# Add-on hangs: a brief summary

I have blogged about clustering BHR hangs before. This post is dedicated to the add-on side of things.

In a way BHR could be seen as a distributed profiler: each user runs a local profiler that samples the stack of a thread only when that thread is hanging for over N milliseconds. Then the stacks are sent to our backend infrastructure through a Telemetry submission.

Imagine to profile Firefox for an hour. At the end of the hour you would like to determine the impact of one of your installed add-ons on the browsing experience. Not simple at all, the add-on might have been in use only for a fraction of your session, say 20 seconds. In that fraction it might have slowed down the browser significantly though. Since you are collecting hangs for the whole session, that signal might eventually be dominated by noise. This means that in most cases, add-on hangs are not going to look like a big deal once aggregated.

I aggregated the stacks and partitioned them in browser and add-on specific ones. For a given add-on Foo, I computed for all sessions the ratio of hangs of Foo over the number of hangs of both Firefox and Foo. Finally I averaged those ratios. This final number gives an idea of the average proportion of hangs due to Foo an user of that add-on can expect to find in his session.

That’s not the whole story though, one can imagine scenarios where an add-on triggers an asynchrounous workload in the browser which will not be accounted to the add-on itself, like garbage collection. In a grantedly less common, but still plausible scenario, an add-on could improve the performances of the browser and in doing so reducing the number of browser specific hangs while increasing the ratio. In general I tend to see the final number as a lower bound and even though it’s not precise, it can help identify bottlenecks.

From the most popular add-ons the ones that have a high ratio are:

1. Lastpass (12%)
2. Hola Better Internet (11%)
3. Avira Antivirus (10%)
4. noscript (9%)

LastPass, for instance, has a ratio of hangs of about 12%, which means that if you had just LastPass installed and no other add-on, on average about 12% of the hangs you would experience would likely be due to LastPass. That’s a lot and and the main culprit seems to be the logic to scan for input fields when the document changes dynamically. That said I am glad I can’t see any popular add-ons with a shockingly high ratio of hangs, which is good.

The numbers shouldn’t be used to mark an add-on as bad or good; they are based on a fallible heuristic and they are meant to give us the tools to prioritize which add-ons we should keep an eye on.

# Next-gen Data Analysis Framework for Telemetry

The easier it is to get answers, the more questions will be asked

In that spirit me and Mark Reid have been working for a while now on a new analysis infrastracture to make it as easy as possible for engineers to get answers to data related questions.

Our shiny new analysis infrastructure is based primarily on IPython and Spark. I blogged about Spark before, I even gave a short tutorial on it at our last workweek in Portland (slides and tutorial); IPython might be something you are not familiar with unless you have a background in science. In a nutshell it’s a browser-based notebook with support for code, text, mathematical expressions, inline plots and other rich media.

The combination of IPython and Spark allows to write data analyses interactively from a browser and seemingly parallelize them over multiple machines thanks to a rich API with over 80 distributed operators! It’s a huge leap forward in terms of productivity compared to traditional batch oriented map-reduce frameworks. An IPython notebook contains both the code and the product of the execution of that code, like plots. Once executed, a notebook can simply be serialized and uploaded to Github. Then, thanks to nbviewer, it can be visualized and shared among colleagues.

In fact, the issue with sharing just the end product of an analysis is that it’s all too easy for bugs to creep in or to make wrong assumptions. If your end result is a plot, how do you test it? How do you know that what you are looking at does actually reflect the truth? Having the code side by side with its evaluation allows more people to  inspect it and streamlines the review process.

This is what you need to do to start your IPython backed Spark cluster with access to Telemetry data:

1. Visit the analysis provisioning dashboard at telemetry-dash.mozilla.org and sign in using Persona with an @mozilla.com email address.
2. Click “Launch an ad-hoc Spark cluster”.
3. Enter some details:
• The “Cluster Name” field should be a short descriptive name, like “chromehangs analysis”.
• Set the number of workers for the cluster. Please keep in mind to use resources sparingly; use a single worker to write and debug your job.
4. Click “Submit”. 
5. A cluster will be launched on AWS preconfigured with Spark, IPython and some handy data analysis libraries like pandas and matplotlib.

Once the cluster is ready, you can tunnel IPython through SSH by following the instructions on the dashboard, e.g.:

ssh -i my-private-key -L 8888:localhost:8888 hadoop@ec2-54-70-129-221.us-west-2.compute.amazonaws.com


Finally, you can launch IPython in Firefox by visiting http://localhost:8888.

Now what? Glad you asked. In your notebook listing you will see a Hello World notebook. It’s a very simple analysis that produces the distribution of startup times faceted by operating system for a small fraction of Telemetry submissions; let’s quickly review it here.

We start by importing a telemetry utility to fetch pings and some commonly needed libraries for analysis: a json parser, pandas and matplotlib.

import ujson as json
import matplotlib.pyplot as plt
import pandas as pd
from moztelemetry.spark import get_pings


To execute a block of code in IPython, aka cell, press Shift-Enter. While a cell is being executed, a gray circle will appear in the upper right border of the notebook. When the circle is full, your code is being executed by the IPython kernel; when only the borders of the circle are visible then the kernel is idle and waiting for commands.

Spark exploits parallelism across all cores of your cluster. To see the degree of parallelism you have at your disposal simply yield:


sc.defaultParallelism



Now, let’s fetch a set of telemetry submissions and load it in a RDD using the get_pings utility function from the moztelemetry library:

pings = get_pings(sc,
appName="Firefox",
channel="nightly",
version="*",
buildid="*",
submission_date="20141208",
fraction=0.1)


That’s pretty much self documenting. The fraction parameter, which defaults to 1, selects a random subset of the selected submissions. This comes in handy when you first write your analysis and don’t need to load lots of data to test and debug it.

Note that both the buildid and submission_date parameters accept also a tuple specifying, inclusively, a range of dates, e.g.:


pings = get_pings(sc,
appName="Firefox",
channel="nightly",
version="*",
buildid=("20141201", "20141202"),
submission_date=("20141202", ""20141208"))


Let’s do something with those pings. Since we are interested in the distribution of the startup time of Firefox faceted by operating system, let’s extract the needed fields from our submissions:

def extract(ping):
os = ping["info"]["OS"]
startup = ping["simpleMeasurements"].get("firstPaint", -1)
return (os, startup)

cached = pings.map(lambda ping: extract(ping)).filter(lambda p: p[1] > 0).cache()


As the Python API matches closely the one used from Scala, I suggest to have a look at my older Spark tutorial if you are not familiar with Spark. Another good resource are the hands-on exercises from AMP Camp 4.

Now, let’s collect the results back and stuff it into a pandas DataFrame. This is a very common pattern, once you reduce your dataset to a manageable size with Spark you collect it back on your driver (aka the master machine) and finalize your analysis with statistical tests, plots and whatnot.

grouped = cached.groupByKey().collectAsMap()

frame = pd.DataFrame({x: log(pd.Series(list(y))) for x, y in grouped.items()})
frame.boxplot()
plt.ylabel("log(firstPaint)")
plt.show()



Finally, you can save the notebook, upload it to Github or Bugzilla and visualize it on nbviewer, it’s that simple. Here is the nbviewer powered Hello World notebook. I warmly suggest that you open a bug report on Bugzilla for your custom Telemetry analysis and ask me or Vladan Djeric to review it. Mozilla has been doing code reviews for years and with good reasons, why should data analyses be different?

Congrats, you just completed your first Spark analysis with IPython! If you need any help with your custom job feel free to drop me a line in #telemetry.

# A/B test for Telemetry histograms

A/B tests are a simple way to determine the effect caused by a change in a software product against a baseline, i.e. version A against version B. An A/B test is essentially an experiment that indiscriminately assigns a control or experiment condition to each user. It’s an extremely effective method to ascertain causality which is hard, at best, to infer with statistical methods alone. Telemetry comes with its own A/B test implementation, Telemetry Experiments.

Depending on the type of data collected and the question asked, different statistical techniques are used to verify if there is a difference between the experiment and control version:

1. Does the rate of success of X differ between the two versions?
2. Does the average value of  Y differ between the two versions?
3. Does the average time to event Z differ between the two versions?

Those are just the most commonly used methods.

The frequentist statistical hypothesis testing framework is based on a conceptually simple idea: assuming that we live in a world where a certain baseline hypothesis (null hypothesis) is valid, what’s the probability of obtaining the results we observed? If the probability is very low, i.e. under a certain threshold, we gain confidence that the effect we are seeing is genuine.

To give you a concrete example, say I have reason to believe that the average battery duration of my new phone is 5 hours but the manufacturer claims it’s 5.5 hours. If we assume the average battery has indeed a duration of 5.5 hours (null hypothesis), what’s the probability of measuring an average duration that is 30 minutes lower? If the probability is small enough, say under 5%, we “reject” the null hypothesis. Note that there are many things that can go wrong with this framework and one has to be careful in interpreting the results.

Telemetry histograms are a different beast though. Each user submits its own histogram for a certain metric, the histograms are then aggregated across all users for version A and version B. How do you determine if there is a real difference or if what you are looking at is just due to noise? A chi-squared test would seem the most natural choice but on second thought its assumptions are not met as entries in the aggregated histograms are not independent from each other. Luckily we can avoid to sit down and come up with a new mathematically sound statistical test. Meet the permutation test.

Say you have a sample of metric $M$ for users of version A and a sample of metric $M$ for users of version B. You measure a difference of $d$ between the means of the samples. Now you assume there is no difference between A and B and randomly shuffle entries between the two samples and compute again the difference of the means. You do this again, and again, and again… What you end up with is a distribution $D$ of the differences of the means for the all the reshuffled samples. Now, you compute the probability of getting the original difference $d$, or a more extreme value, by chance and welcome our newborn hypothesis test!

Going back to our original problem of comparing aggregated histograms for the experiment and control group, instead of having means we have aggregated histograms and instead of computing the difference we are considering the distance; everything else remains the same as in the previous example:


def mc_permutation_test(xs, ys, num):
n, k = len(xs), 0
h1 = xs.sum()
h2 = ys.sum()

diff = histogram_distance(h1, h2)
zs = pd.concat([xs, ys])
zs.index = np.arange(0, len(zs))

for j in range(num):
zs = zs.reindex(np.random.permutation(zs.index))
h1 = zs[:n].sum()
h2 = zs[n:].sum()
k += diff < histogram_distance(h1, h2)

return k / num



Most statistical tests were created in a time where there were no [fast] computers around, but nowadays churning a Monte-Carlo permutation test is not a big deal and one can easily run such a test in a reasonable time.

# ClientID in Telemetry submissions

A new functionality landed recently that allows to group Telemetry sessions by profile ID. Being able to group sessions by profile turns out be extremely useful for several reasons. For instance, as some users tend to generate an enourmous amount of sessions daily, analyses tend to be skewed towards those users.

Take uptime duration; if we just consider the uptime distribution of all sessions collected in a certain timeframe on Nightly we would get a distribution with a median duration of about 15 minutes. But that number isn’t really representative of the median uptime for our users. If we group the submissions by Client ID and compute the median uptime duration for each group, we can build a new distribution that is more representative of the general population:

And we can repeat the exercise for the startup duration, which is expressed in ms:

Our dashboards are still based on the session distributions but it’s likely that we will provide both session and user based distributions in our next-gen telemetry dash.

edit:

• Telemetry does not collect privacy-sensitive data.
• You do not have to trust us, you can verify what data Telemetry is collecting in the about:telemetry page in your browser and in aggregate form in our Telemetry dashboards.
• Telemetry is an opt-in feature on release channels and a feature you can easily disable on other channels.
• The new Telemetry Client ID does not track users, it tracks Telemetry performance & feature-usage metrics across sessions.

# A Telemetry API for Spark

Check out my previous post about Spark and Telemetry data if you want to find out what all the fuzz is about Spark. This post is a step by step guide on how to run a Spark job on AWS and use our simple Scala API to load Telemetry data.

The first step is to start a machine on AWS using Mark Reid’s nifty dashboard, in his own words:

1. Visit the analysis provisioning dashboard at telemetry-dash.mozilla.org and sign in using Persona (with an @mozilla.com email address as mentioned above).
2. Click “Launch an ad-hoc analysis worker”.
3. Enter some details. The “Server Name” field should be a short descriptive name, something like “mreid chromehangs analysis” is good. Upload your SSH public key (this allows you to log in to the server once it’s started up).
4. Click “Submit”. 
5. A Ubuntu machine will be started up on Amazon’s EC2 infrastructure. Once it’s ready, you can SSH in and run your analysis job. Reload the webpage after a couple of minutes to get the full SSH command you’ll use to log in.

Now connect to the machine and clone my starter project template:


git clone https://github.com/vitillo/mozilla-telemetry-spark.git
cd mozilla-telemetry-spark && source aws/setup.sh



The setup script will install Oracle’s JDK among some other bits. Now we are finally ready to give Spark a spin by launching:

sbt run


The command will run a simple job that computes the Operating System distribution for a small number of pings. It will take a while to complete as sbt, an interactive build tool for Scala, is downloading all required dependencies for Spark on the first run.


import org.apache.spark.SparkContext
import org.apache.spark.SparkContext._
import org.apache.spark.SparkConf

import org.json4s._
import org.json4s.jackson.JsonMethods._

import Mozilla.Telemetry._

object Analysis{
def main(args: Array[String]) {
val conf = new SparkConf().setAppName("mozilla-telemetry").setMaster("local[*]")
implicit val sc = new SparkContext(conf)
implicit lazy val formats = DefaultFormats

val pings = Pings("Firefox", "nightly", "36.0a1", "*", ("20141109", "20141110")).RDD(0.1)

var osdistribution = pings.map(line => {
((parse(line.substring(37)) \ "info" \ "OS").extract[String], 1)
}).reduceByKey(_+_).collect

println("OS distribution:")
osdistribution.map(println)

sc.stop()
}
}



If it the job completed successfully, you should see something like this in the output:

OS distribution:
(WINNT,4421)
(Darwin,38)
(Linux,271)


To start writing your job, simply customize src/main/scala/main.scala to your needs. The Telemetry Scala API allows you to define a RDD for a Telemetry filter:

val pings = Pings("Firefox", "nightly", "36.0a1", "*", ("20141109", "20141110")).RDD(0.1)


The above statement retrieves a sample of 10% of all Telemetry submissions of Firefox received on the 9th and 10th of November for any build-id of nightly 36. The last 2 parameters of Pings, i.e. build-id and date, accept either a single value or a tuple specifying a range.

If you are interested to learn more, there is going to be a Spark & Telemetry tutorial at MozLandia next week! I will briefly go over the data layout of Telemetry and how Spark works under the hood and finally jump in a hands-on interactive analysis session with real data. No prerequisites are required in terms of Telemetry, Spark or distributed computing.

Time: Friday, December 5, 2014, 1 PM
Location: Belmont Room, Marriott Waterfront 2nd, Mozlandia

Update
I have uploaded the slides and the tutorial.

# Clustering Firefox hangs

Jim Chen recently implemented a system to collect stacktraces of threads running for more than 500ms. A summary of the aggregated data is displayed in a nice dashboard in which the top N aggregated stacks are shown according to different filters.

I have looked at a different way to group the frames that would help us identify the culprits of main-thread hangs, aka jank. The problem with aggregating stackframes and looking at the top N is that there is a very long tail of stacks that are not considered. It might very well be that by ignoring the tail we are missing out some important patterns.

So I tried different clustering techniques until I settled with the very simple solution of aggregating the traces by their last frame. Why the last frame? When I used k-means to cluster the traces I noticed that, for many of the more interesting clusters the algorithm found, most stacks had the last frame in common, e.g.:

• Startup::XRE_Main, (chrome script), Timer::Fire, nsRefreshDriver::Tick, PresShell::Flush, PresShell::DoReflow
• Startup::XRE_Main, Timer::Fire, nsRefreshDriver::Tick, PresShell::Flush, PresShell::DoReflow
• Startup::XRE_Main, EventDispatcher::Dispatch, (content script), PresShell::Flush, PresShell::DoReflow

Aggregating by the last frame yields clusters that are big enough to be considered interesting in terms of number of stacktraces and are likely to explain the most common issues our users experience.

Currently on Aurora, the top 10 meaningful offending main-thread frames are in order of importance:

1. PresShell::DoReflow accounts for 5% of all stacks
2. nsCycleCollector::collectSlice accounts for 4.5% of all stacks
3. nsJSContext::GarbageCollectNow accounts for 3% of all stacks
4. IPDL::PPluginInstance::SendPBrowserStreamConstructor accounts for 3% of all stacks
5. (chrome script) accounts for 3% all stacks
6. filterStorage.js (Adblock Plus?) accounts for 2.7% of all stacks
7. nsStyleSet::FileRules accounts for 2.7% of all stacks
8. IPDL::PPluginInstance::SendNPP_Destroy accounts for 2% of all stacks
9. IPDL::PPluginScriptableObject::SendHasProperty accounts for 2% of all stacks
10. IPDL::PPluginScriptableObject::SendInvoke accounts for 1.7% of all stacks

Even without showing sample stacks for each cluster, there is some useful information here. The elephants in the room are clearly plugins; or should I say Flash? But just how much do “plugins” hurt our responsiveness? In total, plugin related traces account for about 15% of all hangs. It also seems that the median duration of a plugin hang is not different from a non-plugin one, i.e. between 1 and 2 seconds.

But just how often does a hang occur during a session? Let’s have a look:

The median number of hangs for a session amounts to 5; the mean is not that interesting as there are big outliers that skew the data. Also noteworthy is that the median duration of a session is about 16 minutes.

As one would expect, the number of hangs tend to increase as the duration of a session does:

Tha analysis was run on a week’s worth of data for Aurora (over 50M stackframes) and similar results where obtained on previous weeks.

There is some work in progress to improve the status quo. Aaron Klotz’s formidable async plugin initialization is going to eliminate trace 4 and he might tackle frame 8 in the future. Furthermore, a recent improvent in cycle collection is hopefully going to reduce the impact of frame 2.

# Recommending Firefox add-ons with Spark

We are currently evaluating possible replacements for our Telemetry map-reduce infrastructure. As our current data munging machinery isn’t distributed, analyzing days worth of data can be quite a pain. Also, many algorithms can’t easily be expressed with a simple map/reduce interface.

So I decided to give Spark another try. “Another” because I have played with it in the past but I didn’t feel it was mature enough to be run in production. And I wasn’t the only one to think that apparently. I feel like things have changed though with the latest 1.1 release and I want to share my joy with you.

What is Spark?

In a nutshell, “Spark is a fast and general-purpose cluster computing system. It provides high-level APIs in Java, Scala and Python, and an optimized engine that supports general execution graphs. It also supports a rich set of higher-level tools including Spark SQL for SQL and structured data processing, MLlib for machine learning, GraphX for graph processing, and Spark Streaming.”

Spark primary abstraction is the Resilient Distributed Dataset (RDD), which one can imagine as a distributed pandas or R data frame. The RDD API comes with all kinds of distributed operations, among which also our dear map and reduce. Many RDD operations accept user-defined Scala or Python functions as input which allow average Joe to write distributed applications like a pro.

A RDD can also be converted to a local Scala/Python data structure, assuming the dataset is small enough to fit in memory. The idea is that once you chopped the data you are not interested in, what you are left with fits comfortably on a single machine. Oh and did I mention that you can issue Spark queries directly from a Scala REPL? That’s great for performing exploratory data analyses.

The greatest strength of Spark though is the ability to cache RDDs in memory. This allows you to run iterative algorithms up to 100x faster than using the typical Hadoop based map-reduce framework! It has to be remarked though that this feature is purely optional. Spark works flawlessly without caching, albeit slower. In fact in a recent benchmark Spark was able to sort 1PB of data 3X faster using 10X fewer machines than Hadoop, without using the in-memory cache.

Setup

A Spark cluster can run in standalone mode or on top of YARN or Mesos. To the very least for a cluster you will need some sort of distributed filesystem, e.g. HDFS or NFS. But the easiest way to play with it though is to run Spark locally, i.e. on OSX:


brew install spark
spark-shell --master "local[*]"



The above commands start a Scala shell with a local Spark context. If you are more inclined to run a real cluster, the easiest way to get you going is to launch an EMR cluster on AWS:


aws emr create-cluster --name SparkCluster --ami-version 3.3 --instance-type m3.xlarge \
--instance-count 5 --ec2-attributes KeyName=vitillo --applications Name=Hive \
--bootstrap-actions Path=s3://support.elasticmapreduce/spark/install-spark



Then, once connected to the master node, launch Spark on YARN:


yarn-client /home/hadoop/spark/bin/spark-shell --num-executors 4 --executor-cores 8 \
--executor-memory 8g —driver-memory 8g



The parameters of the executors (aka worker nodes) should obviously be tailored to the kind of instances you launched. It’s imperative to spend some time understanding and tuning the configuration options as Spark doesn’t automagically do it for you.

Now what?

Time for some real code. Since Spark makes it so easy to write distributed analyses, the bar for a Hello World application should be consequently be much higher. Let’s write then a simple, albeit functional, Recommender Engine for Firefox add-ons.

In order to do that, let’s first go over quickly the math involved. It turns out that given a matrix of the rankings of each user for each add-on, the problem of finding a good recommendation can be reduced to matrix factorization problem:

The model maps both users and add-ons to a joint latent factor space of dimensionality $F$. Both users and add-ons are thus seen as vectors in that space. The factors express latent characteristics of add-ons, e.g. if a an add-on is related to security or to UI customization. The ratings are then modeled as inner products in that space, which is proportional to the angle of the two vectors. The closer the characteristics of an add-on align to the preferences of the user in the latent factor space, the higher the rating.

But wait, Firefox users don’t really rate add-ons. In fact the only information we have in Telemetry is binary: either a user has a certain add-on installed or he hasn’t. Let’s assume that if someone has a certain add-on installed, he probably likes that add-on. That’s not true in all cases and a more significant metric like “usage time” or similar should be used.

I am not going to delve into the details, but having binary ratings changes the underlying model slightly from the conceptual one we have just seen. The interested reader should read this paper. Mllib, a machine learning library for Spark, comes out of the box with a distributed implementation of ALS which implements the factorization.

Implementation

Now that we have an idea of the theory, let’s have a look at how the implementation looks like in practice. Let’s start by initializing Spark:


val sc = new SparkContext(conf)



As the ALS algorithm requires tuples of (user, addon, rating), let’s munge the data into place:


val ratings = sc.textFile("s3://mreid-test-src/split/").map(raw => {
val parsedPing = parse(raw.substring(37))
(parsedPing \ "clientID", parsedPing \ "addonDetails" \ "XPI")
}).filter{
// Remove sessions with missing id or add-on list
case (JNothing, _) => false
case (_, JNothing) => false
case (_, JObject(List())) => false
case _ => true
}.map{ case (id, xpi) => {
}}.filter{ case (id, addonList) => {
// Remove sessions with empty add-on lists
}}.flatMap{ case (id, addonList) => {
// Create add-on ratings for each user
}}



Here we extract the add-on related data from our json Telemetry pings and filter out missing or invalid data. The ratings variable is a RDD and as you can see we used the distributed map, filter and flatMap operations on it. In fact it’s hard to tell apart vanilla Scala code from the distributed one.

As the current ALS implementation doesn’t accept strings for the user and add-on representations, we will have to convert them to numeric ones. A quick and dirty way of doing that is to hash the strings:


// Positive hash function
def hash(x: String) = x.hashCode & 0x7FFFFF

val hashedRatings = ratings.map{ case(u, a, r) => (hash(u), hash(a), r) }.cache



We are nearly there. To avoid overfitting, ALS uses regularization, the strength of which is determined by a parameter $\lambda$. As we don’t know beforehand the optimal value of the parameter, we can try to find it by minimizing the mean squared error over a pre-defined grid of $\lambda$ values using k-fold cross-validation.


// Use cross validation to find the optimal number of latent factors
val folds = MLUtils.kFold(hashedRatings, 10, 42)
val lambdas = List(0.1, 0.2, 0.3, 0.4, 0.5)
val iterations = 10
val factors = 100 // use as many factors as computationally possible

val factorErrors = lambdas.flatMap(lambda => {
folds.map{ case(train, test) =>
val model = ALS.trainImplicit(train.map{ case(u, a, r) => Rating(u, a, r) }, factors, iterations, lambda, 1.0)
val usersAddons = test.map{ case (u, a, r) => (u, a) }
val predictions = model.predict(usersAddons).map{ case Rating(u, a, r) => ((u, a), r) }
val ratesAndPreds = test.map{ case (u, a, r) => ((u, a), r) }.join(predictions)
val rmse = sqrt(ratesAndPreds.map { case ((u, a), (r1, r2)) =>
val err = (r1 - r2)
err * err
}.mean)

(model, lambda, rmse)
}
}).groupBy(_._2)
.map{ case(k, v) => (k, v.map(_._3).reduce(_ + _) / v.length) }



Finally, it’s just a matter of training ALS on the whole dataset with the optimal $\lambda$ value and we are good to go to use the recommender:


// Train model with optimal number of factors on all available data
val model = ALS.trainImplicit(hashedRatings.map{case(u, a, r) => Rating(u, a, r)}, factors, iterations, optimalLambda._1, 1.0)

def recommend(userID: Int) = {
val top = predictions.top(10)(Ordering.by[Rating,Double](_.rating))
}

recommend(hash("UUID..."))



I omitted some details but you can find the complete source on my github repository.

To submit the packaged job to YARN run:


spark-submit --class AddonRecommender --master yarn-client --num-executors 4 \



So what?

Question is, how well does it perform? The mean squared error isn’t really telling us much so let’s take some fictional user session and see what the recommender spits out.

For user A that has only the add-on Ghostery installed, the top recommendations are, in order:

• NoScript
• Web of Trust
• Symantec Vulnerability Protection
• Better Privacy
• LastPass
• DuckDuckGo Plus
• HTTPS-Everywhere
• Lightbeam

One could argue that 1 out of 10 recommendations isn’t appropriate for a security aficionado. Now it’s the turn of user B who has only the Firebug add-on installed:

• Web Developer
• FiddlerHook
• Greasemonkey
• ColorZilla
• User Agent Switcher
• McAfee
• RealPlayer Browser Record Plugin
• FirePHP
• Session Manager

There are just a couple of add-ons that don’t look that great but the rest could fit the profile of a developer. Now, considering that the recommender was trained only on a couple of days of data for Nightly, I feel like the result could easily be improved with more data and tuning, like filtering out known Antivirus, malware and bloatware.