Blog

On the new pricing scheme of Google App Engine

phraktle — Sat, 10 Sep 2011 01:09:40 +0000

Yesterday I’ve been to a meetup in San Francisco with some Google engineers working on App Engine. We’ve been mostly chatting about the upcoming new pricing scheme, which is causing a considerable amount of frustration in the GAE community. Not surprising, considering how 1) the pricing model has fundamentally changed 2) people can expect to pay significantly more. The reasoning is that Google wants to find a sustainable business model for App Engine. Previously the pricing was clearly “strategic”, now Google seems to be focused on building a better business as also evidenced by the recent project cleanups. While that’s all well, I feel Google has made a few mistakes here.

First of all, the way the pricing change was communicated to customers was rather unfortunate: most people only realized what it really means when Google rolled out a comparison tool that shows what your previous statements would be priced at now. The results were quite shocking, with at least an order of magnitude difference for many GAE apps. For example, one app we host would go from $0.30/day to $22/day. Several others have reported similar results on the newsgroup. While Google provided hints for optimization, telling your customers they will need to pay 10 to 50 times more, lest they make some changes soon, is a sure-fire way to alienate them. Let’s call this a tactical mistake.

The other mistake in my opinion is more strategic. The fundamental change in the billing structure is to move from actual CPU usage to instance hours. There are many good reasons to rationalize this approach (particularly if you work at Google :), but I think this just killed one of the key differentiators off App Engine. Yes, there are still major differences from other cloud solutions such as Amazon EC2, but in terms of pricing model they just made GAE directly comparable. People can now easily calculate if Amazon's RDS + Elastic Beanstalk provides a decent alternative to GAE.

From a developer standpoint, an essential merit of App Engine used to be that you’d only worry about the resources consumed by executing your code. Now you have to worry about the resourced used while not running your code (ie. idle instances) – in fact this is the source of most of the expenses people got shocked with. Everyone loves a simple abstraction, that’s why we have garbage collection in Java, for example. It takes a lot of experience to tune the garbage collector settings properly – ideally this is something that your DevOps team should take care of instead of the average developer. Analogously, tuning the instance allocation algorithm on GAE should be the concern of Google’s engineers and not be exposed to developers. We love App Engine precisely because there’s no need for configuring your kernel settings, tuning the filesystem, optimizing garbage collection options, etc. Just make the infrastructure work, as if by magic :)

In the meantime, for the developers cast out of Eden, here are a few simple things you should do to minimize the costs incurred by idle instances:

enable concurrent requests (only supported in Java, coming soon for Python 2.7!)
turn off “Always On” (this forces GAE to always keep 3 extra instances idle, which is a huge waste of resources for a smaller app)
reduce “Maximum Idle Instances” to one (or a low number that you’re willing to pay for)
increase “Minimum Pending Latency” to something acceptable for your use-case.

And of course the usual optimization rules still apply: reduce your latency, use asynchronous calls where it makes sense, etc.

Thanks to Google for the free beers and snacks! You really should buy a round for all the App Engine developers out there to increase the peace :)

Update: Peter Magnusson, Director of Engineering for Google App Engine, has posted an announcement, pushing the date for the new billing scheme to Nov 1. Definitely worth reading the thread for some additional perspective!

Personalization Expands Filter Bubbles

phraktle — Sun, 12 Jun 2011 08:18:45 +0000

Thanks for all the votes on my first post! The most significant interest was in stories of successful Hungarian online ventures. However, since facts take actual work to research, I’ll go for a quick mental masturbation in the meantime. The subject is Eli Pariser’s TED talk on Filter Bubbles. (If you’re not familiar with the topic, go ahead and watch it, I’ll wait 9 minutes).

The talk, in the classic alarmist TED fashion, conjures a straw man arguing that personalization leads to a world where people only see what they “want”, as determined by obscure algorithms devised by large for-profit corporations. After skimming Eli’s book, it’s clear that he does make some much better considered points as well. However, most of these revolve around basic issues of privacy, data ownership, transparency, etc - most of which the EFF has been doing a decent job of explaining and representing. I am not here to argue with any of these points. However, the primary meme he’s broadcasting (that personalization traps people) is based on a naive view of how these methods work. I will argue, that the opposite is true, and put things into perspective.

Let’s take a look at the basic structure of the problem first: there’s quite a lot of information out there, and not nearly enough time available for us individually. The standard-issue human mind is not yet capable of containing the universe all at once (apart from brief glimpses on acid).

Even the small part of the cosmos that comes knocking to our senses directly is too much to process, so drastic filtering and ranking is inevitable. You already exist in a robust filter bubble.

Ok, with that cleared up, how do we improve the human condition? Phase one: collect underpants. Or perhaps collect and organize the world’s information, external to our marvelous, yet feeble brains (and please store a backup of our mind while you are at it :). Phase three: profit.

Hey, what’s phase two? Good question!

In an open society no single narrative of reality can be considered correct. We should encourage the co-existence of different filtering and ranking approaches and allow people to discover and explore different perspectives. Traditionally, this exploration was mostly limited to one’s immediate physical and social neighborhood, expanding with improved means of transportation. Today this still plays a larger formative role than the time you spend online: it is the primary source of deeper shared experiences. The horizon has grown with written and broadcast media, connecting a select few to larger masses.

Now, the Internet made knowledge accessible in an unprecedented way and enabled many-to-many connections. Due to the amount of information, you are again inevitably in a filter bubble online. The exciting news is that in this new medium we can implement novel methods of exploration, going beyond the random and top-down approaches available in the physical world. We can use algorithms to expand our bubble, to point out new relationships based on new triggers, collective intelligence, etc. The goal of personalization and recommendations is discovery, as Greg Linden also points out. They create serendipity. We are still in the early stages of research on personalization and we are not even doing a great job at filtering utter garbage. Still, we can already do better than pure random or expert recommendations. There are many experiments going on out there to find useful new approaches for discovery, filtering and ranking. And that’s a good thing.

The real harm is not personalization - it’s limiting accessibility and control of information. We need to fight censorship, maintain net neutrality, people need to own their personal data, companies need to provide transparency on how they use it, and so on.

So, folks, keep your information diet healthy: be curious, travel a lot, talk to people - and check your Facebook sometimes, until something better undoubtedly replaces it. And don’t expect algorithms to commit random acts of poetic terrorism yet.

Hello, World!

phraktle — Mon, 30 May 2011 13:06:51 +0000

Welcome to my blog! I am starting this as an experiment to practice publishing a bit more.

In the spirit of lean methods, I would prefer to write about stuff that people actually care about :) So, I am asking you to vote in this survey (also shown at the bottom) on what you would like to hear more about. Here are some topics that occurred to me that I am motivated to write a post on:

Filter Bubbles - Eli Pariser’s TED talk (and book, which I have yet to read :) warns about what he perceives as the risks of personalization. I’ll explain why I think he fundamentally misses the point of recommendations and how they really work.

Metrics in e-commerce - I could write about the challenges of trying to measure the effectiveness of features on e-commerce sites. How do you know a particular feature adds value? Is it possible to measure changes in revenue using A/B tests?

Real-world recommender systems vs. Netflix Prize - The Netflix Prize did get a lot of attention from press and academia, offering $1M for an improved algorithm for predicting ratings. How relevant are the results of this contest to practical recommender systems? Not a whole lot, I would argue.

Budapest New Tech Meetup - We started this monthly get-together in 2007, and it’s still going strong, having become a “staple” of the local geekdom. I could talk about how it came about and what we have learned in the last few years.

Startup scene in Hungary - In the last few years, the local infrastructure for startups has been growing: more entrepreneurs, more investors, more events, more hype. I could take a look at the landscape and review some of the key organizations, events and people that you could connect with if you’re a startup or a wantrepreneur.

Successful online ventures from Hungary - There are only a handful of tech companies and entrepreneurs around here that can truly be considered successful internationally, such as LogMeIn (public on NASDAQ), Indextools (acquired by Yahoo!), Vatera (acquired by Naspers). One could perhaps also mention Ustream and Prezi as well, or the unmentionable LiveJasmin :). I could do some research to tell more about their story.

Adaptive agile - Over the last six years at EU Edge I tried to apply agile development principles on the projects we did for clients of various sizes. While I consider Scrum a good foundation (as it provides some good patterns), one has to tweak the methodology to fit a given project. I could talk about some observations and rules of thumbs on choosing the right practices.