Category Archives: Internships

Root Cause Analysis for Hadoop Applications

Parth Shah and Thai Bui


One of the reasons why Hadoop jobs are hard to operate is their inability to provide clear, actionable error diagnostic messages for users. This stems from the fact that Hadoop consists of many interrelated components. When a component fails or behaves poorly, the failure will be cascaded to its dependent components which causes a job to fail.

This blog post is an attempt to help to solve that problem by created a user-friendly, self-serving and actionable Hadoop diagnostic system.

Our Goals

Due to its complex nature, the project was split into multiple components. First, we prototyped a diagnostic tool to help debug Hadoop application failure by providing a clear root cause analysis and save engineering time. Second we purposely inflict failures on a cluster (via a method called chaos testing) and collected the data to understand how certain log messages map to errors. Finally, we examined regular expression as well as natural language processing (NLP) techniques to automatically provide root cause analysis in production.

To document this, we organized the blog in the following sections:

  • Error Message Analytics Portal
    • To provide a quick glance at the known root cause.
  • Datadog Dashboard
    • To calculate failure rates related to the unknown root cause and known root cause.
    • Separate infrastructure failure from a missing data failure (missing partition).
  • Data Access
    • All relevant logs messages from services like Yarn, Oozie, HDFS, hive server, etc. were collected and stored under an S3 bucket with an expiration policy.
  • Chaos Data Generation
    • Using chaos testing, we produced actual errors related to memory, network, etc. This was done to understand relationships between log messages and root cause errors.
    • A service was made to create an efficient and simple way to run chaos tests and collect its corresponding/related log data.
  • Diagnostic Message Classification
    • Due to the simple and repetitive nature of log messages (low in entropy), we built a natural language processing model that classified specific error types of an unknown failure.
    • Tested the model on the chaos data for the specific workload at Bazaarvoice

Error Message Analytics Portal

Bazaarvoice provides an internal portal tool for managing analytics applications by the end-users. The following screenshot demonstrates an example message of a “Job failures due to missing data”. This message is caught by using a simple regular expression of the stack trace. A regular expression works because there is only one way a job could fail due to missing data.

DataDog Dashboard

What is a partition?

Partitioning is a strategy of dividing a table into related parts based on date, components or other categories. Using a partition, it is easy and faster to query a portion of data. However, the partition a job is querying can sometimes not be available which causes a job to fail by our design. The dashboard below calculates metrics and keeps track of jobs that failed due to an unavailable partition.

The dashboard classifies failed jobs as either a partition failure (late/missing data) or an unknown failure (Hadoop application failure). Our diagnostic tool attempts to find the root cause of the unknown failures since a late or missing data is an easy problem to solve. 

Data Access

Since our clusters are powered by Apache Ambari, we leveraged and enhanced Ambari Logsearch Logfeeder to ship logs of relevant services directly to S3 and partitioned the data by as shown in the raw_log of the Directory Tree Diagram below. However, as the dataset got bigger, partitioning was not enough to efficiently query the data. To speed up read performance and to iterate faster, the data was later converted into ORC format.

Convert JSON logs to ORC logs

DROP TABLE IF EXISTS default.temp_table_orc;
CREATE EXTERNAL TABLE default.temp_table_orc (
cluster STRING,
file STRING,
thread_name STRING,
level STRING,
event_count INT,
type STRING,
hour INT,
component STRING
LOCATION 's3a://<bucket_name>/orc-log/${workflowYear}/${workflowMonth}/${workflowDay}/${workflowHour}/';
INSERT INTO default.temp_table_orc SELECT * FROM bazaar_magpie_rook.rook_log
WHERE year=${workflowYear} AND month=${workflowMonth} AND day=${workflowDay} AND hour=${workflowHour};

Sample Queries for Root Cause Diagnosis

SELECT t1.log_message,
  (SELECT log_message,
     ( SELECT log_message,
              row_number() over (partition BY log_message
                                 ORDER BY logtime) AS r
      FROM bazaar_magpie_rook.rook_log
      WHERE cluster_name = 'cluster_name'
        AND DAY = 28
        AND MONTH=06
        AND LEVEL = 'ERROR' ) S
   WHERE S.r = 1 ) t1
  (SELECT log_message,
          COUNT(log_message) AS Frequency
   FROM bazaar_magpie_rook.rook_log
   WHERE cluster_name = 'cluster_name'
     AND DAY = 28
     AND MONTH=06
   GROUP BY log_message) t2 ON t1.log_message = t2.log_message
WHERE t1.logtime BETWEEN '2019-06-28 13:05:00' AND '2019-06-28 13:30:00'
ORDER BY t1.logtime
LIMIT 400;

SELECT log_message, logtime, type
FROM bazaar_magpie_rook.rook_log WHERE level = 'ERROR' AND type != 'logsearch_feeder' AND logtime BETWEEN '2019-06-24 09:05:00' AND '2019-06-24
ORDER BY logtime
LIMIT 1000;

SELECT log_message, COUNT(log_message) AS Frequency
FROM  bazaar_magpie_rook.rook_log
WHERE cluster_name = 'dev-blue-3' AND level = 'ERROR' AND logtime BETWEEN '2019-06-27 13:50:00' AND '2019-06-29 14:30:00'
GROUP BY log_message
ORDER BY COUNT(log_message) DESC

Chaos Data Generation

The chaos data generation process makes up a bulk of this project as well is the most important. It stems from the process of chaos testing to experiment on software and infrastructure in order to understand the systems ability to withstand unexpected or turbulent conditions. This concept of testing was first introduced by Netflix in 2011. The following pseudocode explains how it works.

Submit a normal job through an API on a specific cluster

Create a list of IP addresses for all the testable nodes in that cluster (such as the workers nodes).

Once we have all associated nodes, inject failures into them (stress test memory or network)

While the job is not done

    Let the job run

At this point, the job has either failed or succeeded

Stop all stress tests

Gather the job details such as start time, end time, status, and most importantly its associated stress test type (memory, packet_corruption, etc.). 

Store job details and its report in JSON format

Job Report Example

"duration":"45 Minutes",
"nominalTime":"2018-10-27 07:00:00",
"errorMessage":"Error: Main class [org.apache.oozie.action.hadoop.Hive2Main], exit code [2]",
"startTime":"2019-07-25 18:45:37",
"stopTime":"2019-07-25 19:31:18","failedAction":"hive-action",

We followed the same process to generate chaos data on different types of injected failures such as:

  • High Memory Utilization on the Host
  • Packet Corruption
  • Packet Loss
  • High Memory Util on a Container

While there are certainly more errors the model is capable of learning and classifying, for the purpose of prototyping we kept the type of failures to 2-3 categories.

Diagnostic Message Classification

In this section, we explore 2 types of error classification methods, a simple regex and supervised learning.

Short Term Solution with Regex

There are many ways to analyze text for different patterns. One of these ways is called Regular Expression Matching (regex). Regex is “special strings representing a pattern to be matched in search operation”. One use of regex is finding a keyword like “partition”. When a job fails due to missing partitions its error message usually looks something like this “Not all partitions are available even after 16 attempts“. The regex for this specific case would look like \W*((?i)partition(?-i))\W* .

A regex log parser could easily be able to identify this error and do what is necessary to fix the problem. However, the power of regex is very limiting when it comes to analyzing complex log messages and stack traces. When we find a new error, we would have to manually hardcode a new regular expression to pattern match the new error type which is very error-prone and tedious in the long run.

Since the Hadoop eco-system includes many components, there are many combinations of log messages that can be produced; a simple regex parser would be difficult to work here because it can not process general similarities between different sentences. This is where natural language processing helps.

Long Term Solution with Supervised Learning

The idea behind this solution is to use natural language processing to process log messages and supervised learning to learn and classify errors. This model should work well with log data vs. normal language due since machine logs are more structured and low in entropy. You can think of entropy as a measure of how unstructured or random a set of data is. The English language tends to have high entropy relative to machine logs due to its sometimes illogical sentence structure. On the other hand, machine-generated log data is very repetitive which makes it easier to model and classify.

Our model required pre-processing of logs which are called tokenization. Tokenization is the process of taking a set of text and breaking it up into its individual words or tokens. Next, the set of tokens were used to model their relationship in high dimensional space using Word2Vec. Word2Vec is a widely popular model used for learning vector representation of words called, “word embeddings” (word2vec). Finally, we use the vector representation to train a simple logistic regression classifier using the chaos data generated earlier. The diagram below shows a similar training processing from Experience Report: Log Mining using Natural Language Processing and Application to Anomaly Detection by Christophe Bertero, Matthieu Roy, Carla Sauvanaud and Gilles Tredan.

In contrast to the diagram above, since we only trained the classifier on error logs, it was not possible to classify a normal system as it should produce no error logs. Instead, we trained the data on different types of stressed systems. Using the dataset generated by chaos testing, we were able to identify the root cause for each of the error message for each failed job. This allowed us to make our training dataset as shown below. We split our dataset into a simple 70% training and 30% for testing.

Subsequently, each log message was tokenized to create a word vector. All the tokens were inputted into a pre-trained word2vec model which mapped out each word in a vector space, creating a word embedding. Each line was represented as a list of vectors and an average of all the vectors in a log message represents the feature vector in high dimensional space. We then inputted each feature vector and its labels into a classification algorithm such as logistic regression or random forest to create a model that could predict the root cause error from a logline. However, since failure often contains multiple error log messages, it would not be logical to simply reach a root cause conclusion from just a single line of the log from a long error log. A simple way to tackle this problem would be to window the log and input the individual lines from the window into the model and output the final error as the most common error outputted by all the lines in a window. This is a very naive way of breaking apart long error log so more research would have to be done to process error logs without losing the valuable insight of their dependent messages.

Below are a few examples of error log messages and their tokenized version.

Example Raw Log

java.lang.RuntimeException: org.apache.thrift.transport.TSaslTransportException: No data or no sasl data in the stream at org.apache.thrift.transport.TSaslServerTransport$Factory.getTransport(
    at org.apache.thrift.server.TThreadPoolServer$
    at java.util.concurrent.ThreadPoolExecutor.runWorker(
    at java.util.concurrent.ThreadPoolExecutor$
    at [?:1.8.0_191] Caused by: org.apache.thrift.transport.TSaslTransportException: No data or no sasl data in the stream at
    at ~[hive-exec-] at org.apache.thrift.transport.TSaslServerTransport$Factory.getTransport(

Tokenized Raw Log



The model accurately predicted the root cause of failed job with 99.3% accuracy in our test dataset. At an initial glance, the metrics calculated on the test data look very promising. However, we still need to evaluate its efficiency in production to get a more accurate picture. The initial success in this proof of concept warrants further experimentation, testing, and investigation for using NLP to classify errors. 

Training Data 70% (Top 20 Rows)

Test Data Results 30% (Top 20 Rows)


In order to implement this tool in production, data engineers will have to automate certain aspects of the data aggregation and model building pipeline

Self Reporting Errors

A machine learning model is only as good as its data. For this tool to be robust and accurate, any time engineers encounter a new error or an error that the model does not know about, they should report their belief of the root cause so the corresponding errors logs get tagged with the specific root cause error. For instance, when a job fails for a similar reason, the model will be able to diagnose and classify its error type. This can be done through a simple API, form, Hive query that takes in the job id and its assumed root_cause. The idea behind is this that by creating a form, we could manual label log messages and errors in case the classifier fails to correctly diagnose the true problem. 

The self-reported error should take the form of the chaos error to ensure the existing pipeline will work.

Automating Chaos Data Generation

The chaos tests should run on a regular basis in order to keep the model up to date. This could be done by creating a Jenkins job that automatically runs on a regular cadence. Often times, running a stress test will make certain nodes to be unhealthy. It causes our automation to be unable to SSH into the nodes again to stop the stress test such as when the stress test is interfering with connectivity to the node (see the error below). This can be solved by creating a time limit for the stress test, so the script does not have to ssh again once the job has finished. In the long run, as the self-reported errors grow, the model should rely on chaos test generated test less. Chaos tests produce logs for very extreme situations which could be not typical for a normal production environment. 

java.lang.RuntimeException: Unable to setup local port forwarding to
    at com.bazaarvoice.rook.infra.util.remote.Gateway.connect(
    at com.bazaarvoice.rook.infra.regress.yarn.RootCauseTest.kill_memory_stress_test(

Cross-Platform Mobile SDK Testing

This Bazaarvoice blog entry is co-authored by Tanvir Pathan as part of a Bazaarvoice internship project on the Bazaarvoice Mobile Team.

Automated testing of native mobile applications has long been a pain point in the world of mobile app development. If you are creating and distributing apps or open source SDKs across two or more major platforms (Android and iOS in our case), you can easily find yourself duplicating efforts to test the same source and business logic across different technology stacks. For example, if you have experienced developers and testers using Xcode for iOS apps, they may tend to automate testing with Instruments Automation, where Android developers and testers may automate with Espresso or UIAutomator. This becomes an expensive proposition for development and maintenance of unit tests, which can be costly as your test coverage increases.

Test strategy can also vary depending on the type of mobile app development your shop pursues: native, hybrid, cross-compile, mobile web. Hence, the selection of test tools will vary depending on how you build and deploy apps.

In this blog post, we’ll detail a novel solution to cross-platform testing of our native SDKs, along with some background of other mobile tool offerings. Our solution focuses on cross-platform open source mobile SDK testing utilizing Cordova to wrap our SDKs in a generic JavaScript interface, and Calabash to drive our cross-platform behavioral tests.

If you want to check out the full solution, the Cordova plugin and description on how to execute Calabash can be found in our Github repository.

How to View Mobile Application Development

Non-mobile app developers typically don’t actually know the difference between a web app, native, or hybrid app. If you work in any business that supports some kind of mobile solution (and you probably wouldn’t be reading this if you didn’t) it’s really important to understand some fundamental differences. It’s very easy to just throw out the word “mobile” in conversation and not realize there’s multiple parts to this elephant!


The table below presents four general categories of mobile application development. Keep these categories in mind when talking about “mobile” in general and don’t fall in the trap of the blind men and the elephant.

Mobile Development Types & Tools

Native Application Development Developers creating purely native apps will write in the language supported by their target platform. For iOS apps, developers can write in Swift and/or Objective-C, while Android developers can write in Java (and C/C++ for lower level execution)
Cross-compilation Developers can also write apps for multiple platforms in a single language, such as Java Script. Cross-compilation tools will take a common language and actually convert it to the target language of the native platform. In this case, while developers aren’t writing in the native language, the tools create real native apps. Some of the most common tools are: Appcelerator (JavaScript), Xamarin (C#), React Native (Java Script)
Hybrid Applications Hybrid apps utilize Web Views to display content, typically written with common web development technologies such as Java Script, HTML, and CSS. Hybrid apps will typically have a “bridge” that allows javascript code to communicate with the native libraries to do things like access the camera, location services, or contacts. Cordova (aka PhoneGap) as the application container. Developers choose their favorite UI layer to work with Corodva: ionic, Sencha Touch, jQuery Mobile, Onsen, Framework 7.
Web Applications / Mobile Web A web application isn’t as much an application as it is a mobile optimized web site. Hence, you won’t find a Web Application in the App or Google Play Store, you just fire up your favorite mobile web browser and load the site.

Cross Platform Mobile UI Testing Tools

When developing for native mobile, developers will typically write unit tests to check individual pieces of functionality and business logic, perhaps even employ certain mocking techniques to test networking and user interface capabilities without the need for a full application. However, when it comes to full system testing of full applications and SDKs, making the right selection can be a tough process. However, if cross-platform testing is your objective and you want to write all your tests in one common scripting language, the options narrow quickly.

While there are platform-specific UI automation frameworks for Android (Robotium, UiAutomation) and iOS (Instruments Automation, Keep It Functional, EarlGrey), there currently only two (that we are aware of) that allow us to test cross-platform with a common script.

Appium Appium lets developers write tests for applications without having to add any additional code the applications. It works with native, hybrid, and mobile web applications.
Calabash Calabash is owned and maintained by Xamarin, and provides cross-platform testing for native or hybrid apps. Unit tests can be written in Ruby with Cucumber.

2 Birds, One Stone


Making the decision to use Cordova and Calabash was fairly easy. First we already distribute our BBVSD via frameworks and libraries for iOS and Android. Second we know some of our customers are creating hybrid applications with Corodva. So we immediately thought: what if we could create a test environment that not only tests our SDK deliverables, but also provides our clients with an easy avenue to integrate Bazaarvoice services into their own Cordova app. Win! Win! As well, because we already use Cucumber extensively at Bazaarvoice, we decided to leverage our already strong in-house expertise and utilize an automation framework that is internally familiar.

Calabash Unit Tests

Another great thing about using Calabash at Bazaarvoice is that we already have an internal framework developed on top of Cucumber. Because Calabash layers on top of Cucumber, the paradigm and philosophy of writing human readable test cases still applies. The test cases utilize Behavior Driven Development  modeling tools to add meaning to your mobile app testing.

Let’s say you are creating the same app for multiple platforms. Typically, you would have to write completely different sets of code to run similar tests. With Calabash, this is not the case. You write one set of code tests and make slight adjustments depending on the platforms in question and you are done!  Best of all, in addition to Calabash being free, the test cases are super easy to write as a developer and even easier to read for others who may be interested in checking out the health of the project.

Needless to say, Calabash provides a lot of benefits for cross platform testing. Lets take a look at an example test case from the BVSDK Cordova Plugin project. Let’s go through a simple scenario based on the following app screens shots from the iOS simulator.

bvsdk_build_simulator bvsdk_running_simulator

Say you wanted to count the number of products that were recommended by our Product Recommendations API. If you were doing it manually you would go through the following steps:

  1. Wait for the app to launch
  2. Make sure you receive a success alert and press ok
  3. Click the Recommendations tab
  4. Then count how many products there are and compare them to what you were expecting

Now how would we code this? Calabash has two essential components: one feature file and one ruby file. The feature file is where you write out the tests and the ruby file is used primarily to make custom functions if needed (although most of what you need comes right out of the box). So returning to our problem, writing out the test case, you simply write down those exact steps in the feature file:

Feature File
Feature: BVSDK Demo App
  Scenario: As a user, I want to get new recommendations
    Given the app has launched
    Then I should see "BVSDK has been built successfully."
    Then I press the OK button
    Then I press Recommendations tab
    Then I check number of products


That’s really all there is to it. Of course, tests can be also written to be more platform specific when needed.

Entering the Matrix – Travis CI

We use Travis CI for all our public repos at Bazaarvoice. It’s awesome. But, we have to support multiple build tools on different virtual machines. Configuring all these build machines with custom tools sounds and build scripts really scary! Freak out!


The really slick thing about Travis CI is that you can test multiple configuration, variations, permutations, salutations, etc, etc, etc, by building a matrix in Travis’ Config file (.travis.yml). For our testing, since XCode only runs on OS X and it’s the only way to build for iOS, we must have an OS X image. For the Android Studio and Gradle build tools, we build against Linux. In addition there’s some common tooling we can install for each build machine. The result is that we can use two different VMs for testing each platform, with just one set of tests. Note in the test result below, the build jobs are defined by the environment variables defined in the Travis config file.


The .travis.yml script looks like this, where we build a matrix with environment variables and platforms:

    - language: android
      env: TO_TEST=ANDROID
      jdk: oraclejdk8
    - os: osx
      osx_image: xcode8
      env: TO_TEST=IOS
  fast_finish: true
    - platform-tools
    - tools
    - android-23
    - build-tools-23.0.3
    - extra-android-m2repository
    - extra-google-m2repository
    - sys-img-armeabi-v7a-android-19
  - rvm install 2.2.0
  - if [ "$TO_TEST" = "ANDROID" ]; then gem install calabash-android; fi
  - if [ "$TO_TEST" = "IOS" ]; then gem install calabash-cucumber; fi
  - if [ "$TO_TEST" = "ANDROID" ]; then chmod 755; fi
  - if [ "$TO_TEST" = "ANDROID" ]; then ./; fi
  - if [ "$TO_TEST" = "ANDROID" ]; then chmod 755; fi
  - if [ "$TO_TEST" = "ANDROID" ]; then ./; fi
  - if [ "$TO_TEST" = "IOS" ]; then chmod 755; fi
  - if [ "$TO_TEST" = "IOS" ]; then ./; fi

BVSDK Cordova Plugin Features

So what if I want to try out the BVSDK Cordova plugin? If you want more info or checkout the source code for the plugin and unit tests, just head over to our Cordova Github repo. There’s plenty of info in the README for running the examples and unit tests.

Open Source Contributions

If you are building a Cordova-based application and want to see other things added just let us know, or better yet submit a pull request and we’ll be happy to review it!

Intern Demo Day

As the summer comes to an end, so do the internships for numerous university students here at Bazaarvoice. This past week, the interns were given an opportunity to present a summary of their accomplishments. This afternoon of presentations, known as the Bazaarvoice “Intern Demo Day”, highlighted the various achievements throughout the company, not just in the R&D department.

The following is a short summary of the great work our interns complete this summer as well as some images from the “Intern Demo Day”.

CHASE PORTER: My project, which I have named “The Great (red)Shift”, is intended to improve data accessibility for computed aggregated counts of various canonical events written to HBase. To do this I designed a data warehouse in Amazon Redshift that I loaded with transformed aggregated counts extracted from the tables in HBase. This makes the counts readily SQL query-able in an incredibly fast system whereas before they had to be computed with performance heavy queries from Raw Logs generated by Cookie Monster. The biggest block for this project was in processing the data from HBase which was stored as serialized bytes and needed to be handled uniquely for different types of canonical events (i.e. pageviews, impressions, features) to translate into a readable form for Redshift.

BEN DEVORE: My product is web crawler written in node.js that scrapes clients’ webpages for product data in order to build their product feeds for them. For many of Bazaarvoice’s smaller clients, building and maintaining their product feed is a significant obstacle in the onboarding process. This tool aims to clear that obstacle by taking this task out of there hands.

STONEY MCCRARY: So I have been fortunate enough to get to work on several different pieces in curations but I am going to talk on what I have been hammering on for the last couple of weeks. More and more of our high volume clients are receiving millions of hits a day and this has caused performance to become a higher priority problem for them. In response to this, we are focusing our efforts on building a new display with performance in mind. Performance for the display centers around only providing the minimal amount of data needed and supply the rest as necessary. The piece I will be showing is the display carousel and how it dynamically loads and dumps the data to allow for faster loading and to keep browser memory low.

ZESHAN ANWAR: Eagle is a dashboard built for our Incubator team. With so many moving parts, it was important we had a summarized ‘birds-eye’ view of the team in one place. Eagle was initially meant to be an aggregation of all our Jenkin builds; a single view of all our jobs across our different Jenkins environments. However, it grew to also include JIRA and GitHub statistics. My other project was optimizing our UI tests by having them run concurrently. Our old UI tests were extremely slow, and by running them in parallel we drastically reduced test times.



BRENDON KELLEY: Testing Framework: This summer my project was to help build out a new testing framework for Curations. The current automation tests used for Curations is Saladhands. Before my internship, there wasn’t much if any automation tests for the submission/direct upload capability of Curations. I worked on creating tests and a CI environment for submission in a new testing framework called Intern. One of tests includes a language translation test using mongoDB as an endpoint to store the various languages’ text. Intern is a javascript based testing framework which will allow developers to contribute to writing tests since Curations is mostly javascript. I’ve also worked on updating and creating new console tests in this framework. The foundation built this summer in Intern will enable the ability to further contribute to the framework.

KRYSTINA DIAO: My main project for the summer was to analyze and report the effectiveness of the implementation of the new Connections Knowledgebase. Through Salesforce, I collected and analyzed the number of cases, time spent on each case, etc. After drawing my conclusions, I decided to present my findings via data visualization methods (JavaScript’s C3 and D3 libraries) and provide actionable insights on how this information can be leveraged. This information is valuable in that it can be used for future product KB decisions, as well as understanding how much time, manpower, and money is saved by having a KB.

Krystina Diaos (3)

Krystina Diaos (2)

Krystina Diaos

MARKO SAVIC: Over the summer, I was a part of the SEO Team. I managed to create a tool on pagesManaged and keywordsManaged feed for every Spotlights client. Generated feeds will be consumed by SeoClarity tools on a daily basis. This helps in identifying search rank gains on the specified keywords and pages where Spotlights are present. The SeoClarity reporting will help in proving out Spotlights value and eventually lead to Spotlights renewal/upticks.Also, I created algorithm tweaks on the PINS (Post Interaction Notification System) Generator that take into account product activeness, product relevancy and review count, and use them to ask the user to write reviews on the most relevant products.

TREVOR NELLIGAN: Here is a description of my project: I worked on the Aperture Component library and many of the projects it supports. Aperture is build in React, and its purpose is to be used as an internal Bazaarvoice tool for constructing web pages. Using Aperture, anyone at Bazaarvoice can easily create a functional, intuitive, Bazaarvoice themed webpage, all with the building blocks Aperture provides.

Using the Aperture library, I helped the construction of numerous pages for the curations beta console. I personally built the interface for a new client-facing template builder, which will allow clients to create curations templates quickly and easily without having to go through an implementation engineer and a long process, as was the case previously. I also supplied custom Aperture components for several projects, like the content curation beta page.

RAMIE RAUFDEEN: The mixer is a component of our product recommendations engine which differentiates shoppers, and optimizes recommendations for them. This is primarily derived from their shopping behavior – in real time. Prior to the mixer, product recommendations were aggregated from multiple sources, using the same algorithm for every shopper. Shoppers are now categorized based off of a set of rules (using the shopper’s profile data), each of the rules map to a plan (which you can think of as an ‘algorithm’). A plan defines how recommendations should be mixed from each of the sources. For example, if source B has proven to have a higher conversion rate for ‘heavy-shoppers’, the plan for ‘heavy-shopper’ would give a higher weighting to source B. We can now target specific types of shoppers when it comes to product recommendations. This also sets the groundwork for a more granular machine learning implementation in the future.


We want to thank all the interns who spent time with us this summer and wish you the best back at school. We look forward to hearing about all the great things you all develop in the future.

If you are interested in an internship at Bazaarvoice, please contact

Using the Bazaarvoice Conversations API to build a UGC template

(This post is by Devin Carr, one of our Summer 2013 interns.)

LightReviewDisplaySmallWorking as a Developer Advocate intern on the Bazaarvoice Developer Relations team has been a great learning opportunity. At the beginning of the Summer I discussed with my mentors, Chas Peacock and Frederick Feibel, what I wanted to learn while interning. We decided on front-end web-development because it is an area in which I had minimal experience. For the rest of the Summer I spent my time creating a simple review display template using the Bazaarvoice Conversations API and some really cool front-end frameworks & libraries. I also researched other ratings and reviews implementations to learn about common industry best practices and incorporated them into what I am calling my LightReviewDisplay Template.

LightReviewDisplay_jsComing from a background limited to Java, I quickly discovered that Javascript is quite a different programming language. I took a few days to learn basic Javascript and HTML from Codecademy to get a grasp of the syntax and functions. Using that knowledge I created the first version of the LightReviewDisplay. I put all of the Javascript and dynamic HTML in one large file that handled loading the reviews and displaying them on the page. During my first code review Chas and Frederick suggested my code was lacking in structure and layout best practices, such as including an Object Oriented Programing model. Instead of most of my code being in just a few functions they recommended I implement RequireJS, self described as a “JavaScript file and module loader”, to encapsulate my code within a maintainable structure. They also suggested that I use Mustache, a HTML template library. Using these two libraries I could divide my code into multiple modules and avoid having large amounts of HTML embedded within my Javascript.

LightReviewDisplay_js_filterThe second version that I made was starting to look more like a modern and modular Javascript project. Chas and Frederick agreed that the increased readability of my refactored project was an improvement, but they felt it could be further improved by imposing more structure. They suggested I research some MVC (Model, View, Controller) frameworks. I looked into Backbone, KnockoutJS, Angular, and Stapes, all very versatile with separated modules for the model (data), the view (page), and the controller (user input). I ended up choosing Stapes because it is lightweight and is a very minimal framework that allows lots of customization.

LightReviewDisplayRatingsSummaryAt this point the project was functional but the user experience needed improvement. I decided to do some research into how a product review page should actually look. I inspected several retailers to see what was common about their layouts and what would be the best way to integrate reviews onto a product page. They all followed a two part system: a summary section at the top of the page and a more detailed section further down. The summary section shows the product’s price, description and a review ratings summary. It typically had the average rating for the product, the total number of reviews, and a rating breakdown per-star. Then towards the bottom of the product page, there was a review module that held all of the consumer reviews. I combined what I had learned about layout with Twitter Bootstrap resulting in an efficient and usable template for displaying UGC.

My Summer project was a great way to get started with Javascript and get familiar with the Bazaarvoice Conversations API. Throughout my internship I learned a lot about being a software engineer, expanded my CS knowledge and got some firsthand experience working amongst a small team of developers. This internship also gave me the opportunity to learn with the latest front-end libraries, architectural patterns, and work with really great developers. A special thank you goes out to Chas and Frederick for taking time out of their days to assist me with any problem I came across.

More about the LightReviewDisplay Template:
Bazaarvoice Inspiration Gallery:

Summer 2013 Intern Science Fair

large_groupEvery year at Bazaarvoice we bring on a new class of Summer interns and put them to work creating innovative (and educational) projects. At the beginning of the Summer interns choose from a list of projects and teams that interest them. From there they are embedded in a team where they spend the rest of the Summer working on their chosen project with help from seasoned professional mentors. At the end of the Summer they set up demos and we invite the whole company to visit with them and see what they have accomplished. Read on to learn about our interns and their fantastic and innovative projects.

Below are some of the 2013 intern science fair projects in the creators own words:

Devin Carr
School: Texas A&MField of study: Electrical Engineering

ccarr_intern_scifairI built a product review template using the Bazaarvoice Platform API as well as some of the latest front-end libraries and architectural patterns. It provides a best practice and simplistic model to assist potential/existing clients and their developers with a base to develop an efficient product page integrated with Bazaarvoice Ratings & Reviews.

Lewis Ren
School: UC BerkeleyField of study: Electrical Engineering & Computer Science

Genie is a product recommendation tool based on user clustering. The idea is that you want to give every user the possibility of a product within the network as a recommendation; however, be able to rank them in such a way that is most relevant to the specific user. By clustering users based on modularity we can determine how big of an impact a specific user will have on the rest of the network. For example, a high modularity node that makes a decision will have a stronger impact on its neighbors, but will propagate out slower and have a small to negligible effect on outer nodes. Similarly, a low modularity node that makes a decision will have a lesser impact, but will radiate outwards much more quickly.

Matt Leibowitz
School: University of DallasField of study: Physics / Computer Science

My project is called Flynn, which is a web page for accessing documents out our data store. Flynn allows developers to bypass a fairly long and involved process (up to ten minutes of looking up documentation and performing curl requests) with a quick and simple web interface.

Perry Shuman

My project is a bookmarklet built to provide debug information about a page with injected BV content: information that could help a client fix implementation issues that they might have, without having to know the inner workings of our code. It also allows previously hidden errors to be displayed to a client, giving them a first line of information to debug broken pages.

Ralph Pina
School: UT AustinField of study: Computer Science

rpina_intern_scifair“Over this past summer I interned with Bazaarvoice’s mobile team. In its effort to make it easier to integrate BV’s services into every platform and outlet their clients use it has build and distributed mobile SDKs (Software Development Kits) to display and submit data to the BV API. I worked to improve the Android SDK, and along with another intern, Devin Carr, wrote a new Windows Phone 8 SDK. I also wrote 2 Windows Phone 8 sample applications to demonstrate how to implement the SDK in an app. Afterwards, we opened sourced all our mobile SDKs under the permissive Apache 2.0 license.”

Rishi Bajekal
School: University of Illinois at Urbana-ChampaignField of study: Computer Science

Built a RESTful web service for serving photos from the data stack and worked on the Conversations 2013 API as part of the that team.

Ryan Morris
School: McCombs School of Business at the University of Texas at AustinField of study: Management Information Systems

rmorris_intern_scifairMy project’s theme was predictive modeling and focused around using a client’s web metrics, review count, and average rating to predict orders for products. To create the predictive model, I pulled the client’s visits, orders, and revenue for a certain time period from their web analytics account. I then pulled the client’s number of approved reviews and average rating for each product from the Bazaarvoice database using mySQL. Finally, I ran the data in the statistical program R to come up with a predictive model. With this model, I was able to analyze the effects that varying review count and average rating had on predicted orders for a product.

Schukey Shan
School: University of WaterlooField of study: Software Engineering

sshan_intern_scifairMy project was the brand search summary, which provides both an API and a UI to summarize data we collect on brands. For example, if we search for a brand “Acme”, the API looks for all the brand names that contain the word “Acme”, creates a composite of the matched brand values, and returns the count for pageviews, ugc mpressions, and unique visitors for each client of each brand value in the composite. This would be useful as an overview of how brand values look like over the network.

Steven Cropp
School: Rensselaer Polytechnic InstituteField of study: Computer Systems Engineering and Computer Science

This Summer I worked on an xml file validation service. Users can set up “rules” that dictate how xml files must be formed, and then apply those rules to our clients xml feeds to generate user friendly reports that outline what is wrong with the xml, and some details about how to fix any issues. This should help smooth out the process our clients go through when importing feeds, saving the support team and our clients time.

Thomas Poepping
School: Carnegie Mellon UniversityField of study: Computer Science

tpoepping_intern_scifair1. A tool to inject spreadsheets of UGC into the system.
2. An addition to the web app that generates a payroll report for all moderators, streamlining the payroll process.
3. A pilot WordPress plugin that moderates WordPress comments, then allows administrators to act on rejected comments.

Hello, World: Thoughts from an Undergrad Intern

Last summer I was given an internship opportunity in the R&D department of Bazaarvoice as a software engineer. Having only finished my freshman year in college, I had no idea what to expect from a tech company of this size. I would have never guessed that on my first day I would be handed a macbook pro with better specs than any computer I had used before. I certainly didn’t anticipate anything like the All Hands conference downtown, or the Seven Year anniversary party. And although I suspected I would pick up a few new skills during my employment, I could have never imagined the breadth of my learning over the past few months.

My first week was spent setting up my development environment, and performing the dev onboarding — writing a game of tic-tac-toe in GWT. For those of you who do not know, GWT is the Google Web Toolkit, which Wikipedia identifies as “an open source set of tools that allows web developers to create and maintain complex JavaScript front-end applications in Java.” Virtually all of my programming experience is in Java, so I was grateful that I would be able to apply that experience. However, all of the programs I had written before had been completely text based. None of my previous projects involved a graphical interface, even one as simple as tic-tac-toe. I recall each day of my first week following the same pattern — I arrived in the morning feeling overwhelmed at the alien concept I was supposed to learn that day; by the late afternoon I had struggled through enough tutorials and examples to feel that I had a decent understanding of the new skill, and before leaving the office I would begin to look at the next subject to learn, and would feel overwhelmed all over again.

Overall I feel that my first week was analogous to any of the computer science classes I have taken in school, albeit at an accelerated pace. I would write some code, and once I felt comfortable with my progress, I would show it to my mentor, who would point out the things I had done correctly, and offer advice for improvement in the areas where it was needed. Like any assignment for school, the tic-tac-toe project was exclusively for my own benefit; no one was going to use my code for anything, but there is no better teacher than practice. The project served its purpose, because by the end of my first week I was developing code for the product.

Development was a totally new experience for me. All of my previous programming involved starting from scratch, and at completion, every line of code was my own. At Bazaarvoice, I was a developer jumping into a code base thousands of lines deep, and there were a dozen other developers constantly shaping it. It was an awesome experience working as a member of a team, rather than working on a project on my own. As a team member I not only gained experience in programming, but also in the team dynamics of software development. I feel that I was able to be a boon to my team, having contributed code in the development of several features. Because I was writing code for features that the team was familiar with, it was very easy to get help from the other members of the team. I was only one degree of separation away from the developer who had the answer to one of my many questions; if the first person I asked did not know the answer, they could immediately direct me to someone who did. This high availability of assistance was a key factor in my ability to put useful code into the product. The code review system — in which every line of code was looked at by at least one other developer before entering the product’s codebase — ensured that my code conformed to the proper coding practices. Because of this, I not only improved the functionality of my code, but also the style.

My education went beyond the software itself. I attended all of the team meetings, which taught me about the “real life” side of software development. I attended meetings which showed me how the developers made sure that the product they were producing was in line with the needs of the customers. Every week involved multiple meetings to estimate the time it would take to implement various features, so that the team’s progress could be projected into the future,  and we could know how long it would take for the product to have certain degrees of functionality. The meetings were one of the most informative aspects of my internship, because they addressed issues that I had never even considered as a CS student.

This was my first internship, so while I can’t offer a personal comparison between working on a personal project and working as a team member, I can say that last summer was an enriching opportunity, which not only expanded my computer science knowledge, but taught me first hand what it means to work as a software engineer, and in my mind, that’s the most valuable thing I can get out of an internship.

This summer I have returned to Bazaarvoice, and this time I will be working on an independent project from the ground up. I am excited to see how this internship compares and contrasts to my first.