Hacker News Reader

Comment by mcav | Hacker News Reader

Wow:

> For this feature, the fully denormalized Cassandra dataset weighs in at 3 terabytes and 76 billion columns.

What's amazing is that 3 terabyes costs about $300 nowadays. Include good components, power, redundancy, maintenance, and it probably comes to $1,000 per year.

So this feature is actually pretty cheap to do, unless I'm missing something.

Every time I read "terabyte", I have to pause, because my brain goes "That's gigantic!" and then a little voice says, "It's the same size as the hard disk in my desktop...it cost $89."

I have the same sort of weird mental disconnect with SD cards. It's just hard to fathom 32GB of data fitting into something about the size of a quarter.

also note that WD (?) have managed to make a terabyte 2.5" disk. Expect high-end laptops to ship with that soon enough. :)

Hey, I know the feeling, it's weird to adjust to but I still haven't filled my 5 Terabytes of space yet so I'm happy. Soon I'm going to offload all my files to a file-server I'm building with an exorbitant amount of HDD space.

Serving up thousands of requests per seconds from 3 TB (while continually adding and updating more data!) is a lot different from just archiving it.

A _lot_. :)

To put it in perspective, 100GB is pretty much the limit of what you can serve from a single mysql machine, for instance. This will vary depending on exact workload but that is a reasonable ball park number.

Of course, in a relational database you wouldn't (generally) be storing the vast majority of the data in question...you'd be querying for it. That's not to trivialize the improvement in performance here, just to point out that this probably would be 100GB in a MySQL database.

Any idea how much per machine can you serve when you use a bigtable/cassandra/dynamo/pnuts-like system instead of MySQL?

http://wiki.apache.org/cassandra/PoweredBy

330GB per machine for Facebook's Cassandra cluster.

3 TB of disk is easy. 3 TB flowing quickly from disk to RAM to CPU cache and back out again is another story.

I am most of the time disappointed by x86s, but i5/i7 looks more promising: http://arstechnica.com/hardware/news/2009/09/intel-launches-...

I was thinking "WOW", as well, but breck's right. This shouldn't really be that surprising any more given the price of disk storage these days.

Reply on news.ycombinator.com