The future of databases is RAM clouds and SSD

Huh, interesting. Any resources where I can learn more about these firmware optimizations? Intuitively, that doesn't make much sense: it would seem that dropping a lot of the software overhead associated with concurrency, etc, would yield a significant improvement.

Having said that, it's also possible that the actual 'software overhead' is pretty much negligible when compared to disk access times.

Would love to learn more on this.

It's not that simple. The FTL addresses two fundamental problems with flash - slow random write performance and wear leveling. It works beautifully on consumer hardware, but performs rather poorly in an enterprise environment. For some workloads you can squeeze a lot more than 5-10% improvement out of the drives. But that's not the interesting part (most of today's workloads are read-intensive anyway).

The interesting part comes from being able to execute existing operations in fundamentally different ways. When you're not restricted to sequential reads you can implement a huge chunk of database internals differently - different concurrency models, different implementations of certain joins, etc. This is where the bulk of the performance improvement comes from.

Once you do this, you realize that a traditional bottleneck (I/O) is completely blown away. Suddenly, all kinds of things nobody ever bothered to optimize become performance bottlenecks (for example, the query plan is traditionally interpreted, not compiled to native code - not good enough once you get an I/O subsystem that's 250 times faster than what you used to have).

Once you get rid of the software bottlenecks, you realize that you're limited by the bus. Since SSDs are highly parallelized devices, you have to change the way you interleave operations for optimal performance. That alone is another order of magnitude.

In three years or so, the flash manufacturers will put chips on the SSDs that will allow you to send down programs to them (not unlike the shaders in video cards). That will likely be a bigger revolution than SSDs themselves. Where we are now is just the beginning.

You're right that we haven't shown any evidence that our approach is anything but a niche. We will. This stuff just takes a long time to build.

We committed to SQL Server a long time ago. It's served us very well and it will take a lot of effort to replace it. To keep up with increasing workloads we've had to purchase reasonably expensive hardware - I think the last server cost us almost $100k. The #1 bottleneck has always been disk IO. To compensate for slow disks you need to increase RAM and make big RAID arrays, both fairly expensive options compared to the Intel SSDs available now.

Upgrading our current hardware with only $10k worth of SSDs will add at least a year or two to the life of our SQL Servers.

The missing discussion point for me is more fundamental - how and when should I break with my RDBMS data modeling habits?

I can't say for sure, but suspect some of the headlong rush into denormalized NoSQL stores is motivated by the ease of the data model. I wonder if once you go down that road, if your "program contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of" what you get in a decent RDBMS. (Forgive my hijacking of Greenspun)

That said, I'm usually working in denormalized models anyway (star schemas in data warehouses for handy-dandy slice/dice/drill), so I'm completely open to different ways of approaching data processing and analysis problems.

I actually feel more free with SQL than limited though. What I would really enjoy would be to have the fully expressive power of SQL embedded in a natural feeling way in a higher level language. I'm interested in Q (of KDB) for this reason, have also been baking my brain in prolog to tear down old data model habits, and hacking some in Ruby and Common Lisp for the same reasons.

Of course, I'm probably breaking my career in the process, since I've made a habit of getting paid to work on RDBMS walking the line between hardware/software performance tuning. But because of that, I'm also glad to see new approaches to RDBMS storage and innovation in the whole DB space.

Interesting size; I was talking to guys at a gay dating site, and their worldwide db was around 60 GB also; it could fit in RAM, on the 4XL amazon ec2 instances.

For now, i'd say just buy more ram. Basecamp apparently serve 100% of their data from memory.

If you haven't already maxed the RAM on the server in question, that's generally the first step.

When (the active working set) of your DB is much larger than the RAM your server can support, either get a bigger server or put heavy random IO loads on SSDs.

You're asking it as an A-xor-B when it's really a case of A-then-B.

Properly, Flash is another tier in the storage hierarchy. Faster than disk in important measures, but more expensive. Slower than RAM, but cheaper, lower power, and durable. Unfortunately, the infrastrucure for taking easy advantage of Flash isn't very mature yet. SSDs make flash act like a hard disk, but OSs and applications generally don't allow transparent fined-grained data management to take advantage of it. I guess ZFS is best in this regard, since it can migrate hot data to fast devices, but I haven't seen good options on Linux, and I haven't seen hardware RAID controllers that could fill the gap. In your DB, you can place tables on different devices, or use table partitioning to separate recent data from cold, but that isn't as easy as adding more RAM.

Ultimately though, cost effectiveness really depends on your workload. MySpace is apparently finding it cost effective to use SSDs for their cache tier. A few months back James Hamilton linked to a study that suggested that for some common workloads, the case for SSDs had to include both price/performance and performance/watt, since it didn't pencil out for each individually. That'll change though, as prices drop and performance improves.

All the major RDBMS do recursive queries nowadays.