asharp

I do find it strange that you can go and get wholesale bw in the US for $1/mbit or less, yet residential ADSL tails, sold with contention ratios in the order of thousands:1 can be sold for like $60/month....

Please note that most of the bw in Asia is to other parts of Asia. There is lots of peering and lots of fibre. Trying to get to the US, on the other hand, and you have a few, very expensive (comparatively) cables. This is why you find cloud companies in korea doing live disk replication into japan. The fibre in the region is just so cheap that it just doesn't matter anymore.

You would probably notice that if you were using local (or even localish) sites, that there would be a dramatic improvement.

What's more interesting is why they're charging per gig not per mb/second.

Also interesting is that this is a routed service as far as I can see....

If they're in the US they probably have access to a HE feed. That's $1/mbit. You'll probably get higher quality transit for < $20/mbit. That's stupidly cheap compared to ec2 prices.

Just as a general note re repeated sig testing errors: Wouldn't it be possible to run a standard A/B testing run over a small but not insignificant number of iterations, and then iterate over that a number of times?

You could then use bayes to find a final estimate of, say, H1. As each high level iteration is fairly small, feedback can be provided to the user, although it couldn't be acted on.

Speaking of which, if we have an expected number of false positives for any given number of test scores, couldn't you take the average number of positives generated as an rv and then try and determine if it is different to the expected number of false positives?

It seems as though this type of error relies on the fact that a single false positive stops the testing rather then continuing on and allowing regression to the mean. By stopping this, it should then stop, or at least reduce this type of error.

It also assumes that the distribution remains constant over time, an assumption that is, well, wrong.

But that's not very interesting. What's interesting is if it's still useful regardless, or if there is some sort of modification that would improve it taking that into account.

Interesting.

Some of the claims made seem to be strange. Adding in additional choices is fine, dealing with multiple choices is fine, modifying each page as you give it to the user is fine, you're just adding in additional assumptions, which, when wrong would completely ruin your test. Similarly these results would completely ruin a bandit algorithm, because it relies on a much larger set of assumptions then a standard a/b test.

One quick example: You lose temporal independence and you are testing X and Y. For the first 15 rounds, X's metric is 100 and Y's is -100. after that, they are reversed. With an epsilon first gambit algorithm with N=15, the algorithm will simply choose X forever.

That said, they are an very interesting set of algorithms and it'd be interesting to see how brittle they are in practice.

Interesting.

So you take the output of the bandit algorithm and use that as an output (ie. X was the "best" design) for the next stage of design?

What type of bandit algorithm do you use?

This is a random question, but why doesn't amazon just have a HVM flag on instance creation?

Is it to try and better DRM the windows machines?

Hmmmm. I couldn't find this server on storm's cloudharmony.com benchmarking page (It only goes up to the 48GB model), but it's interesting comparing nonetheless.

An OrionVM 16 gig of ram server gets an aggregate disk IO score of 156.79 vs. their 170.5 and an IOPS score of 159.79 vs. their 159.99.

And that's with redundant network backed hard disks in a visualised environment with all of those benefits and overheads.

What's also interesting is that this is a pxecloud with local 8 disk SAS raid 10 (Not san storage/etc.)

Overall, very interesting offering.

(NB: I work with orionvm a company that makes IOtastic servers)

Yes, and we are looking for anchor customers :)

Orionvm bills to the millisecond and takes about 15 seconds to come up. They also have the fastest disks around and good cpus. They are currently located in Australia Australia though, which might not be useful.

Disclaimer: I'm a founder at orionvm

Adults especially seem to forget that there is no clear change that happens when kids turn 18. Children turn into adults in a very gradual manner over a large period of time rather heterogeneously. This then causes problems when adults, especially those who have been adults for a large period of time attempt to treat all "children" (including teenagers/etc.) as children, especially when the use of power is used.

To give an example (taken from this page), put yourself into this sort of position. You have somebody who cares for you, wants all that is best for you. Somebody like a spouse. You live together, talk/etc. If you found out that they were spying on you, listening in on what you talk with to your friends, trying to find what websites you visit/etc. You'd consider it a rather massive invasion of privacy and you'd consider the spouse to be at the very least creepy, even if they were aiming to learn this information "for your own good".

Similarly a few posts here are variants on "spy on your kids", using various clandestine methods. I'd imagine that you would find it rather distasteful if your government were caught doing similar things to you, especially if the reasons used were that it was "for your own good", and yet the child's perspective in this isn't being considered.

To illustrate, say you are 18 and one day (age of majority 18), and had any of those things happen to you, you would be, quite understandably, pissed. Now as a 17 year old, your reaction would likely be unchanged, however there would be a lot less understanding and a lot more of a tendency to explain the anger generated and the resulting behavior as simply "poor behavior" of the child.

Just as a random question: Why are people using hash functions designed to be fast for large file digests/etc (ie. SHA) for password hashing?

Isn't this what scrypt/bcrypt/etc. are designed for?

Just a friendly reminder that string addition of salts in hash functions is a Bad Idea (tm), and that you should use HMACs instead.

There is something similar called token payments that is what most people (hopefully) use.

Basically when you get somebodies CCard, you take the details and pass them to the provider (storing them nowhere). You then receive back a token.

You can then bill said token like a standard card, but on compromise you can just invalidate all the tokens and no sensitive data has leaked.

It's listed but it's only used for internal operations (iirc password resets/etc.)

Only the user info section of your profile gets seen by others.

Yeah, however you know more then that per user.

ie. how long the user is on the page, how many pages they view, where they came from (did they come from google/news.yc/etc?)

If you're still stuck, send me an email! (You have no email on your profile page, btw)

Define a series of heuristics.

For example, lets say you have one heuristic that is the number of page views on a particular page, then the number of people who fill out a particular form, then the number of people who.... (Basically various things that should correlate with sales).

The easy way is to then go through and first off measure the correlation between each heuristic and sales and then define some facsimile for the change in sales through something like the sum for all h(corr(h, s)*delta h) with h being the heuristic and s being sales.

You can then use that as a new test heuristic for A/B testing.

That looks cool.

Where did you get the pricing from, to generate your cost estimates (ie. the 10/30c/gig)?

Any particular reason for dedicated databases? Is it just the IO?

Welcome. I'd be happy to help, if I can. Email me if you're interested?

http://orionvm.com.au is cool, easy to use and stupidly fast. (Disks 60% faster then a raid 10 of 15K SAS disks, #1 in a whole suite of benchmarks http://orionvm.com.au/blog/3rd-Party-Performance-Benchmarks/)

Currently Australia only, although if you are in the US and find it interesting, email me.

I would definitely read it. I'd be interested in anything you have to write, and probably a bit more ;)

We can produce a damn fuel cell cheaply.

We can't due to Duponts patents on PEM :(

Talk with a patent lawyer, generally the first consultation is free and is protected under attorney/client privilege.

Once you know roughly how much it will cost and what you would need, you can file for provisional patents (at least here in AUS, IANAL), which are much cheaper and don't end up in the public domain but do establish a priority date.

Then you can easily get any monies required to secure your patents/start work, without having to worry about disclosing too much/etc.

Notice that it is an article from 2007, and most of the points raised here have been fixed to varying degrees.

Only in the case of a very badly configured server.

I believe the orionvm cloud would qualify as "good or even average disk performance" (http://orionvm.com.au/blog/3rd-Party-Performance-Benchmarks/ Benched by cloudharmony.com), as they are 60% faster than a dedicated server with 4 * 15k SAS disks.

Disclaimer: I'm a director at orionvm.

Bcrypt would be more secure.

The problem is that by using a unique salt per user, you can't simply create one salted hash and use it on every user hash simultaneously, however you can still check one password each hash you generate. Bcrypt/scrypt are more secure as it requires much more effort to check each password.