What's That Noise?! [Ian Kallen's Weblog]

All | LAMP | Music | Java | Ruby | What's That Noise? | The Agilist | Commute | Ball
Main | Next page »

20100123 Saturday January 23, 2010

2010: An Odyssey Continues

Hello, 2010. 2009 was a strange year, indeed. I quit my job after 5 years last spring, since then I've been working as a software/infrastructure consultant and playing with a lot of new technologies (of course also leveraging expertise I have with some more established ones). I've also done a lot of thinking. Thinking about cloud computing and compute marketplaces, about energy, mobile and transportation technologies, about life and the meaning of the whole thing. Consulting has had its ups and downs. Looking back at 2009 though, I can safely say I'm glad it's over. Not that it's all bad news, I have my share of things to be pleased about too. My daughter's Bat Mitzvah was beautiful. I've been spending more time with the kids, switched to decaf (I really enjoy the taste of coffee but excess caffeine isn't healthy) and was invited by my old friends in Metallica to some of their seminal events.

Nonetheless, 2010 will no doubt be an upgrade over 2009.

One of the things I'm spending more domestic time around is Odyssey of the Mind, a creative problem solving competition that my son is participating in. My daughter did it previously and went to the world finals with her team twice, my son went last year, and (warning: proud papa bragging alert) placed 4th! This year the coaching torch is passed to me and I'm working with the team facilitating their solution for the long term problem which will be presented at the regional competition next month. We're hoping for a set of repeat victories that will send us to the world finals again this year. I'm new to coaching OotM but my co-coach has coached before and I've helped the teams in supporting roles in years past, so the Odyssey regime is not completely new to us.

This week will present some interesting challenges. My co-coach is physician and will be flying into Haiti to join the relief effort; I'm really happy for her to have this opportunity to be of service. I'll do my best to keep the OotM team moving forward until she returns.

Work will pick up too. Consulting has given me an opportunity to learn Ruby, Rails and a lot of stuff in that technology orbit. I've also been putting Amazon Web Services to heavy use and playing around with Twitter's APIs. I've been looking for opportunities to scratch some Hadoop itches and lately my interests have turned towards programming in Scala, I expect some consulting gigs to shake loose to sate those interests. If the opportunity is right, my entrepreneurial impulses will get the best of me and I'll stop (or scale back) consulting to jump a on new start-up. My desire to set the world on fire will never be satisfied.

Yea, 2010 is gonna be good. Check it out, the Giants have made some changes that look like a credible offense. Already, it's gotta be better than 2009.

( Jan 23 2010, 10:02:26 PM PST ) Permalink
Comments [1]

20091203 Thursday December 03, 2009

Scaling Rails with MySQL table partitioning

Often the first step in scaling MySQL back-ended web applications is integrating a caching layer to reduce the frequency of expensive database queries. Another common pattern is to denormalize the data and index to optimize read response times. Denormalized data sets solve the problem to a point, when the indexes exceed reasonable RAM capacities (which is easy to do with high data volumes) these solutions degrade. A general pattern for real time web applications is the emphasis on recent data, so keeping the hot recent data separate from the colder, larger data set is fairly common. However, purging denormalized data records that have aged beyond usefulness can be expensive. Databases will typically lock records up, slow or block queries and fragment the on-disk data images. To make matters worse, MySQL will block queries while the data is de-fragmented.

MySQL 5.1 introduced table partitioning as a technique to cleanly prune data sets. Instead of purging old data and the service interruptions that a DELETE operation entails, you can break up the data into partitions and drop the old partitions as their usefulness expires. Furthermore, your query response times can benefit from knowing about how the partitions are organized; by qualifying your queries correctly you can limit which partitions get accessed.

But MySQL's partition has some limitations that may be of concern:

So you want to have an auto_increment primary key but partition on the date? Too bad. MySQL's doc on the matter closes woefully, "We are working to remove this limitation in a future MySQL release series." I don't think PostgreSQL's table partitioning has this limitation but I haven't verified that.

When we first looked at MySQL's table partitioning, this seemed like a deal breaker. The platform we've been working with was already on MySQL 5.1 but it was also a rails app. Rails has its own ORM semantics that uses integer primary keys. What we wanted were partitions like this:

Here, data is segmented into daily buckets and as time advances, fresh buckets are created while old ones are discarded. Yes, with some social web applications, a week is the boundary of usefulness that's why there are only 8 buckets in this illustration. New data would be written to the "Now" partition, yesterday's data would be in yesterday's bucket and last Monday's data would be in the "Monday" bucket. I initially thought that the constraints on MySQL's partitioning criteria might get in the way; I really wanted the data partitioned by time but needed to partition by the primary keys (sequential integers) that rails used like this:

In this case, data is segmented by a ceiling on the values allowed in each bucket. The data with ID values less than 100 go in the first bucket, less than 200 in the second and so on.

Consider an example, suppose you had a table like this: 

CREATE TABLE tweets (
  id bigint(20) unsigned NOT NULL AUTO_INCREMENT,
  status_id bigint(20) unsigned NOT NULL,
  user_id int(11) NOT NULL,
  status varchar(140) NOT NULL,
  source varchar(24) NOT NULL,
  in_reply_to_status_id bigint(20) unsigned DEFAULT NULL,
  in_reply_to_user_id int(11) DEFAULT NULL,
  created_at datetime NOT NULL,
  PRIMARY KEY (id),
  KEY created_at (created_at),
  KEY user_id (user_id)
);
You might want to partition on the "created_at" time but that would require having the "created_at" field in the primary key, which doesn't make any sense. I worked around this by leveraging the common characteristic of rails' "created_at" timestamp and integer primary key "id" fields: they're both monotonically ascending. What I needed was to map ID ranges to time frames. Suppose we know, per the illustration above, that we're only going to have 100 tweets a day (yes, an inconsequential data volume that wouldn't require partitioning, but for illustrative purposes), we'd change my table like this: 
ALTER TABLE tweets PARTITION BY RANGE (id) (
  PARTITION tweets_0 VALUES LESS THAN (100), 
  PARTITION tweets_1 VALUES LESS THAN (200), 
  PARTITION tweets_2 VALUES LESS THAN (300), 
  PARTITION tweets_3 VALUES LESS THAN (400), 
  PARTITION tweets_4 VALUES LESS THAN (500), 
  PARTITION tweets_5 VALUES LESS THAN (600), 
  PARTITION tweets_6 VALUES LESS THAN (700), 
  PARTITION tweets_maxvalue VALUES LESS THAN MAXVALUE
);
When the day advances, it's time create a new partition and drop the old one, so we do this: 
ALTER TABLE tweets REORGANIZE PARTITION tweets_maxvalue into (
  PARTITION tweets_7 VALUES LESS THAN (800), 
  PARTITION tweets_maxvalue VALUES LESS THAN MAXVALUE
);
ALTER TABLE tweets DROP PARTITION tweets_0;
Programmatically determining what the new partition name should be (tweets_7) requires introspecting on the table itself. Fortunately, MySQL has metadata about tables and partitions in its "information_schema" database. We can query it like this 
SELECT partition_ordinal_position AS seq, partition_name AS part_name, partition_description AS max_id 
FROM information_schema.partitions 
WHERE table_name='tweets' and table_schema='twitterdb' ORDER BY seq

So that's the essence of the table partition life-cycle. One issue here is that when we reorganize the partition, it completely rewrites it; it doesn't atomically rename "tweets_maxvalue" to "tweets_7" and allocate a new empty "tweets_maxvalue" (that would be nice). If the partition is populated with data, the rewrite is resource consumptive; the more data, the more so. The availability of a table that has a moderate throughput of data (let's say 1 million records daily, not 100) would suffer, locking the table for a long duration waiting for the rewrite is unacceptable for time sensitive applications. Given that issue, what we really want is to anticipate the ranges and reorganize the MAXVALUE partition in advance. When we allocate the new partition, we set an upper boundary on the ID values for but we need to make a prediction as to what the upper boundary will be. Fortunately, we put an index on created_at so we can introspect on the data to determine the recent ID consumption rate like this: 

SELECT UNIX_TIMESTAMP(created_at),MAX(id) FROM tweets WHERE DATE(created_at)=DATE( now() - interval 1 day );
SELECT UNIX_TIMESTAMP(created_at),MAX(id) FROM tweets WHERE DATE(created_at)=DATE( now() - interval 2 day );
SELECT UNIX_TIMESTAMP(created_at),MAX(id) FROM tweets WHERE DATE(created_at)=DATE( now() - interval 3 day );
SELECT UNIX_TIMESTAMP(created_at),MAX(id) FROM tweets WHERE DATE(created_at)=DATE( now() - interval 4 day );
The timestamps (epoch seconds) and ID deltas give us what we need to determine the ID consumption rate N; the new range to allocate is simply max(id) + N.

OK, so where's the rails? All of this so far has been transparent to rails. To take advantage of the partition access optimization opportunities mentioned above, it's useful to have application level knowledge of the time frame to ID range mappings. Let's say we just want recent tweets from a user M and we know that the partition boundary from a day ago is N, we can say 

Tweet.find(:all, :conditions => ["user_id = M AND id > N"])
This will perform better than the conventional (even if we created an index on user_id and created_at). 
Tweet.find(:all, :conditions => ["user_id = M AND created_at > now() - interval 1 day"])
Scoping the query to a partition in this way isn't unique to rails (or even to MySQL, with Oracle we can explicitly name the partitions we want a query to hit in SQL). Putting an id range qualifier in the WHERE clause of our SQL statements will work with any application environment. I chose to focus on rails here because of the integer primary keys that rails requires and the challenge that poses. To really integrate this into your rails app, you might create a model that has the partition metadata.

I hope this is helpful to others who are solving real time data challenges with MySQL (with or without rails), I didn't turn up much about how folks manage table partitions when I searched for it. There's an interesting article about using the MySQL scheduler and stored procedures to manage partitions but I found the complexity of developing, testing and deploying code inside MySQL more of a burden than I wanted to carry, so I opted to do it all in ruby and integrate it with the rails app. If readers have any better techniques for managing MySQL table partitions, please post about it!

Table partitioning handles a particular type of data management problem but it won't answer all of our high volume write capacity challenges. Scaling write capacity requires distributing the writes across independent indexes, sharding is the common technique for that. I'm currently investigating HBase, which transparently distributes writes to Hadoop data nodes, possibly in conjunction with an external index (solr or lucene), as an alternative to sharding MySQL. Hadoop is sufficiently scale free for very large workloads but real time data systems has not been its forte. Perhaps that will be a follow up post.

Further Reading:

MySQL partitions tutorial
View more documents from Giuseppe Maxia.

   

( Dec 03 2009, 12:08:37 AM PST ) Permalink
Comments [1]

20090910 Thursday September 10, 2009

Hub-a-dub-bub, feed clouds in a tub

While at Technorati, I observed a distinct shift around summer of 2005 in the flow of pings from largely worthwhile pings to increasingly worthless ones. A ping is a simple message with a URL. Invented by Dave Winer, it was originally implemented with XML-RPC but RESTful variants also emerged. The intention of the message is "this URL updated." For a blog, the URL is the main page of the blog; it's not the feed URL, a post URL ("permalink") or any other resource on the site - it's just the main page. Immediately, that narrow scope has a problem; as a service that should do something with the ping, the notification has no more information about what changed on the site. A new post? Multiple new posts? Um, what are the new post URL(s)? Did the blogroll change? And so on. Content fetching and analysis is cheap; network latencies, parsing content and doing interesting stuff with it is easy at low volumes but is difficult to scale.

In essence, a ping is a cheap message to prompt an expensive analysis to detect change. It's an imbalanced market of costs and benefits. Furthermore, even if the ping had a richer payload of information, pings lack another very important component: authenticity. These cheap messages can be produced by anybody for any URL and the net result I observed was that lots of people produced pings for lots of things, most of which weren't representative of real blogs (other types of sites or just spam) or real changes on blogs, just worthless events that were resource intensive to operate on. When I left Technorati in March (2009), we were getting around 10 million pings a day, roughly 95% of them of no value. Pings are the SMTP of the blogosphere; weak identity systems, spammy and difficult to manage at scale.

Besides passively waiting for pings, the other method to find things that have changed is to poll sites. But polling is terribly inefficient. To address those inefficiencies, FriendFeed implemented Simple Update Protocol. SUP is reminiscent of another Dave Winer invention, changes.xml however SUP accounts for discovery and offers a more compact format.

But SUP wasn't the first effort to address the aforementioned deficiencies, 2005 saw a lot of activity around something called "feedmesh." Unfortunately, the activity degenerated into a lot of babble; noise I attribute to unclear objectives and ego battles but I'm sure others have their own interpretation. The feedmesh discussion petered out and little value beyond ping-o-matic and other ping relayers emerged from it. Shortly thereafter SixApart created their atom stream, I think spearheaded by Brad Fitzpatrick. The atom stream is essentially a long lived HTTP connection that streams atom elements to the client. The content flow was limited to SixApart's hosted publishing platforms (TypePad, LiveJournal and Vox) and the reliability wasn't that great, at least in the early days, but it was a big step in the right direction for the blog ecosystem. The atom stream was by far the most efficient way to propogate content from the publishing platform to interested parties such as Technorati operating search, aggregation and analytic systems. It eliminates the heavyweight chatter of cheap ping messages and the heavyweight process that follows: fetch page, discover feed, fetch feed, inspect contents and do stuff with it.

So here we are in 2009 and it feels like deja-vu all over again. This time Dave is promoting rssCloud. rssCloud does a waltz of change notification; the publisher notifies a hub, the hub broadcasts notifications to event subscribers and the subscriber does the same old content fetch and analysis cycle above. rssCloud seems fundamentally flawed in that it is dependent on IP addresses as identifiers. Notification receivers who must use new a IP address must re-subscribe. I'm not sure how an aggregator should distribute notificaton handling load across a pool of IP addresses. The assumption is that notification receivers will have stable, singular IP addresses; rssCloud appears scale limited and a support burden. The focus on specification simplicity has its merits, I think we all hate gratuitous complexity (observe the success of JSON over SOAP). However, Dave doesn't regard system operations as an important concern; he'll readily evangelize a message format and protocol but the real world operability is Other Peoples Problem.

Pubsubhubub (hey, Brad again) has a similar waltz but eliminates the content fetch and analysis cycle ergo it's fundamentally more efficient than rssCloud's waltz. Roughly, Pubsubhubub is to the SixApart atom stream what rssCloud is to old school XML-RPC ping. If I were still at Technorati (or working on event clients like Seesmic, Tweetdeck, etc), I would be implementing Pubsubhubbub and taking a pass on rssCloud. With both systems, I'd be concerned with a few issues. How can the authenticity of the event be trusted? Yes, we all like simplicity but looking at SMTP is apt; now mail systems must be aware of SPF, DKIM, SIDF, blahty blah blah. Its common for mail clients to access MTA's over cryptographically secure connections. Mail is a mess but it works, albeit with a bunch of junk overlaid. I guess this is why Wave Protocol has gathered interest. Anyway, Pubsubhubbub has a handshake during the subscription processes, though that looks like a malicious party could still spin up endpoints with bogus POSTs. I'd like to see an OAuth or digest authentication layer in the ecosystem. Yea, it's a little more complicated, but nothing onerous... suck it up. Whatever. Brad knows the authenticity rap, I mean he also invented OpenID (BTW, we adopted OpenID at Technorati back in 2007). At Technorati we had to implement ping throttling to combat extraneous ping bots; either daemons or cron jobs that just ping continuously, whether the URL has changed or not. You can't just blacklist that URL, we don't know it was the author generating those pings, there's no identity authenticity there. We resorted to blocking IP addresses but that scales poorly and creates other support problems. We had 1000's of domain names blocked and whole class C networks blocked but it was always a wackamole exercise; so much for an open blogosphere, a whitelist is the only thing that scales within those constraints. Meanwhile, back to rssCloud and Pubsubhubbub, what event delivery guarantees can be made? If a data center issue or something else interrupts event flow, are events spooled so that event consumption can be caught up? How can subscribers track the authenticity of the event originators? How can publishers keep track of who their subscribers are? Well, I'm not at Technorati anymore so I'm no longer losing any sleep over these kinds of concerns but do I care about the ecosystem. For more, Matt Mastracci is posting interesting stuff about Pubsubhubbub and a sound comparison was posted to Techcrunch yesterday.

     

( Sep 10 2009, 12:35:37 PM PDT ) Permalink


20090802 Sunday August 02, 2009

Business Models in Consulting, Contracting and Training

Since leaving Technorati last spring, I've been working independently with a few entrepreneurs on their technical platforms. This has mostly entailed working with tools around Infrastructure-As-A-Service (AWS), configuration management (Chef and RightScale), search (Solr) and learning a lot about programming and going into production with ruby and rails. I've spoken to a number of friends and acquaintances who are working as consultants, contractors, technical authors and trainers. Some are working as lone-wolves and others working within or have founded larger organizations. I'm always sniffing for where the upside opportunities are and the question that comes to my mind is: how do such businesses scale?

A number of technology services companies that I've taken notice of have been funded in the last year or so including OpsCode, Reductive Labs, Cloudera and Lucid Imagination. I think all of these guys are in great positions; virtual infrastructure (which is peanut butter to the chocolate of IaaS), big data and information retrieval technologies provide the primordial goo that will support new mobile, real time and social software applications. They are all working in rapid innovation spaces that hold high value potentials but also new learning and implementation challenges that rarefy specialized knowledge.

Years ago when I was working with Covalent Technologies, we tried to build a business around "enterprise open source" with Apache, Tomcat, etc as the basis. Packaging and selling free software is difficult. On the one hand, offering a proven technology stack configuration to overcome the usual integration and deployment challenges as well as providing a support resource is really valuable to Fortune 500's and such. However, my experience there and observations of what's happened with similar businesses (such as Linuxcare and SpikeSource) has left me skeptical how big the opportunity is. After all, while you're competing with the closed-source proprietary software vendors, you're also competing with Free.

The trend I'm noticing is the branching out away from the packaging and phone support and into curriculum. Considering that most institutional software technology education, CS degrees, extended programs, etc have curricula that are perpetually behind the times, it makes sense that the people who possess specialized knowledge on the bleeding edge lead the educational charge. Lucid Imagination, Cloudera and Scale Unlimited are illustrating this point. While on-premise training can be lucrative, I think online courseware may provide a good answer to the business scale question.

For myself, I'm working with and acquiring knowledge in these areas tactically. Whatever my next startup will be, it should be world-changing and lucrative. And I'll likely be using all of these technologies. Thank goodness these guys are training the workforce of tomorrow!

     

( Aug 02 2009, 08:57:08 PM PDT ) Permalink


20090521 Thursday May 21, 2009

Configuring Web Clouds with Chef

I'm not generally passionate about network and system operations, I prefer to focus my attention and creativity on system and software architectures. However, infrastructure provisioning, application deployment, monitoring and maintenance are facts of life for online services. When those basic functions aren't functioning well, then I get passionate about them. When service continuity is impacted and operations staff are overworked, it really bothers me; it tells me that I or other developers I'm working with are doing a poor job of delivering resilient software. I've had many conversations with folks who've accepted as a given that development teams and operations teams have friction between them; some even suggest that they should. After all, so goes that line of thinking, the developers are graded on how rapidly they implement features and fix bugs whereas the operators are graded on service availability and performance. Well, you can sell that all you want but I won't buy it.

In my view, developers need to deliver software that can be operated smoothly and operators need to provide feedback on how smoothly the software is operating; dev and ops must collaborate. I accept as a given that developers

Conversely, I accept as a given that operators I don't want to oversimplify, there's more to the obligations that dev and ops have to each other in order to collaborate effectively. What I've noticed though is that a lot of operators might be skilled with configuring specific server infrastructure or performing OS analysis but configuration management and automation requires using really good tools that they lack. I've seen situations where the available tools are perceived as too complicated and so tools are developed that usually consist of a lot of specialized shell scripts (or perhaps it's just plain old NIH). Cfengine is a good start but the reports I have is that it's difficult to work with and, if you're not very careful, may automatically manage to misconfigure your systems. Puppet was developed to be a more powerful system for configuration management but the feedback I've seen on it is adding new functionality is hard; it has its own configuration language and when you want to extend it you have to deal with a lot of complicated mechanisms. Chef was developed to answer that frustration; by making the configuration language a DSL on top of an already widely used scripting language (ruby), the chef code base provides an easier way to extend it and the chef codebase itself is reportedly an order of magnitude smaller and simpler than puppet's (cavaet: I generally distrust SLoC metrics but just sayin').

So I've been giving Chef a test-drive for this infrastructure-on-EC2 management project that's been cooking. The system implemented the following use cases:

  1. Launch web app servers on EC2 with Apache, Passenger, RoR (+other gems) and overlay a set of rails apps out of git
  2. Launch a pair of reverse proxies (with ha-proxy) in front of the app servers - and reconfigure them when the set of app servers is expanded or contracted
  3. Configure the proxy for failover with heartbeat
  4. Add new rails apps to the set of app servers
  5. Updating/rolling back rails apps
The system is enabled through a combination of the EC2 API (via RightAWS) and Chef's REST API as well as using chef-deploy (think: Capistrano run by a system provisioning agent) to augment Chef's functionality. So far, it seems to be working great!

There's a lot of energy in the Chef community (check out Casserole), combined with monitoring, log management and cloud technologies, I think there's a lot of IT streamlining ahead. Perhaps the old days of labor and communication intensive operations will give way to a new era of autonomic computing. I'll post further about some of the mechanics of working with ruby, rails, chef, EC2, chef-deploy and other tools in the weeks ahead (particularly now that EC2 has native load balancing, monitoring and auto-scaling capabilities). I'll also talk a bit about this stuff at a Velocity BoF. If you're thinking about attending Velocity, O'Reilly is offering 30% off to the first 30 people to register today with the code vel09d30 today (no I'm not getting any kinduva kickback from O'Reilly). And you can catch Infrastructure in the Cloud Era with Adam Jacob (Opscode), Ezra Zygmuntowicz (EngineYard) to learn more about Chef and cloud management.

               

( May 21 2009, 12:30:07 PM PDT ) Permalink
Comments [2]

20090331 Tuesday March 31, 2009

Cloning VMware Machines

I bought a copy of VMware Fusion on special from Smith Micro (icing on the cake: they had a 40% off special that week) specifically so I could simulate a network of machines on my local MacBook Pro. While I've heard good things about Virtual Box, one of the other key capabilities I was looking for from MacIntosh virtualization software was the ability convert an existing Windows installation to a virtual machine. VMware reportedly has the best tools for that kind of thing. I have an aging Dell with an old XP that I'd like to preserve when I finally decide to get rid of the hardware; when it's time to Macify, I'll be good to go.

I started building my virtual network very simply, by creating a CentOS VM. Once I had my first VM running, I figured I could just grow the network from there; I was expecting to find a "clone" item in the Fusion menus but alas, no joy. So, it's time to hack. Looking around at the artifacts that Fusion created, a bunch of files in a directory named for the VM, I started off by copying the directory, the files it contained that had the virtual machine name as components of the file name and edited the metadata files ({vm name}.vmdk/.vmx/.vmxf). Telling Fusion to launch that machine, it prompted if this was a copy or a moved VM - I told it that it was copied and the launch continued. Both launched VM's could ping each other so voila: my virtual network came into existence.

I've since found another procedure to create "linked clones" in VMware Fusion. It looks like this will be really useful for my next scenario of having two different flavors of VM's running on my virtual network. The setup I want to get to is one where I can have "manager" host (to run provisioning, monitoring and other management applications) and cookie-cutter "worker" hosts (webservers, databases, etc). Ultimately, this setup will help me tool up for cloud platform operations; I have more Evil Plans there.

So all of this has me wondering: why doesn't VMware support this natively? Where's that menu option I was looking for? Is there an alternative to this hackery that I just overlooked?

         

( Mar 31 2009, 09:12:22 AM PDT ) Permalink
Comments [2]

20090312 Thursday March 12, 2009

Going to Metallica's Rock and Roll Hall of Fame Induction

Those 25 things you should know about me memes circulating rarely interest me (honestly, I don't care that you have a collection of rare El Salvadoran currency). However, one thing that my friends know but regular readers may not is that I have a fairly eclectic background. Did you know that I used to hang around the art department's hot glass studio in college to blow glass? Did you know that I learned to program in Pascal when I was in college and hated it? Yea, yea, I don't care much anymore either. But anyway, back in the 80's I was friends with this Danish dude from LA who shared my interest in the underground heavy metal scene that was burgeoning, particularly in Britain ("New Wave of British Heavy Metal" AKA NWOBHM) and Europe. We used to trade records and demos (the first Def Leppard 3 song EP on 9" vinyl, I was tired of it so I traded him for a bunch of Tygers of Pan Tang and other crap I didn't own already). I think he, like myself, used to pick up copies of Melody Maker and Sounds at the local record store to read about what was going on overseas. Eventually, Kerrang! came out providing fuller coverage of the metal scene, complete with glossy pictures. But in the meantime back in San Francisco, I helped a friend of mine (Ron Quintana) operate his fanzine Metal Mania (don't be confused, the name was re-appropriated by various larger publishing concerns at different times in the years since but none of them had any relationship to the original gangstas).

Back in the day, Howie Klein was a muckety muck in the music industry, haunting the local clubs like The Old Waldorf and Mabuhay Gardens. Howie hooked us up with a show on KUSF. I dubbed the show Rampage Radio, it ran in the wee hours every Saturday night (right after Big Rick Stuart finished up his late night reggae show with those dudes from Green Apple Records on Clement Street). In between hurling insults at "album oriented rock" and big-hair metal bands (posers!), we played a lot of stuff you couldn't hear anywhere else. Among the many obscure noises we aired were demos from East Bay metalheads Exodus. Amazingly, Rampage Radio is still on the air. Well, that Danish kid and one of the guys I befriended from Exodus were Lars Ulrich and Kirk Hammett, respectively. In short order, they would be playing together in a band Lars named Metallica (after haggling with Ron about not taking that name for the 'zine).

I eventually lost interest in the metal scene (not enough innovation, too much sound-alike derivatives to keep me listening); even though the music from then is still on my playlist, my repertoire has broadened widely (talk to me about gypsy style string jazz, please). I've been peripherally in touch with friends from back then. Over the years, I'd go to a few Metallica shows but the guys are always mobbed at the backstage parties, there's not much of an opportunity to actually talk about anything. Anyway, we have little in common now. I develop software and crazy assed online services; they tour the world to perform in front of throngs. And I don't drink Jaegermeister anymore. In 2000, I introduced one of the friends I've stayed in touch with, Brian Lew, who also had a fanzine Back In The Day, to editors at salon.com (where I was working at the time). He contributed a great article expressing a sentiment that I shared, dismay at Metallica's war on Napster. I don't think I've actually talked to Lars in 15 years. After seeing news coverage of him ranting about how people (his most valued asset: his fans) where ripping him off, I'm not sure I wanted to. But I think we're all over that now, let's just play Rock Band and fuhgedaboutit.

So here we are decades later and Metallica hasn't just warped the music industry, they are the industry. They're up there with Elvis and the Beatles and all of that (except, barring Cliff Burton, they're not dead). Last week, Brian pings me that Q-Prime (Metallica's management company in New York) is trying to reach me. After a few phone calls, it turns out that Metallica is honoring a handful of us old-schoolers by inviting us to a big shindig in Cleveland for their Rock and Roll Hall of Fame induction next month. How cool is that?! I'm still kind blown away that this is really happening (am I being punkd??).

So, I may be leaving Technorati but I'm going to the Rock and Roll Hall of Fame! w00t! That tune keeps humming through my conscience, "...living in sin with a safety pin, Cleveland rocks! Cleveland rocks!" but the way it sounds in my head, it's ganked up, roaring from a massive PA and a wall of Marshall stacks. So now you know what my plans will be in a few weeks and now you've learned a dozen or so things about me (if not 25) that you may not have known before.

           

( Mar 12 2009, 12:22:51 PM PDT ) Permalink


20090310 Tuesday March 10, 2009

More Changes At Technorati (this time, it's personal)

My post last week focused on some of the technology changes that I've been spearheading at Technorati but this time, I have a personal change to discuss. When I joined Technorati in 2004, the old world of the web was in shambles. The 1990's banner-ads-on-a-CPM-basis businesses were collapsed. The editorial teams using big workflow-oriented content management system (CMS) infrastructure (which I worked on in the 90's) were increasingly eclipsed by the ecosystem of blogs. Web 2.0 wasn't yet the word on everyone's lips. But five years ago, Dave Sifry's infectious vision for providing "connective tissue" for the blog ecosystem, tapping the attention signals and creating an emergent distributed meta-CMS helped put it there. Being of service to bloggers just sounded too good, so I jumped aboard.

Through many iterations of blogospheric expansion, building data flow, search and discovery applications, dealing with data center outages (and migrations) and other adventures, it's been a long strange trip. I've made a lot of fantastic friends, contributed a lot of insight and determination and learned a great deal along the way. I am incredibly proud of what we've built over the last five years. However today it's time for me to move on, my last day at Technorati will be next week.

Technorati has a lot of great people, technology and possibilities. The aforementioned crawler rollout provides the technology platform with a better foundation that I'm sure Dorion and the rest of the team will build great things on. The ad platform will create an abundance of valuable opportunities for bloggers and other social media. I know from past experiences what a successful media business looks like and under Richard Jalichandra's leadership, I see all of the right things happening. The ad platform will leverage Technorati's social media data assets with the publisher and advertiser tools that will make Technorati an ad delivery powerhouse. I'm going to remain a friend of the company's and do what I can to help its continued success, but I will be doing so from elsewhere.

I want to take a moment to thank all of my colleagues, past and present, who have worked with me to get Technorati this far. The brainstorms, the hard work, the arguments and the epiphanies have been tremendously valuable to me. Thank You!

I'm not sure what's next for me. I feel strongly that the changes afoot in cloud infrastructure, open source data analytics, real time data stream technologies, location based services (specifically, GPS ubiquity) and improved mobile devices are going to build on Web 2.0. These social and technology shifts will provide primordial goo out of which new innovations will spring. And I intend to build some of them, so brace yourself for Web 3.0. It's times like these when the economy is athrash that the best opportunities emerge and running for cover isn't my style. The next few years will see incumbent players in inefficient markets crumble and more powerful paradigms take their place. I'm bringing my hammer.

( Mar 10 2009, 02:06:20 PM PDT ) Permalink
Comments [4]

20090304 Wednesday March 04, 2009

Welcome to the Technorati Top 100, Mr. President

Since its inception just 6 weeks ago, the White House Blog has had a tremendous center of gravity. I noted the volume of links coming in to the White House Blog after the first week. This is an existential moment for the blogosphere because today the White House Blog has 3216 links from 2323 blogs. And so it's official: the White House Blog has reached the Technorati Top 100.

I find myself reflecting on what the top 100 looked like four years ago, after the prior presidential inauguration, and what it looks like today; the blogosphere is a very different place. Further down memory lane, who recalls when Dave Winer and Instapundit were among the top blogs? Yep, most of the small publishers have been displaced by those with big businesses behind them. Well, at least BoingBoing endures but Huffpo and Gizmo better watch out, here comes Prezbo.

       

( Mar 04 2009, 10:59:16 PM PST ) Permalink


New Crawlers At Technorati

A lot of changes are afoot at Technorati. Over the last year or so, we've been looking inward at the infrastructure and asking ourselves, "How can we do this better?". The data spigot that Technorati builds on was the first thing to focus on, it's a critical part in one leg of the back-end infrastructure tripod. The tripod consists of data acquisition, search and analytics Technorati; while the ping handling and queuing are relatively simple affairs the crawler is the most sophisticated of the data acquisition subsystems. It's proper functioning is critical to the functioning of the other legs; when it doesn't function well, search and analytics don't either (GIGO="garbage in/garbage out").

As Dorion mentioned recently, we're retiring the old crawler. Why are we giving the old crawler getting an engraved watch and showing it to the door? Well, old age is one reason. The original spider is a technology that dates back to 2003, the blogosphere has changed a lot since then and we have a much better developed understanding of the requirements. The original spider code has presented a sufficient number of GIGO-related and code maintenance challenges to warrant a complete re-thinking. It contrasts starkly with the replacement.

Data model
There are a lot of ways to derive structural information out of the pages and feeds that a blog presents. The old spider used event driven parses, building a complex state as it went with flat data structures (lists and hashes). The new one uses the composed web documents to populate a well-defined object model; all crawls normalize the semi-structured data found on the web to that model.
Crawl persistence
The old spider was hard-wired to persist the aforementioned data structure elements to relational databases (sharded MySQL instances) while it was parsing, so that the flow of saving parsed data was closely coupled with parsing events, forsaking transactional integrity and consuming costly resources. The new spider composes and saves its parse result as a big discreet object (not collections of little objects in an RDBMS). This reduced the hardware footprint by an order of magnitude.
Operational visibility
Whether a blog's page structure was understood (or not), the feed was well formed (or not) or any of the many other things that determine the success or quality of a blog's crawl was opaque under the old spider. With the new spider, detailed metadata and metrics are tracked during the crawl cycles. This better enables the team to support bloggers and extend the system's capabilities.
Unit tests
Wherever you have complex, critical software you want to have unit tests. The old spider had almost no unit tests and was developed in a way that made testing the things that mattered most exceptionally difficult. The new spider was developed with a test harness upfront, it now has hundreds of tests that validate thousands of aspects of the code. The test are uniformly invoked by the developers and automatically whenever the code is updated (AKA under continuous integration).
The old spider didn't leverage packages to logically separate the different concerns (fetching, parsing, validation, change determination, etc), the aforementioned flat data structures, mingling of concerns and absence of unit tests made changing it exceedingly difficult. Now, we have a whole that is greater than the sum of the parts; having a well defined data model, sensible persistence, operational visibility and unit tests has added up to an order of magnitude improvement across several dimensions. The real benefits are seen when we've shown that the system is easy to change, I mentioned this several weeks ago when I noted the ease with which we could adapt custom requirements to crawl the White House blog.

Another change that we've made is to the legacy assumption that everything that pings is a blog. That assumption proved to be increasingly untenable as the ping meme spread amongst those who didn't really understand the difference between some random page and a blog, nefarious publishers (spammers) and other perpetrators of spings. Over 90% of the pings hitting Technorati are rejected outright because they've been identified as invalid pings. A large portion of the remainder are later determined to be invalid but we now have a rigorous system in place for filtering out the noise. We've reduced the spam level considerably (as mentioned in a prior post). For instance, there's a whole genre of splogs that are pornography focused (hardcore pictures, paid affiliate links, etc) that previously plagued our data; now we've eliminated a lot of that nonsense from the index.

Here are a pair of charts showing the daily occurrence of a particular porn term in the index.

6 month retrospective as of November 3rd, 2008:
6 month retrospective today, 5 months later:

As you can see, that's an order of magnitude reduction; 90% of the occurrences of that term was spam.

So what's next for the crawler? We've got some stragglers on the old spider, we're going to migrate them over in the next few days. There are still a lot of issues to shake out, as with any new software (for instance, there are still some error recovery scenarios to deal with). But it's getting better all of the time (love that song). We'll be rolling out new tools internally for identifying where improvements are needed, ultimately we'd like to enable bloggers to help themselves to publish, get crawled, be found and recognized more effectively. And there are more changes afoot, stay tuned.

         

( Mar 04 2009, 08:31:16 PM PST ) Permalink
Comments [2]

20090227 Friday February 27, 2009

Code Is Social

Like many software developers, I confess that I have found myself on occasion coming back to old code that I wrote asking, "What was I thinking?" Where it's my code, this WWIT question doesn't happen very often these days but in general I can't count the number of times I've encountered and had to work on code that was not written to be read. I think sometimes programmers write terse code and regard its brevity as a badge of honor, "If I'm wizardly enough to write this, then only True Wizards will read it." Or maybe it's just laziness or hurriedness, these code mysteries are akin to omitting comments, API documentation and other communication artifacts. When I see non-descriptive variable names, gratuitous indirection, excessive right indenting, monkey patching or unnecessary cyclomatic complexity, it's almost anti-social behavior; it's a communication fail more than a functional one. Likewise, gratuitous verbosity stifles communication in the opposite manner; this isn't pre-school - grasp of the ABC's is assumed. So spelling out what code is doing in this belabored fashion is just silly: 

# an array to collect permalinks
permalinks = []
# loop over the feed entries
for entry in feed.entries:
    permalinks.append(entry.link)
Whereas this is clear 
# extract an array of permalinks from the feed entries
permalinks = [ entry.link for entry in feed.entries ]
OK, I'm assuming the reader knows what a Python list comprehension does. The first one is using a lot of vertical space to satisfy a very simple intent. I often find the opposite problem, excessive brevity, is authored by those enamored with their language's idioms. Software written with scripting languages often exhibit this; Perl is famous for expressiveness (I say this lovingly as a repentant x-Perl Wizard) but even the languages with adherents claiming their tongue is more "readable" have those same users donning wizard hats, trying to be clever. Ever tried to maintain Python code riddled with nested list comprehensions containing lambdas? Ruby, similar idiomatic norms abound, 'nuf said.

I've appreciated celebrations of wizardry (see A folding language) but there's more to wizardry than meta-programming and brevity. Coding like a wizard doesn't mean being so clever that only other wizards can collaborate. In my view, a true wizard has the wisdom to steer clear of verbose indulgences and terse spells; the wizard walks the middle path of clarity so that the code is not dumbed-down but the apprentice will grasp the intent. The wizard's code should read as poetry.

When code is unsocial (or anti-social), the quality suffers. Complex software needs a gene pool - lots of eyeballs, lots of variant perspectives and experiences. A small gene pool leads to in-bred ideas. Thus code from a lone wolf (even a kick-ass wolf) will usually be of lower quality than code developed by a plurality (unless its a plurality of novices, then all bets are off).

My plea to fellow crafters of bits: please code for clarity. Don't be so brief that your intentions are unclear. And don't be so garrulous that your intention is lost in the verbiage. Again, I'm not claiming innocence of these sins of code. But over the years I've become considerably more aware of the costs and benefits in the choices between brevity and verbosity. Perhaps clarity is in the eye of beholder or perhaps more narrowly, in the eye of the author. But I try to look at my own code objectively and ask, "If I don't see this code for six months and then come back to it to do some maintenance, will today's intent be clear?" I hope the code I write will be approachable by those who come behind me to work on it, especially if it's me lest I ask myself the WWIT question.

Sigh, I'm venting because I just got side tracked refactoring some program code (and its single test) that lacked clarity. Thanks for indulging me this far. I'm gonna go listen to some old Social Distortion now, have a great weekend!

         

( Feb 27 2009, 06:26:15 PM PST ) Permalink
Comments [1]

20090127 Tuesday January 27, 2009

Inbloguration: One Week Later

As mentioned last week on the Technorati blog, Technorati was crawling the new White House blog within a day of its launch. Most of the blogosphere doesn't require individual customization in our crawling framework but in some special cases, it must be done. The White House blog is a prime example of why this is so but I'm pleased to report that Technorati's new crawling technology makes what was impossible with our old crawler easy with the new one.

Given the volume of moderation that'd be required, it doesn't surprise me that the posts don't take comments. But there are other more basic blogging practices of concern:

The absence of full content feeds or full content on the front page isn't a problem for Technorati's new crawler; it can crawl the post pages and extract content to index from those.

So how has the blog done in the last week? Well, it emerged in the top 1000 just five days after its inception. Keep an eye on the blog's Technorati blog info page - it's currently ranked 882 (1,876 links from 1440 blogs). Many of the links are to main blog page, citing its existence. But the majority of the links were to the kick-off post (date line reads: "Tuesday, January 20th, 2009 at 12:01 pm") and the inaugural address. Here are the posts and their link counts:

It's really encouraging to see the White House, specifically Macon Phillips, posting updates about President Obama's policies and political activities. Going forward, I hope to see more consistent publishing practices. For instance, should slug words be separated by underscores or hyphens? Should the proper names in URL slugs be mixed case (e.g. Timothy_Geithner_sworn_in) or down cased (e.g. now-comes-lilly-ledbetter)? It's less important which one but just pick one and stick to it! Also, the White House Blog should use durable URLs for permalinks: there's a posting with the path /blog_post/PressBriefingHighlights/, another with /president-obama-delivers-your-weekly-address - what are they going to use for the next post with highlights from a press briefing or next week's weekly address? If Mr. Phillips needs any further advise on effective blog publishing technology, I'm easy to find and happy to be of service.

             

( Jan 27 2009, 12:41:39 PM PST ) Permalink


20081225 Thursday December 25, 2008

Downturn, what downturn?

Just the other day, Data Center Knowledge asked Are Colocation Prices Heading Higher? My immediate reaction was, that's a stupid question: last time VC funding went into hibernation, data center space was suddenly cheap and abundant. The article suggested that companies operating their own data centers will run to the colos as a cost cutting measure. Maybe, but I'm not so sure. Data center migrations can be expensive, risky operations. Methinks that the F500's inclined to undertake a migration would have done so already. The article cited a report emphasizing a shift from capital expenses to operating expenses.

Tier 1 says demand for data center space grew 14 percent over the past 12 months, while supply grew by just 6 percent, "exacerbating an already lopsided supply/demand curve."
However, Tier 1 attributed the demand, "especially, (to) the primacy of the Internet as a vehicle for service and application delivery." With the litany of Techcrunch deadpool reports, I'm finding it difficult to believe that the data center space supply/demand will continue skewing.

Sure, it's not all bad news. Fred Wilson reports that Union Square Ventures will be Investing In Thick and Thin. Acknowledging that, "it is easier to invest in thin times. The difficult business climate starts to separate the wheat from the chaff and the strong companies are revealed." Wilson goes on to say

I don't feel that its possible, or wise, or prudent to attempt to time these (venture investment) cycles.
Our approach is to manage a modest amount of capital (in our case less than $300 million across two active funds) and deploy it at roughly $40 million per year, year in and year out no matter what part of the cycle we are in.
That way we'll be putting out money at the top of the market but also at the bottom of the market and also on the way up and the way down. The valuations we pay will average themselves out and this averaging allows us to invest in the underlying value creation process and not in the market per se.
Yes, the economy is gyrating in pain, but the four horsemen aren't galloping nearby. So take a pill, catch your breath and deal with it: the sun will come out, just don't bother trying to time it too carefully.

Now, there's no shortage of reasons for gloom and doom: mega-ponzi schemes collapsing, banks and real estate combusting, the big 3 in various states of failure, yet BMW North America will raise list prices 0.7%. Before the complete credit breakdown, real estate volume was actually rising in a lot of places (ergo: prices were aligning supply and demand). I was at a William-Sonoma store in Albuquerque the other day, the place was mobbed. My point is that while the economy is retrenching (or the country is rebooting), the detritus will be separated (Wilson's wheat from chaff) and data center space should be cheap and abundant. Everything seems fine to me. At least until the next bubble.

For those of you observing that sort of thing, Merry Christmas!

     

( Dec 25 2008, 01:20:14 PM PST ) Permalink


20081223 Tuesday December 23, 2008

Another Hadoop-backended Database: CloudBase

This post to one of the Hadoop mailing lists caught my eye, Announcing CloudBase-1.1 release. Wait, wasn't Cloudbase the embedded database company that IBM acquired several years back but ended up donating the product to the Apache Software Foundation as Derby? No, not that Cloudbase. This is apparently another project that aims to provide data warehousing on top of Hadoop.

I've been watching the emergence of HBase, Hypertable and most recently the proposed incubation of Facebook's Cassandra with great interest. The first two are modeled from Google's BigTable but all are essentially horizontally scalable column oriented databases. The developers of these systems explicitly steer away having their technologies pegged as relational databases, with the refrain: "We don't do joins." What the CloudBase project aims to do is not model themselves on BigTable but to explicitly support joins between tables built on top of an HDFS cluster. It looks like they've posted extensive documentation and have released a JDBC driver, pretty cool! This is the most interesting database initiative I've seen since GreenPlum announced their support for mapreduce.

Yes, as far as scale-out data analytics, we live in interesting times.

                 

( Dec 23 2008, 04:02:21 PM PST ) Permalink
Comments [2]

20081221 Sunday December 21, 2008

Python SVN Bindings, Trac and mod_python

I have some code I want to noodle on outside of work. Since I'm on a holiday break, I'm doing a bit of that (yes, this is what I do for fun, so?). In the past, I had used my own private CVS server for those kinds of things but these days, I could just as well live without CVS. I decided to roll a subversion server into my Apache build (the latest Apache + other modules aren't in the yum repositories for my distro, so I roll my own). While I'm putting a subversion server up, why not trac, too? Heh, that's where things got stuck.

When I installed the subversion dependencies (specifically, neon), I just used vanilla build params. After installing subversion, I was surprised that Trac couldn't access it. It turns out that the litmus test was this: 

$ python
Python ...
>>> from svn import core
...it failed miserably. Various recompile efforts seemed to move the problem around. I saw a variety of the symptoms described in the Trac-Subversion integration docs troubleshooting section. The missing gss_delete_sec_context symbol error was apparently the telltale critical one, it originated from neon having been compiled without SSL support. The neon compile config that led to success was 
./configure --enable-shared --enable-static --with-ssl=openssl
Then the real key was to completely start over with the subversion compile, not just the swig python bindings. 
make clean
./configure \
 --with-berkeley-db=/usr/local/BerkeleyDB.4.7 \
 --prefix=/usr/local \ 
 --with-apxs=/usr/local/httpd2.2.11/bin/apxs \
 --with-apr=/usr/local/apr \
 --with-apr-util=/usr/local/apr
make
make swig-py
make check-swig-py
make install
make install-swig-py
ldconfig
Only then did the litmus test above pass. One of the things about this setup that is kind of a nuisance is that the python bindings didn't get installed into site-packages, therefore mod_python was quite unhappy. Also, trac seemed to want to put its eggs in the root directory. So the Apache server's envvars script has these variables exported to work around those issues 
PYTHONPATH=/usr/local/lib/svn-python
LD_LIBRARY_PATH=/usr/local/lib/svn-python/libsvn
PYTHON_EGG_CACHE=/data1/egg_cache
The result (including the requisite httpd.conf tweaks) is a working subversion 1.54 and trac 0.11 setup. It was more fiddling for the evening than I'd hoped for and I'm not sure my foibles and remedies were optimal (clearly, I missed an RTFM somewhere) but I hope this resolution helps at least one reader.

Happy Hannukah and winter solstice!

             

( Dec 21 2008, 02:55:24 PM PST ) Permalink