans the corresponding list of hotspot boxes.
Google themselves, in their GLayer() code, add the sophistication of performing a quadtree search to speed up the search for hotspots within a tile, but other people who have implemented this strategy in their own code reckon that's not necessary, and a linear scan of the hotspot list is fast enough.
I've no idea how to extend that to handling cursor over polyline detection.