Arbitrary configuration networked disk is Hard, probably harder than anything else AWS does. Enterprise still spends a lot on storage (generally big hardware RAIDs, dedicated storage area networks, etc.). 1 GB of enterprise storage might cost 50-100x the base per-GB cost of the cheapest SATA disk, by the time you include all the costs. No AWS customer wants to pay that much of a premium.

You can often build a storage system for a very specific purpose more more cheaply (maybe 3x the base drive cost?), like the Backblaze pods, or just figure the cost of adding 4-8 x SATA drives and software RAID 10 to each of your servers, but then you get to try to solve the filesystem problem, and have weird performance characteristics. Plus, if you think EBS reliability is bad (it is), wait for the joy of statistically significant number of spinning rust drives in anything not designed with enough redundancy, and where some failures are correlated (trays failing at the same time, controller bugs, ...).

One of Google's big advantages was always their Google File System. And GFS is a lot better for the specific loads in search than for the kind of loads applications like gmail impose.

Local instance storage is really the only performant and cost effective thing. You can hang a lot more direct attached disk off your own server than off any AWS instance (1.7TB).

The other area where AWS kind of falls down is "mad RAM" servers. I can buy 288GB RAM servers (18-slot) cheaply now (well, 144GB is cheap, 288GB is less so), and 2TB RAM is available (64-slot). AWS tops out at 68GB for about 5x the lease payments on these servers (so maybe 2x the loaded cost after Reserved Instance discounts)

Wondering that too. I really like the idea of having hosting in the cloud. But I'm not going to move anything critical there until the storage gets better and more predictable. All these stories you read about having to stripe several EBS volumes together to get even half-decent performance seems to add a lot of complexity/extra engineering for something that ought to just work out of the box. Seems to me that its unlikely that EBS will be any good until their entire infrastructure gets upgraded to 10gbit ethernet.

Meanwhile we still have 1 1/2 year left on our current 3 year datacenter contract(we run on 6 server vmware setup hooked up to a hitachi SAN on which we rent storage space). So Amazon have around 1 year left to come up with some good reasons for us to switch. Otherwise we will probably just grab another 3 years with the current provider.

It's a hard problem to solve, because you're selling disk based on capacity. Problem is, capacity is easy to figure out, but capacity at a specific performance profile is hard.

In an enterprise datacenter, a "fast" Tier-1 or Tier-2 SAN runs anywhere from $10-40/GB/Year to the end user. A "cheap" (and much slower) Tier-3 SAN runs anywhere from $1-15/GB, depending on utilization, availability, etc.

I don't think even those rates are possible with Amazon (with the exception of S3 object storage), because they don't know anything about their users. In the enterprise, you segment your users based on anticipated peak performance requirements. Amazon has no idea what your performance profile is -- and no idea whether it will change tomorrow.

From hearsay (I don't have direct knowledge) EBS is an EMC SAN solution. So Amazon can't really do anything directly to fix this. They probably have put pressure on EMC for years.