[nostrademons]

As has been pointed out elsewhere on the thread, complexity doesn't scale linearly. It's far easier to write ten sites that each gets 30 qps than it is to write one site that gets 300 qps, and it's easier to write ten sites that get 300 qps than one site that gets 3000 qps.

Twitter's fundamental problem also is a harder one to scale than something like Heroku or Wordpress. For those hosted sites, you can shard easily by host, so that each of the 100,000 Heroku-hosted sites can get its own EC2 instance(s) and behave pretty much independently. You can't do that when the point of your site is that any action might instantly be broadcast to thousands of followers. High-fanout writes are not an easy problem to solve.

[beagle3]

But if you design it well enough, it does. Here's a copy-pasta from a reddit comment of mine from a few months ago: http://www.reddit.com/r/programming/comments/b2u6t/twitter_o...

------------------

Color me unimpressed.

At some point, I was collecting 40GB/day of financial data (and that's after bzip2ing them .. probably 200GB/day before); This was done on hardware costing $30K (which was two equivalent machines with 4GB, each having 20*1TB in a raid configuration -- this was a hot-backup configuration) and the operation run (coded, supervised, administered) by 2 people.

I'm extrapolating from your numbers: Let's say you have 70GB over 14 days = 5GB/day; Let's assume Twitter has 100GB/day of text twits (which, incidentally, means ~1 billion tweets per day which I highly doubt, as they took few years to get to the 1 billion mark, and last I heard they were at less than 100 million twits/day)

Then, at this day and age (numbers selected for 2 years ago, when they had their last infrastructure revision), what you do is buy 20 servers with 8GB of memory each (for, say $5K each), plus a little redundancy, and store all the latest twits in memory, and the most popular user's older twits as well; everything else on disk. Throw in cheap web front-ends that don't even need a local disk, load balancing, and a gigabit ethernet backplane. You're still under $200K in equipment.

Yes, the code is not going to be trivial, but for $100K and 3 months you can get a stunt programmer (I know a few who can do it and won't charge as much).

A run-of-the-mill RDBMS is the wrong tool for this job; Basically, run of the mill anything; but that does not make it incredibly hard.

I think $300K for hardware and software, you can get a Twitter clone that performs as well.

Twitter is successful, but that's not thanks to good engineering.

[nostrademons]

Be careful extrapolating success scaling in one domain to success scaling in another.

I've also done the 40 GB/day of NYSE TAQ data financial analysis thing, and the 1000+ trades/second real-time financial analytics thing. And I work on Google Search, and have a passing familiarity with how other Google products scale.

The scaling challenges of batch financial models vs. real-time financial processing vs. information retrieval vs. email vs. social products are very different. Even going from a model of the web where it's static and changes every few months (like Google of 2004) to one where sites get update every few minutes and users expect to see the updates immediately in search results (like Google of today) requires vastly different technology.

The main thing about scaling that I've learned from working at a couple places that require it is to go into it with a fresh mind each time, and really pay attention to what the requirements are and what you can cut corners on. There're some general principles you should know (eg. Jeff Dean's "Numbers you should know", memory is much faster than disk, cut out layers of abstraction that you don't need), but in order to apply them effectively, you really need to pay attention to the details of your problem domain.

If you think you can solve Twitter's scaling problems, they're hiring, they're pre-IPO, and they're probably giving out decent chunks of stock.

[beagle3]

> Be careful extrapolating success scaling in one domain to success scaling in another.

I agree about that.

> go into it with a fresh mind each time, and really pay attention to what the requirements are and what you can cut corners on. There are some general principles you should know: memory is much faster than disk, cut out layers of abstraction that you don't need, etc - but in order to apply them effectively, you really need to pay attention to the details of your problem domain.

(slightly edited) - This is golden.

However:

> If you think you can solve Twitter's scaling problems, they're hiring, they're pre-IPO, and they're probably giving out decent chunks of stock.

I know I can solve Twitter's scaling problems (I don't think the solution I posted is the end-all-be-all, and for all I know that might not be where their scale problem is -- it is just perceived and argued about this part, which is not very hard).

However, Twitter's abysmal uptime (for the kind of sevice they are providing) had no bearing on their growth in 2008-2009. And even if by re-architecting Twitter they can save $2M/year on operations, it would be dumb to do that before they're in the black for a while and can identify their real profit and loss centers.

Also, those stock are not worth quite as much as people think when you take everything into account. (I've got a successful exit as a non-founder behind me; I'm intimately familiar with all the gory details including taxes, dilution, etc -- Whether options or RSUs, if you are granted anything of value, you have to pay full taxes AT THE TIME OF THE GRANT).

My point was only to show how non-impressive the problem twitter is (supposedly) facing. It's a repeating discussion:

  - Twitter sucks
  - No they don't, they do xxx and it's damn hard

Don't know why I even bother anymore. A company that had (maybe still has?) their millions-of-views-a-day pages created in Ruby doesn't care about solving scale issues.

At Google, you guys throw out closing paragraph tags from the main page when it is clear it renders fine without them.