[recuter]

Understandably for a PAAS they tell you to treat memcache as if it wasn't there and make no promises as to the quota.

The reason is that they can twiddle things around behind the scenes without anybody architecting their app around a particular memcache performance fingerprint. If you think about it, that's how you are supposed to treat something ephemeral as memcache anyway, best effort availability, not to be relied on.

I wouldn't worry about this too much. Their datastore being a sort of blackbox is more of a problem. Check out the Khan academy dev blog, this is not a bottleneck.

[thrownaway2424]

That's not a very useful service if the memcache can disappear. It's difficult to properly provision a service when the average case costs X and the worst case costs 1000X because the cache layer is "ephemeral". When the cache is working your cost utilization will look like shit and your financial management is going to ask you to cut costs. Then you'll do that and when the cache suddenly disappears, you'll suffer an outage because you are out of capacity.

So I think it's not very reasonable to offer no-SLA components of a hosted platform. It's either there with some assurances, or it might as well not exist at all.

[dylanvee]

(Khan Academy dev here.) The datastore being a black box is a slight concern to me personally, but the really great tradeoff is letting Google handle the responsibility of keeping our persistence layer, among many other services, up and running. They're quite good at doing that.