The details have finally been posted in that thread. And while $6500 is a lot of money, we have to realize that this is a lot of writing - the headline doesn't give the full sense of it.
For those unfamiliar with GAE, a ListProperty is really a collection of properties. The author is using the property as a geohash with a significant number of values, plus he has additional multiproperty indexes defined, plus he's doing a rewrite (delete + write). All combined it appears to be ~460 writes per entity.
So what we're talking about is $6500 for 6.5 billion writes... exactly what is printed on the sales brochure. Is that a lot? Most datastores don't charge by the operation so I don't have a lot to compare it to. It seems expensive but not crazy, especially considering that the data is replicated via PAXOS to 3+ datacenters with automatic loadbalancing and failover.
"Geoboxing is a technique used to search for entities near a point on the earth in a database that can only perform equality queries (like App Engine)"
So their implementation is a compromise on account of GAE's limitations, and they have to pay through the nose to use it. This is when I'd be looking at hosting some features outside of GAE, which is what we do with Full-text Search.
No single platform or environment encompasses every use case. GAE doesn't offer R-trees or other spatial indexes; for that matter neither does MongoDB or MySQL.
Geohashing is a reasonable solution for some spatial problem domains; it's one solution along the spectrum of "precalculate a lot up front and make queries cheap" vs "write in a cheap & easy format but make queries more expensive". Pre-calculation strategies are usually more scalable when you have large query loads, but they suck bigtime if you need to fully recalculate a large body of data (as the original blog author is doing).
Maybe the blogger would be better off using PostGIS; but then, scaling and synchronizing a large cluster of PostGIS systems is nontrivial. The issues here are too application-specific to draw any positive or negative conclusions about appengine.
For those unfamiliar with GAE, a ListProperty is really a collection of properties. The author is using the property as a geohash with a significant number of values, plus he has additional multiproperty indexes defined, plus he's doing a rewrite (delete + write). All combined it appears to be ~460 writes per entity.
So what we're talking about is $6500 for 6.5 billion writes... exactly what is printed on the sales brochure. Is that a lot? Most datastores don't charge by the operation so I don't have a lot to compare it to. It seems expensive but not crazy, especially considering that the data is replicated via PAXOS to 3+ datacenters with automatic loadbalancing and failover.