A Short Script for Logging into Interactive Elastic MapReduce Clusters

Elastic MapReduce is great, but the latencies can be painful.  For me, this is especially true when I'm in the early stages of developing a new job and need to make the transition from code on my local machine to code running in the cloud -- the ~5 minute period between starting up a cluster and actually being able to log on to it is too long to sit there staring at a blank screen and too short to effectively context switch to something else in a useful way.

My current solution is to allow myself to get distracted but to drag myself back to my EMR session as soon as it's available.  Adding some simple polling plus a sticky growl notification to my interactive-emr-startup script does the trick quite nicely:


#!/bin/bash


if [ -z "$1" ]; then
  echo "Please specify a job name"
  exit 1
fi


elastic-mapreduce \
  (... with all of my favorite options ...) \
| tee ${TMP_FILE}


JOB_ID=`cat ${TMP_FILE} | awk '{print $4}'`
rm ${TMP_FILE}


# poll for WAITING state
JOB_STATE=''
MASTER_HOSTNAME=''
while [ "${JOB_STATE}" != "WAITING" ]; do
  sleep 1
  echo -n .
  RESULT=`elastic-mapreduce --list | grep ${JOB_ID}`
  JOB_STATE=`echo $RESULT | awk '{print $2}'`
  MASTER_HOSTNAME=`echo $RESULT | awk '{print $3}'`
done
echo Connecting to ${MASTER_HOSTNAME}...


growlnotify -n "EMR Interactive" -s -m "SSHing into ${MASTER_HOSTNAME}"


ssh $MASTER_HOSTNAME -i ~/.ssh/emr-keypair -l hadoop -L 9100:localhost:9100

One of my personal productivity goals for the year is finding little places like this that I can optimize with a short script.  This particular one has rescued me from the clutches of HN more than once!

On Code Reviews and Developer Feedback

There's a great post from last week at 37signals, Give it five minutes:

While he was making his points on stage, I was taking an inventory of the things I didn’t agree with. And when presented with an opportunity to speak with him, I quickly pushed back at some of his ideas. I must have seemed like such an asshole.

His response changed my life. It was a simple thing. He said “Man, give it five minutes.” I asked him what he meant by that? He said, it’s fine to disagree, it’s fine to push back, it’s great to have strong opinions and beliefs, but give my ideas some time to set in before you’re sure you want to argue against them. “Five minutes” represented “think”, not react. He was totally right. I came into the discussion looking to prove something, not learn something.

There’s also a difference between asking questions and pushing back. Pushing back means you already think you know. Asking questions means you want to know. Ask more questions.

This is such a great outlook and a great way to approach the discussion of feedback for code reviews and design reviews.

It's surprising how little time development teams devote to training, or even internal discussion on effective feedback. As developers, we are constantly engaged in this kind of communication: white-boarding sessions, spec reviews, design reviews, code reviews. We're expected to give and receive feedback on a daily basis, but few of us are properly prepared for it. Not only do we lack the training, but we have many negative examples to draw from. Who hasn't been a part of a design review where tempers flare? Properly giving feedback is something that requires constant attention and practice. Receiving feedback can be just as difficult.

Culture of Communication

One of the major pillars of our engineering culture at bizo is "the 3 Cs": Communication, Communication, Communication.

We've tried hard to build a team of engineers that are eager to receive feedback, humble about their abilities, objective and gracious with their feedback, and freely giving of their own knowledge and experience. We see communication as a prerequisite for building a world-class team and developing high-quality code. You often hear the phrase "strong opinions, weakly held," and that is the kind of culture we have tried to build.

Communication is hard. It takes real team agreement and commitment to continued work to keep this culture alive and well. It's important the team views effective communication as important and that the culture supports it.

Code Reviews

Code reviews are something that can easily be approached from the wrong perspective, both as an author or reviewer.

As a reviewer, it can be easy to jump in and argue, to try and push 'your' solution (even though it may be equivalent), to push back instead of asking questions and trying to understand.

As an author, it's far too easy to get attached to your code, to your specific solution/naming/etc. It's also easy to feel like each comment is an attack on your ability, and that by accepting the feedback, this somehow means that you were wrong or did a bad job. Of course, nothing could be further from the truth!

At Bizo, we perform code reviews for every change. They are a major part of our culture of communication. In order to perform effective code reviews, it's important to have some shared guidelines that help support effective communication.

Here are some guidelines we've found to be helpful for performing code reviews:

What is a code review

• A careful line-by-line critique of code by peers
• happens in a non-threatening context
• goal is cooperation and mutual learning, not fault finding

Code reviews are a team exercise to improve understanding and make the code better!

When people think of code reviews they usually think of catching bugs. Code reviews do occasionally catch bugs or potential performance problems, but this is rare.

Just as important is fostering a shared understanding of the code and exposure to new approaches, techniques, and patterns. Seeing how your peers program is a great way to learn from them.

Ensuring coding standards and style guides is another way code reviews help. Working on a team it's important to keep readability and quality high using a shared vocabulary.

As an author

As an author, it's important to view each comment as a new opportunity to improve your code. Instead of jumping into defense mode, take a step back and think. Try to approach the code again for the first time with this new perspective. Your team has a lot of experience and varied backgrounds -- draw from them! They are there to help you. Use the gift of their experience and knowledge to improve the code.

Trust the team, and view all comments as action items. Some changes can seem arbitrary, especially when it comes to naming and organization. Unless there's a strong reason, tend to agree with your reviewers. If a reviewer finds something confusing, it is confusing! Code spends most of its life in maintenance and programming is a team sport. Remember that they are your audience, and you want them to be able to understand your code at 4am after a system crash.

As a reviewer

As a reviewer, it's important to take the time to understand the code, think, and ask questions to understand the code before providing feedback. The author probably spent a lot more time thinking about the problem and the approach over the course of the project.

Be strict on coding standard and style guide violations. The real cost of software is maintenance (80% according to Sun). It's important the code is easily understood by the team.

Be gentle on personal preferences. If it's not a standard violation and just a matter of personal preference, defer to the author. It's okay to present your perspective, but mention that it's just a preference and not meant to be taken as an action item.

Trust the author. It's often the case that there are many valid approaches to a problem. It's great to present alternative approaches and discuss pros/cons of various approaches. If you see alternative solutions, bring them up! When discussing alternatives, make sure to listen to the author. Remember they are the subject matter expert and you are working together on the same team.

It takes work!

Communication is hard! It's easy to screw-up. It's easy to go into attack or defense mode when you're passionate about what you're doing. It's really something we all need to remind ourselves to work on every day. It's something we need to periodically remind ourselves as a team. Try to view each review as an opportunity to practice these guidelines. Just remember to take a step back, think, and ask questions.

Fault Tolerant MongoDB on EC2

While working on a project at Bizo I needed to connect a Rails app to a MongoDB backend both of which run in Amazon's Cloud (EC2). At Bizo we have a policy to not use non Amazon services when possible (to limit risk) - so we normally run most of our services straight off of EC2. I'd like to share what I've learned as best practices throughout the experience as I hope it might save some time and frustration for others.

Primer

Replica sets are the preferred way to run a distributed, fault tolerant MongoDB service. But as with any distributed system, nodes will eventually fail. Now replica sets are pretty good at handling failures, but they can't save you if too many nodes fail.
Specifically a replica set requires a minimum of two nodes to function at all times (1 primary and 1 secondary node). Thus a good rule of thumb is to run **at least 3 nodes** in a replica set, that way if a node fails your database service doesn't go down with it. The Rails app I was working with doesn't experience enormous amounts of traffic so 3 m1.large (64bit) nodes were sufficient for my needs. What follows is a rundown of our setup and how it handles common needs of fault tolerant systems.

Best Practices

Minimize Failure with AutoScaling, Availability Zones, CloudWatch and EBS Volumes

• Use Autoscaling Groups, CloudWatch and EBS Volumes to replace failed nodes as soon as they go down. Since we run three nodes, our replica set is insulated from failure due to a single node crashing. But if two nodes crash the replica set goes with them. To solve this we use Cloudwatch alarms to trigger the Autoscaling Group whenever a node goes down - that way a new replacement node is automatically brought online within a few minutes of a failure to reduce the risk of nodes sequentially failing. Additionally each node stores it's data on an EBS Volume (network attachable hard drive) - that way when a node fails, it's replacement doesn't startup with missing data - it simply mounts the previous node's EBS.
• To protect against multiple nodes failing simultaneously run each node in a separate availability zone. The above isn't sufficient to protect against things like hardware failures as all 3 instances could wind up on the same hardware. Running each node in a separate availability zone guarantees that our mongo instances run with a reasonable amount of separation (eg. they don't all end up on the same hardware box). Ideally you'd run each node in its own region (separate data center), but this causes headaches trying to configure firewalls as Amazon does not allow security groups to be used across multiple regions (see security below). So unless you want to setup a VPN for cross region communication - you're probably better off just running in separate availability zones.
  Assuming you've created the group and are running three nodes, each in a separate availability zone you can configure the auto scaling group using Amazon's command line tools like so:
  

  
  as-update-auto-scaling-group my-mongo-service \
  --region us-east-1 \
  --availability-zones us-east-1a us-east-1b us-east-1c \
  --max-size 3 \
  --min-size 3 \
  --desired-capacity 3
  
  

  Now if any node fails then a new one will startup to take its place in the proper availability zone.
  

If Everything Fails, have backups

It's always good to have backups just in case something really bad happens. Fortunately since we use EBS Volumes this is really easy - we create nightly snapshots of the primary node's EBS Volume on our cron server using Amazon's command line tools (ec2-create-snapshot). These snapshots are persisted to S3 and we can easily restore our replica set from these backups.

Use Elastic IPs

As nodes fail and are replaced you want both your Replica set and your database clients to be able to find and connect to the new nodes. The easiest way to do this in Amazon is to use Elastic IPs - special* static ip addresses that can be assigned to individual instances. Since each instance runs in a separate availability zone we just need one Elastic IP per zone. When a new node starts up to replace a failed instance, it checks which zone it was started in and assigns itself the matching Elastic IP. Both the client and replica set configuration should point at the Elastic IP address - that way failures and startups of new nodes will be seamless to your app. This is because cross-security group openings in the firewall need to use internal (not external) addresses and the DNS url in the console will resolve to an internal ip address from an EC2 instance.

Security

This is where the headaches can start. Ideally you want to restrict access to your MongoDB instances to just your client application using Amazon's security groups. The way we normally set this up is to give your mongo instances a security group, say mongo-db-prod and your client app a security group, say cool-app-prod. Then mongo-db-prod would grant access on port 27017 (default mongodb port) to security group: cool-app-prod. Unfortunately what's not documented very well is that if you use the **external Elastic IP** addresses in your configuration it **will not work** with security groups! Instead you have to use the Elastic IPS DNS url (found in Amazon's web console) for security groups to work properly.

A Final Caveat

One thing to be careful of is if you require more than 5 nodes in a replica set you'll run into a problem using Elastic IPs - Amazon by default only allows 5 eips per region. You'll need to either ask Amazon to increase this limit on your account or seek out an alternative setup.
Well, that's it but If you have another setup for running MongoDB on EC2 I'd love to here it. Until next time.

Building a Product in Just 8 Hours

Recently at Bizo, we decided to try a new kind of hack day. Previously during hackdays our engineers worked individually on their own project(s). But on our last hack day we decided to try something new – The 8 Hour Product Challenge.
We would build and launch a completely new product in the course of a normal workday (9-5pm). “Launching” meant this product had to be running publicly on the internet by 5pm – no excuses, no “Wait! I need 5 more minutes” – whatever was there had to be deployed. In short the experience was fantastic and I can’t wait to do it again. Here’s a breakdown of the experience:

Initial Meeting 9:30am

Organizing developers for a meeting of any kind is like trying to heard cats. But if it’s a meeting before 11:00am you’re not herding regular cats, you’re herding sleepy, fat cats with one leg and half an ear. Somehow after a lot of cattle prodding by our VP of engineering our team eventually managed to shuffle its way into the conference room like the decaffeinated zombies we were and get to work.
We decided to build a stealth product. The system would use Bizo’s rich business data to personalize special content for visitors based on things like their industry, company size and seniority. The goal for the end of the day was to have a small webapp up and running on Amazon’s servers.
After a bit more discussion, we decided to split tasks up into five groups of two engineers:

• Data Discovery Team – Find relevant items for users by using our B2B business data & network.
• Scraping Team – given a url representing an item scrape the page contents and store them for later use
• Data Classification Team – extract relevant data from the HTML source of the previously scraped urls.
• Backend Team – backend architecture for the webapp that fetches serves the content from that was generated in the previous steps
• Frontend Team – frontend design + javascript that makes the app functional

Engineers were assigned more or less randomly, with the exception of myself – I was assigned to the frontend team directly. During the course of our initial planning meeting, we (myself included) often found ourselves becoming sidetracked with feature bloat, premature scalability concerns and a myriad of other things not essential to our MVP. Fortunately one of my coworkers (Stephen) was smart enough to enforce timeboxing the meeting to one hour – eventually we got things back on topic and designed the critical components before time ran out.

Start Work 10:30am

Once the teams were assigned, we all jumped in and started to work on our relevant tasks. My partner, Darren and I immediately started out by sketching ideas on paper for our design. I can’t stress enough how important sketching is for being able to rapidly prototype a product – a trick I picked up back when I interned over at ZURB. Only after we had some solid sketches did we move into Photoshop mockups. Meanwhile the other teams were all furiously programming their parts of the application:

• Data Discovery Team Worked out a simple ranking algorithm for data and started writing the Hive script to extract it.
• URL Scraping Team Started out in Scala hacking up a script to scrape & download url content.
• Data Extraction Team Decided to try out the Pismo gem to extract summaries and titles from scraped html content.
• Backend Team Was working on getting a sweet Scalatra webapp up and running

Lunch Time & Status Updates 12:30pm

By lunchtime everything seemed to be coming along nicely. On the frontend had completed our Photoshop mockup and had just began writing some basic css styles. All the other teams reported making good progress on their tasks, with no major snags in the foreseeable future (betcha you never heard that one before…).

Afternoon 1:30pm

My frontend partner and I powered through our post lunch food coma and were able to move into begin wiring up the ui using CoffeeScript in conjunction with Dependence.js. My teammate and I decided to give pair programming a shot. He has always been more of an Emacs kind of guy, while I prefer vi, but in the interests of learning I decided to try Emacs for the rest of the day – I now know why he’s always so worried about contracting carpel tunnel syndrome :).
Using some sample data generated by the first three teams we were able to get a rough ui working pretty quickly. Our side of things turned out to be pretty straightforward and involved three AJAX requests. One was to retrieve a list of items grouped by segment from the Scalatra web server, one to get a list of the current targetable segments from the Bizo API and another call to the API to retrieve a visitor’s business segments (bizographics).
The only snag we hit was running into a race condition – originally we attempted executing all three requests simultaneously. In reality we had to wait to get the list of segments before getting the visitor’s profile. Darren and I just looked at each other and shrugged, then we indented the third API request a few spaces in our CoffeeScript code – race condition solved! Yes that’s correct we fixed a race condition by indenting some code, don’t judge – it was a hackday.
The other teams all seemed to be doing well, the scraping team discovered the Scala Collection’s magical par(), which turns normal data structures into parallel ones, they almost peed their pants with joy. At this point the backend team had completed the Scalatra app and was working on setting up our eventual deployment to EC2 using our custom infrastructure, cowboy.

The Home Stretch & Deployment 4:30pm

Right around 4:30 we ran into a major problem. There had been a miscommunication regarding the necessary format of the JSON file needed by the frontend, and our data was coming through to the app in a format that just wouldn’t work. We had to scramble and hash things out with the other teams quickly before we hit our 5pm deadline. Thanks to a major push by the Data Extraction Team we were finally able to get everything in place. The product worked! – it wasn’t the most polished app ever, but what we had accomplished in just one day was pretty amazing. The product was deployed on EC2 and presented internally within Bizo – it was met with a lot of excitement and Bizo will probably be releasing it publicly in the weeks to come.

Closing Thoughts

Overall the experiment turned out to be a smash hit. Looking back on the experience there are a bunch of things we could have done better. In retrospect we got sidetracked a bit too much on non essential feature ideas when we really should have been spending time clarifying the format of the data as it passed through each team’s project – this was something that came back and bit us near the end of the day. But mishaps aside it’s really amazing what you can accomplish with 9 other talented people in a single day.
I can’t recommend trying this with your team enough – what do you think, is your engineering team up for The 8 Hour Product Challenge? If you’re interested in the technologies we used throughout the day, see below.

Appendix, Technologies Used

• News Discovery Team – Hive, Amazon EC2, Amazon S3
• URL Scraping Team – Scala
• Data Extraction Team – jRuby, gems of note: pismo, right_aws (for Amazon S3)
• Backend Team – Scalatra, Amazon S3
• Frontend Team – Photoshop, HTML5, Sass, CoffeeScript, jQuery, dependence.js

work at Bizo (looking for some good engineers)

We’re a small, disciplined team that gets a lot done. Our platform processes billions of page views monthly and 100s of terabytes of data so we have lots of fun problems to tackle. We believe in teamwork and communication: comments, design reviews, code reviews for every change, weekly tech talks. We believe in giving developers ownership over projects. We believe Engineering is more than coding. We have fun and keep the beer fridge well stocked.

We have customers, are well funded and recently named the forth fastest growing private company in the San Francisco Bay Area.

We are looking for motivated problem solvers with an entrepreneurial / hacker spirit.

If you're a reader of this blog, you already know our technology stack. Some highlights: Scala, Java, Javscript, Ruby, AWS (pretty much every service), Hadoop/Hive, GWT, MongoDB, Solr, etc.

If you're interested, please apply on stackoverflow.

Using GenericUDFs to return multiple values in Apache Hive

A basic user defined function (UDF) in Hive is very easy to write: you simply subclass org.apache.hadoop.hive.ql.exec.UDF and implement an evaluate method.  We've previously written about this strategy, and it works well for most simple cases.

The first case where this breaks down is when you want to return multiple values from your UDF.  For me, this often arises when we have serialized data stored in a single Hive field and want to extract multiple pieces of information from it.

For example, suppose we have a simple Person object (leaving out all of the error checking code):

case class Person(val firstName: String, val lastName: String)


object Person {
  def serialize(p: Person): String = {
    p.firstName + "|" + p.lastName
  }


  def deserialize(s: String): Person = {
    val parts = s.split("|")
    Person(parts(0), parts(1))
  }
}

We want to convert a data table containing these serialized objects into one containing firstName and lastName columns.

create table input(serializedPerson string) ;

load data local inpath ... ;

create table output(firstName string, lastName string) ;


So, what should our UDF and query look like?


Using the previous strategy, we could create two separate UDFs:


insert overwrite table output
select firstName(serializedPerson), lastName(serializedPerson)
from input ;


Unfortunately, the two invocations will have to separately deserialize their inputs, which could be expensive in less trivial examples.  It also requires writing two separate implementation classes whose only difference is which field to pull out of your model object.


An alternative is to use a GenericUDF and return a struct instead of a simple string.  This requires using object inspectors to specify the input and output types, just like in a UDTF:


class DeserializePerson extends GenericUDF {
  private var inputInspector: PrimitiveObjectInspector = _

  def initialize(inputs: Array[ObjectInspector]): StructObjectInspector = {

    this.inputInspector = inputs(0).asInstanceOf[PrimitiveObjectInspector]

    val stringOI =    PrimitiveObjectInspectorFactory.getPrimitiveJavaObjectInspector(STRING)

    val outputFieldNames = Seq("firstName", "lastName")

    val outputInspectors = Seq(stringOI, stringOI)

    ObjectInspectorFactory.getStandardStructObjectInspector(outputFieldNames,

      outputInspectors)

  }

  def getDisplayString(children: Array[String]): String = {

    "deserialize(" + children.mkString(",") + ")"

  }

  def evaluate(args: Array[DeferredObject]): Object = {

    val input = inputInspector.getPrimitiveJavaObject(args(0).get)

    val person = Person.deserialize(input.asInstanceOf[String])

    Array(person.firstName, person.lastName)

  }

}


Here, we're specifying that we expect a single primitive object inspector as an input (error handling code omitted) and returning a struct containing two fields, both of which are strings.  We can now use the following query:


create temporary function deserializePerson as 'com.bizo.udf.DeserializePerson' ;


insert overwrite table output
select person.firstName, person.lastName
from (
  select deserializePerson(serializedPerson)
  from input
) parsed ;


This query deserializes the person only once but gives you access to both of the values returned by the UDF.


Note that this method does not allow you to return multiple rows -- for that, you still need to use a UDTF.

Clustering of sparse data using python with scikit-learn

Coming from a Matlab background, I found sparse matrices to be easy to use and well integrated into the language. However, when transitioning to python’s scientific computing ecosystem, I had a harder time using sparse matrices. This post is intended to help Matlab refugees and those interested in using sparse matricies in python (in particular, for clustering)

Requirements:

scikit-learn (2.10+)

numpy (refer to scikit-learn version requirements)
scipy (refer to scikit-learn version requirements)

Sparse Matrix Types:
There are six types of sparse matrices implemented under scipy:

1. bsr_matrix -- block sparse row matrix
2. coo_matrix -- sparse matrix in coordinate format
3. csc_matrix -- compressed sparse column matrix
4. csr_matrix -- compressed sparse row matrix
5. dia_matrix -- sparse matrix with diagonal storage
6. dok_matrix -- dictionary of keys based sparse matrix
7. lil_matrix -- row-based linked list sparse matrix

For more info see: (http://docs.scipy.org/doc/scipy/reference/sparse.html):

When to use which matrix:
The following are scenarios when you would want to choose one sparse matrix type over the another:

• Fast Arithmetic Operation: csc_matrix, csr_matrix
• Fast Column Slicing (e.g., A[:, 1:2]): csc_matrix
• Fast Row Slicing (e.g., A[1:2, :]) csr_matrix
• Fast Matrix vector products: csr_matrix, bsr_matrix, csc_matrix
• Fast Changing of sparsity (e.g., adding entries to matrix): lil_matrix, dok_matrix
• Fast conversion to other sparse formats: coo_matrix
• Constructing Large Sparse Matrices: coo_matrix

Clustering with scikit-learn:

With the release of scikit-learn 2.10, one of the useful new features is the support for sparse matrices with the k-means algorithm. The following is how you would use sparse matrices with k-means:

Full Matrix to Sparse Matrix

Example-1

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

from numpy.random import random

from scipy.sparse import *

from sklearn.cluster import KMeans

# create a 30x1000 dense matrix random matrix.

D = random((30,1000))

# keep entries with value < 0.10 (10% of entries in matrix will be non-zero)

# X is a "full" matrix that is intrinsically sparse.

X = D*(D<0.10) # note: element wise mult

# convert D into a sparse matrix (type coo_matrix)

# note: we can initialize any type of sparse matrix.

# There is no particular motivation behind using

# coo_matrix for this example.

S = coo_matrix(X)

labeler = KMeans(k=3)

# convert coo to csr format

# note: Kmeans currently only works with CSR type sparse matrix

labeler.fit(S.tocsr())

# print cluster assignments for each row

for (row, label) in enumerate(labeler.labels_):

print "row %d has label %d"%(row, label)

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

One of the issues with Example-1 is that we are constructing a sparse matrix from a full matrix. It will often be the case that we will not be able to fit a full (although intrinsically sparse) matrix in memory. For example, if the matrix X was a 100000x1000000000 full matrix, there could be some issues. One solution to this is to somehow extract out the non-zero entries of X and to use a smarter constructor for the sparse matrix.

Sparse Matrix Construction

In Example-2, we will assume that we have X's data stored on some file on disk. In particular, we will assume that X is stored in a csv file and that we are able to extract out the non-zero data efficiently.

Example-2

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

import csv
from scipy.sparse import *
from sklearn.cluster import KMeans

def extract_nonzero(fname):
"""
extracts nonzero entries from a csv file
input: fname (str) -- path to csv file
output: generator<(int, int, float)> -- generator
producing 3-tuple containing (row-index, column-index, data)
"""
for (rindex,row) in enumerate(csv.reader(open(fname))):
for (cindex, data) in enumerate(row):
if data!="0":
yield (rindex, cindex, float(data))

def get_dimensions(fname):
"""
determines the dimension of a csv file
input: fname (str) -- path to csv file
output: (nrows, ncols) -- tuple containing row x col data
"""
rowgen = (row for row in csv.reader(open(fname)))
# compute col size
colsize = len(rowgen.next())
# compute row size
rowsize = 1 + sum(1 for row in rowgen)
return (rowsize, colsize)

# obtain dimensions of data
(rdim, cdim) = get_dimensions("X.csv")

# allocate a lil_matrix of size (rdim by cdim)
# note: lil_matrix is used since we be modifying
# the matrix a lot.
S = lil_matrix((rdim, cdim))

# add data to S
for (i,j,d) in extract_nonzero("X.csv"):
S[i,j] = d

# perform clustering
labeler = KMeans(k=3)
# convert lil to csr format
# note: Kmeans currently only works with CSR type sparse matrix
labeler.fit(S.tocsr())

# print cluster assignments for each row
for (row, label) in enumerate(labeler.labels_):
print "row %d has label %d"%(row, label)

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

What to do when Sparse Matrices aren't supported:

When sparse matrices aren't supported, one solution is to convert the matrix to a full matrix. To do this, simply invoke the todense() method.