SlideShare a Scribd company logo
Mitigating User Experience from 'Breaking Bad’ 
@Twitter | PK, AK 2014 1 
The Twitter Approach 
Piyush Kumar, Arun Kejariwal 
(@pi_kumar, @arun_kejariwal) 
Performance and Capacity Engineering @ Twitter 
September 2014
@Twitter | PK, AK 2014 2 
Internet Usage 
• Increasing User Base and Data Traffic 
 >4 Billion Internet users by 2018, up from 2.5 B in 2013 [1] 
 9.2 B mobile subscriptions by 2019 [2] 
 10x growth in mobile data traffic between 2013 and 2019 [2] 
 Sales of wearable devices to reach 111.9 million units by 2018 [3] 
• Organic growth 
 Over 271 M monthly active users* 
 Over 173 B Timeline Views* 
[1] http://www.cisco.com/c/en/us/solutions/service-provider/visual-networking-index-vni/index.html#~overview (June 2014) 
[2] http://www.ericsson.com/res/docs/2014/ericsson-mobility-report-june-2014.pdf 
[3] http://online.wsj.com/articles/apple-plans-multiple-designs-for-smartwatch-1403245062 
* https://investor.twitterinc.com/releasedetail.cfm?ReleaseID=862505
@Twitter | PK, AK 2014 3 
Internet Usage (contd.) 
• Real-time consumption 
[2] 
[1] 
[1] http://socialtimes.com/twitter-succeeded-espn-failed_b199578 [2] http://online.wsj.com/articles/airlines-take-to-twitter-to-keep-world-informed-1406125253
Systematic Capacity Planning: Why Bother? 
• Evolving product landscape 
@Twitter | PK, AK 2014 4 
 Agile development 
 New products 
 Experimentation 
 New features 
 A/B Testing 
• Organic user growth 
• Growing user engagement
@Twitter | PK, AK 2014 5 
Factors Impacting Capacity Planning 
• Breakout(s) 
 Unexpected ramp up
Factors Impacting Capacity Planning (contd.) 
• Breakout(s) 
@Twitter | PK, AK 2014 6 
 Unexpected mean shift
@Twitter | PK, AK 2014 7 
Breakout Detection: How? 
• Visual 
 Large number of services 
 Not scalable 
 @ Twitter – Algorithmic approach 
 Detect breakout(s) via statistical learning
@Twitter | PK, AK 2014 8 
Breakout Detection: Why Bother? 
• Detect cold and hot services 
 Cold 
 Optimization 
 Hot 
 Higher traffic 
 Operational 
 Efficiency 
 Minimize performance impact 
• Detect Performance Regressions 
• Characterize user engagement 
 Assess an A/B Test 
 Roll out for a new platform
@Twitter | PK, AK 2014 9 
Breakout Detection (contd.) 
• Key Questions 
 How do we characterize a breakout? 
 Input 
o Time Series 
 Magnitude 
o Sustained increase or decrease 
 Direction 
o Positive – Rising services 
o Negative – Cold services 
 Account for characteristics of time series 
 Local spikes 
o Removed using rolling median 
• Commonly used in computer vision for noise reduction, called edge thinning 
 Seasonality 
o May induce false positive(s)
@Twitter | PK, AK 2014 10 
Existing Approaches 
• Support and Resistance 
 Donchian Channels 
• Limitations 
 Establishing support and resistance zones
@Twitter | PK, AK 2014 11 
Existing Approaches (contd.) 
• MACD (Moving Average Convergence Divergence) 
 Finance 
 MACD: Moving Average Convergence Divergence. Used in finance. 
o Difference between fast (12 day) and slow (26 day) EMA (Exponential Moving Average) 
 Signal 
o 9 day EMA of MACD 
 Histogram 
o MACD - Signal
@Twitter | PK, AK 2014 12 
Existing Approaches (contd.) 
• MACD (Moving Average Convergence Divergence) 
• Limitations 
 How to determine the appropriate period of MACD? 
 Inaccurate
@Twitter | PK, AK 2014 13 
Our Approach 
• Applying the change point detection algorithm 
 Rolling Median of the raw time series 
 Raw time series 
 Very susceptible to the noise or outliers 
 Moving average 
 SMA or EMA 
 Mean is susceptible to outliers 
o SMA and EMA are sensitive to noise for a small window size [1] 
[1] http: 
//www.mathworks.com/help/signal/examples/signal-smoothing.html 
Useful 
Not useful 
Not useful
@Twitter | PK, AK 2014 14 
Our Approach (contd.) 
• Step 1: Compute rolling median 
 Definition 
 Take the median of a given window size, use it as a single point. That window slides over every point in 
the time series. 
 Properties of rolling median 
 Robust to anomalies 
o Better capture the underlying trend 
o Strong noise attenuation property 
o Produces a consistent and unbiased estimate of input distribution [1] 
[1] Non Linear Signal Processing – Ganzalo R. Arce, Page 84
@Twitter | PK, AK 2014 15 
Our Approach (contd.) 
• Hypothesis testing 
To determine if a significant change point exists or not. 
 H0 – Null Hypothesis 
 No change point exists 
 H1 – Alternative Hypothesis 
 One significant change point exists 
 Test Statistic 
 Test statistic is a quantity calculated from our sample of data. 
 Its value is used to decide whether or not the null hypothesis should be rejected in our hypothesis test 
 Critical Value/Threshold 
 Threshold to which test statistic is compared to determine if the null hypothesis should be rejected.
@Twitter | PK, AK 2014 16 
Our Approach (contd.) 
• Step 2: Change-point detection 
 Hypothesis testing [Change point Detection - David Hinkley, 1970] 
 Maximize Log-likelihood Estimation (MLE) 
 Incoming data is represented by data, y1:n ∈ (y1, …, yn ) 
 For a model with a change point at τ1where τ1 ∈ {1, 2, ….n-1}, the maximum likelihood is 
 p(.) is the probability density function, Θ – MLE of parameter 
 Inference 
o Test Statistic 
H1 – Alternative 
Hypothesis 
o If λ > c, then we reject the null hypothesis, where ‘c’ is the threshold. 
o Value of c is still an open question in statistical research
@Twitter | PK, AK 2014 17 
Our Approach (contd.) 
Input time series 
• Illustration – Detecting mean shift 
Anomaly, not mean shift 
Rolling median, input to 
change-point algorithm 
Input Time series
@Twitter | PK, AK 2014 18 
Comparison evaluation 
Time Series Input 
Change point Algorithm vs. MACD and Donchian Channels 
Donchian Channels 
MACD 
Change point Algorithm
@Twitter | PK, AK 2014 19 
Our Approach (contd.) 
• Detecting positive/negative breakouts 
 Hot services (initial mean < final mean) 
 Cold services (initial mean > final mean) 
• Minimizing false positives 
 Define mean shift threshold 
 Measuring mean shift in the presence of anomalous data
@Twitter | PK, AK 2014 20 
Our Approach: Evaluation (contd.) 
• Ramp up
@Twitter | PK, AK 2014 21 
Our Approach: Evaluation (contd.) 
• Mean shift
@Twitter | PK, AK 2014 22 
Our Approach: Evaluation (contd.) 
• Detecting Breakout in presence of anomalies 
Anomaly 
Mean shift happening here, 
not visible to the human eye
@Twitter | PK, AK 2014 23 
Our Approach: Evaluation (contd.) 
• How things look like after anomaly removal 
~17% jump!
@Twitter | PK, AK 2014 24 
Our Approach: Evaluation (contd.) 
• Robust in the presence of anomalies 
Input time series with lots of spikes 
Rolling median of the input time series
@Twitter | PK, AK 2014 25 
Our Approach: Evaluation (contd.) 
• Other Data – Stocks
@Twitter | PK, AK 2014 26 
Our Approach: Evaluation (contd.) 
• Gold Price
@Twitter | PK, AK 2014 27 
Our Approach (contd.) 
• Boundary conditions and Limitations 
 Address sensitivity during the training period 
 Minimize false positives 
 Threshold change 
 Non normally distributed data
@Twitter | PK, AK 2014 28 
Future Work 
• Heuristic 
 Length of breakout segment >10% of total length 
• Learn the threshold over time
@Twitter | PK, AK 2014 29 
Acknowledgments 
• Capacity Engineering Team at Twitter
@Twitter | PK, AK 2014 30

More Related Content

What's hot

SplunkLive! Customer Presentation - Cequint
SplunkLive! Customer Presentation - CequintSplunkLive! Customer Presentation - Cequint
SplunkLive! Customer Presentation - CequintSplunk
 
7 Predictive Analytics, Spark , Streaming use cases
7 Predictive Analytics, Spark , Streaming use cases7 Predictive Analytics, Spark , Streaming use cases
7 Predictive Analytics, Spark , Streaming use cases
DataWorks Summit/Hadoop Summit
 
Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...
Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...
Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...
Impetus Technologies
 
Kingspan sensorv1.2 05th October 2015
Kingspan sensorv1.2 05th October 2015Kingspan sensorv1.2 05th October 2015
Kingspan sensorv1.2 05th October 2015
Smart Dublin
 
Open Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, Paris
Open Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, ParisOpen Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, Paris
Open Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, Paris
OW2
 
Use Machine Learning to Get the Most out of Your Big Data Clusters
Use Machine Learning to Get the Most out of Your Big Data ClustersUse Machine Learning to Get the Most out of Your Big Data Clusters
Use Machine Learning to Get the Most out of Your Big Data Clusters
Databricks
 
Big image analytics for (Re-) insurer
 Big image analytics for (Re-) insurer Big image analytics for (Re-) insurer
Big image analytics for (Re-) insurer
Flavio Trolese
 
How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...
How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...
How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...
Kai Wähner
 
Finding Changes in Real Data
Finding Changes in Real DataFinding Changes in Real Data
Finding Changes in Real Data
Ted Dunning
 
Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...
Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...
Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...
Impetus Technologies
 
Innovating With Data and Analytics
Innovating With Data and AnalyticsInnovating With Data and Analytics
Innovating With Data and Analytics
VMware Tanzu
 
Real-time Puppies and Ponies - Evolving Indicator Recommendations in Real-time
Real-time Puppies and Ponies - Evolving Indicator Recommendations in Real-timeReal-time Puppies and Ponies - Evolving Indicator Recommendations in Real-time
Real-time Puppies and Ponies - Evolving Indicator Recommendations in Real-time
Ted Dunning
 
Structuring EMR Data For Analytics
Structuring EMR Data For AnalyticsStructuring EMR Data For Analytics
Structuring EMR Data For Analytics
Brandon Stange
 
Cheap learning-dunning-9-18-2015
Cheap learning-dunning-9-18-2015Cheap learning-dunning-9-18-2015
Cheap learning-dunning-9-18-2015
Ted Dunning
 
Level Up Your Security Skills in Splunk Enterprise
Level Up Your Security Skills in Splunk EnterpriseLevel Up Your Security Skills in Splunk Enterprise
Level Up Your Security Skills in Splunk Enterprise
Splunk
 
Deep Learning for Public Safety in Chicago and San Francisco
Deep Learning for Public Safety in Chicago and San FranciscoDeep Learning for Public Safety in Chicago and San Francisco
Deep Learning for Public Safety in Chicago and San Francisco
Sri Ambati
 
big image analytics- ai meets big data
big image analytics- ai meets big databig image analytics- ai meets big data
big image analytics- ai meets big data
FIAT/IFTA
 
Storing Sensor data using User Defined Types
Storing Sensor data using User Defined TypesStoring Sensor data using User Defined Types
Storing Sensor data using User Defined Types
Andy Cobley
 

What's hot (18)

SplunkLive! Customer Presentation - Cequint
SplunkLive! Customer Presentation - CequintSplunkLive! Customer Presentation - Cequint
SplunkLive! Customer Presentation - Cequint
 
7 Predictive Analytics, Spark , Streaming use cases
7 Predictive Analytics, Spark , Streaming use cases7 Predictive Analytics, Spark , Streaming use cases
7 Predictive Analytics, Spark , Streaming use cases
 
Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...
Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...
Real-time Streaming Analytics: Business Value, Use Cases and Architectural Co...
 
Kingspan sensorv1.2 05th October 2015
Kingspan sensorv1.2 05th October 2015Kingspan sensorv1.2 05th October 2015
Kingspan sensorv1.2 05th October 2015
 
Open Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, Paris
Open Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, ParisOpen Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, Paris
Open Source and Data-Driven Economy, OW2con'18, June 7-8, 2018, Paris
 
Use Machine Learning to Get the Most out of Your Big Data Clusters
Use Machine Learning to Get the Most out of Your Big Data ClustersUse Machine Learning to Get the Most out of Your Big Data Clusters
Use Machine Learning to Get the Most out of Your Big Data Clusters
 
Big image analytics for (Re-) insurer
 Big image analytics for (Re-) insurer Big image analytics for (Re-) insurer
Big image analytics for (Re-) insurer
 
How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...
How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...
How to Apply Machine Learning with R, H20, Apache Spark MLlib or PMML to Real...
 
Finding Changes in Real Data
Finding Changes in Real DataFinding Changes in Real Data
Finding Changes in Real Data
 
Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...
Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...
Real-time Streaming Analytics for Enterprises based on Apache Storm - Impetus...
 
Innovating With Data and Analytics
Innovating With Data and AnalyticsInnovating With Data and Analytics
Innovating With Data and Analytics
 
Real-time Puppies and Ponies - Evolving Indicator Recommendations in Real-time
Real-time Puppies and Ponies - Evolving Indicator Recommendations in Real-timeReal-time Puppies and Ponies - Evolving Indicator Recommendations in Real-time
Real-time Puppies and Ponies - Evolving Indicator Recommendations in Real-time
 
Structuring EMR Data For Analytics
Structuring EMR Data For AnalyticsStructuring EMR Data For Analytics
Structuring EMR Data For Analytics
 
Cheap learning-dunning-9-18-2015
Cheap learning-dunning-9-18-2015Cheap learning-dunning-9-18-2015
Cheap learning-dunning-9-18-2015
 
Level Up Your Security Skills in Splunk Enterprise
Level Up Your Security Skills in Splunk EnterpriseLevel Up Your Security Skills in Splunk Enterprise
Level Up Your Security Skills in Splunk Enterprise
 
Deep Learning for Public Safety in Chicago and San Francisco
Deep Learning for Public Safety in Chicago and San FranciscoDeep Learning for Public Safety in Chicago and San Francisco
Deep Learning for Public Safety in Chicago and San Francisco
 
big image analytics- ai meets big data
big image analytics- ai meets big databig image analytics- ai meets big data
big image analytics- ai meets big data
 
Storing Sensor data using User Defined Types
Storing Sensor data using User Defined TypesStoring Sensor data using User Defined Types
Storing Sensor data using User Defined Types
 

Viewers also liked

Github pagesでRPubsにサヨナラ!
Github pagesでRPubsにサヨナラ!Github pagesでRPubsにサヨナラ!
Github pagesでRPubsにサヨナラ!
yutannihilation
 
Com t'ho explico
Com t'ho explicoCom t'ho explico
匿名バイナリ配布集団rwinlib
匿名バイナリ配布集団rwinlib匿名バイナリ配布集団rwinlib
匿名バイナリ配布集団rwinlib
yutannihilation
 
Anteojito, revista completa, agosto 11 de 1966
Anteojito, revista completa,  agosto 11 de 1966Anteojito, revista completa,  agosto 11 de 1966
Anteojito, revista completa, agosto 11 de 1966
Martin Alberto Belaustegui
 
Tric y Trake 15 junio 1967
Tric y Trake  15 junio 1967Tric y Trake  15 junio 1967
Tric y Trake 15 junio 1967
Martin Alberto Belaustegui
 
Mapa ilustrado de Estados Unidos
Mapa ilustrado de Estados Unidos Mapa ilustrado de Estados Unidos
Mapa ilustrado de Estados Unidos
Martin Alberto Belaustegui
 
La diversidad de los seres vivos
La diversidad de los seres vivosLa diversidad de los seres vivos
La diversidad de los seres vivos
Martin Alberto Belaustegui
 
A Systematic Approach to Capacity Planning in the Real World
A Systematic Approach to Capacity Planning in the Real WorldA Systematic Approach to Capacity Planning in the Real World
A Systematic Approach to Capacity Planning in the Real World
Arun Kejariwal
 
A Tool for Practical Garbage Collection Analysis In the Cloud
A Tool for Practical Garbage Collection Analysis In the CloudA Tool for Practical Garbage Collection Analysis In the Cloud
A Tool for Practical Garbage Collection Analysis In the CloudArun Kejariwal
 
Mistery box
Mistery boxMistery box
山水
山水山水
山水
julia chuang
 
Formació en competències
Formació en competènciesFormació en competències
Formació en competències
CESIRE - Dept d'Educació - GENCAT
 
押してダメなら引いてみろ! ggplot2逆引きプロジェクト
押してダメなら引いてみろ! ggplot2逆引きプロジェクト押してダメなら引いてみろ! ggplot2逆引きプロジェクト
押してダメなら引いてみろ! ggplot2逆引きプロジェクト
yutannihilation
 
dplyrとは何だったのか
dplyrとは何だったのかdplyrとは何だったのか
dplyrとは何だったのか
yutannihilation
 
St patricks day
St patricks daySt patricks day
St patricks day
begomesonada
 
TeX原稿からEPUBを作りたい
TeX原稿からEPUBを作りたいTeX原稿からEPUBを作りたい
TeX原稿からEPUBを作りたい
Keiichiro Shikano
 
東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法
東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法
東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法
Nagi Teramo
 
Tidyverseとは
TidyverseとはTidyverseとは
Tidyverseとは
yutannihilation
 

Viewers also liked (20)

CTFとは
CTFとはCTFとは
CTFとは
 
Github pagesでRPubsにサヨナラ!
Github pagesでRPubsにサヨナラ!Github pagesでRPubsにサヨナラ!
Github pagesでRPubsにサヨナラ!
 
Com t'ho explico
Com t'ho explicoCom t'ho explico
Com t'ho explico
 
匿名バイナリ配布集団rwinlib
匿名バイナリ配布集団rwinlib匿名バイナリ配布集団rwinlib
匿名バイナリ配布集団rwinlib
 
Anteojito, revista completa, agosto 11 de 1966
Anteojito, revista completa,  agosto 11 de 1966Anteojito, revista completa,  agosto 11 de 1966
Anteojito, revista completa, agosto 11 de 1966
 
Tric y Trake 15 junio 1967
Tric y Trake  15 junio 1967Tric y Trake  15 junio 1967
Tric y Trake 15 junio 1967
 
Mapa ilustrado de Estados Unidos
Mapa ilustrado de Estados Unidos Mapa ilustrado de Estados Unidos
Mapa ilustrado de Estados Unidos
 
La diversidad de los seres vivos
La diversidad de los seres vivosLa diversidad de los seres vivos
La diversidad de los seres vivos
 
A Systematic Approach to Capacity Planning in the Real World
A Systematic Approach to Capacity Planning in the Real WorldA Systematic Approach to Capacity Planning in the Real World
A Systematic Approach to Capacity Planning in the Real World
 
A Tool for Practical Garbage Collection Analysis In the Cloud
A Tool for Practical Garbage Collection Analysis In the CloudA Tool for Practical Garbage Collection Analysis In the Cloud
A Tool for Practical Garbage Collection Analysis In the Cloud
 
Mistery box
Mistery boxMistery box
Mistery box
 
山水
山水山水
山水
 
Formació en competències
Formació en competènciesFormació en competències
Formació en competències
 
押してダメなら引いてみろ! ggplot2逆引きプロジェクト
押してダメなら引いてみろ! ggplot2逆引きプロジェクト押してダメなら引いてみろ! ggplot2逆引きプロジェクト
押してダメなら引いてみろ! ggplot2逆引きプロジェクト
 
dplyrとは何だったのか
dplyrとは何だったのかdplyrとは何だったのか
dplyrとは何だったのか
 
St patricks day
St patricks daySt patricks day
St patricks day
 
TeX原稿からEPUBを作りたい
TeX原稿からEPUBを作りたいTeX原稿からEPUBを作りたい
TeX原稿からEPUBを作りたい
 
東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法
東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法
東京R非公式おじさんが教える本当に気持ちいいパッケージ作成法
 
Tidyverseとは
TidyverseとはTidyverseとは
Tidyverseとは
 
RをAWSで使おう
RをAWSで使おうRをAWSで使おう
RをAWSで使おう
 

Similar to Mitigating User Experience from 'Breaking Bad': The Twitter Approach [Velocity New York 2014]

Sequence-to-Sequence Modeling for Time Series
Sequence-to-Sequence Modeling for Time SeriesSequence-to-Sequence Modeling for Time Series
Sequence-to-Sequence Modeling for Time Series
Arun Kejariwal
 
Modelling and Detecting Changes in User Satisfaction
Modelling and Detecting Changes in User SatisfactionModelling and Detecting Changes in User Satisfaction
Modelling and Detecting Changes in User Satisfaction
Julia Kiseleva
 
Change management infra basic - quick guideline v1.1 tj
Change management infra basic - quick guideline v1.1 tjChange management infra basic - quick guideline v1.1 tj
Change management infra basic - quick guideline v1.1 tj
Tijs -T.J.- van Velthoven, MBA - AVAILABLE
 
Powerpoint RM 2022.pdf
Powerpoint RM 2022.pdfPowerpoint RM 2022.pdf
Powerpoint RM 2022.pdf
AlEmranIsmail1
 
LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC
LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC
LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC
Julian Kalac P.Eng
 
Case Study: It’s All About Data – And the Customer
Case Study: It’s All About Data – And the CustomerCase Study: It’s All About Data – And the Customer
Case Study: It’s All About Data – And the Customer
Jill Kirkpatrick
 
AITPM Conference Presentation -Callan Stirzaker
AITPM Conference Presentation -Callan StirzakerAITPM Conference Presentation -Callan Stirzaker
AITPM Conference Presentation -Callan Stirzaker
JumpingJaq
 
Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...
Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...
Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...
Peter Lindgren
 
SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...
SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...
SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...
Michael Kehoe
 
Are we there yet?
Are we there yet?Are we there yet?
Are we there yet?
Mike Burrows
 
Methodology Perspective and Narrative
Methodology Perspective and NarrativeMethodology Perspective and Narrative
Methodology Perspective and Narrative
Larry Dalton
 
ntroduction of Signal such as sinosoidal signals, definition of signals
ntroduction of Signal such as sinosoidal signals, definition of signalsntroduction of Signal such as sinosoidal signals, definition of signals
ntroduction of Signal such as sinosoidal signals, definition of signals
DrAjayKumarYadav4
 
'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014
'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014
'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014
ColomboCampsCommunity
 
WECC_CIPC_CIPv5_SurveyReport_JBB_12152015
WECC_CIPC_CIPv5_SurveyReport_JBB_12152015WECC_CIPC_CIPv5_SurveyReport_JBB_12152015
WECC_CIPC_CIPv5_SurveyReport_JBB_12152015Dr. Joseph Baugh, PMP
 
Value Stream Mapping
Value Stream MappingValue Stream Mapping
Value Stream Mapping
TKMG, Inc.
 
Teaching with Tableau
Teaching with TableauTeaching with Tableau
Teaching with Tableau
Kristen Sosulski
 
Value Stream Mapping – Stories From the Trenches
Value Stream Mapping – Stories From the TrenchesValue Stream Mapping – Stories From the Trenches
Value Stream Mapping – Stories From the Trenches
DevOps.com
 
Estimating and planning Agile projects
Estimating and planning Agile projectsEstimating and planning Agile projects
Estimating and planning Agile projects
Murray Robinson
 
Improve phase lean six sigma tollgate template
Improve phase   lean six sigma tollgate templateImprove phase   lean six sigma tollgate template
Improve phase lean six sigma tollgate templateSteven Bonacorsi
 
Improve phase lean six sigma tollgate template
Improve phase   lean six sigma tollgate templateImprove phase   lean six sigma tollgate template
Improve phase lean six sigma tollgate templateSteven Bonacorsi
 

Similar to Mitigating User Experience from 'Breaking Bad': The Twitter Approach [Velocity New York 2014] (20)

Sequence-to-Sequence Modeling for Time Series
Sequence-to-Sequence Modeling for Time SeriesSequence-to-Sequence Modeling for Time Series
Sequence-to-Sequence Modeling for Time Series
 
Modelling and Detecting Changes in User Satisfaction
Modelling and Detecting Changes in User SatisfactionModelling and Detecting Changes in User Satisfaction
Modelling and Detecting Changes in User Satisfaction
 
Change management infra basic - quick guideline v1.1 tj
Change management infra basic - quick guideline v1.1 tjChange management infra basic - quick guideline v1.1 tj
Change management infra basic - quick guideline v1.1 tj
 
Powerpoint RM 2022.pdf
Powerpoint RM 2022.pdfPowerpoint RM 2022.pdf
Powerpoint RM 2022.pdf
 
LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC
LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC
LEAN SPEED vs SIX SIGMA QUALITY by JULIAN KALAC
 
Case Study: It’s All About Data – And the Customer
Case Study: It’s All About Data – And the CustomerCase Study: It’s All About Data – And the Customer
Case Study: It’s All About Data – And the Customer
 
AITPM Conference Presentation -Callan Stirzaker
AITPM Conference Presentation -Callan StirzakerAITPM Conference Presentation -Callan Stirzaker
AITPM Conference Presentation -Callan Stirzaker
 
Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...
Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...
Catch! Workshop concept 2 - Improving travel plan monitoring with better, mor...
 
SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...
SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...
SRECon-Europe-2017: Reducing MTTR and False Escalations: Event Correlation at...
 
Are we there yet?
Are we there yet?Are we there yet?
Are we there yet?
 
Methodology Perspective and Narrative
Methodology Perspective and NarrativeMethodology Perspective and Narrative
Methodology Perspective and Narrative
 
ntroduction of Signal such as sinosoidal signals, definition of signals
ntroduction of Signal such as sinosoidal signals, definition of signalsntroduction of Signal such as sinosoidal signals, definition of signals
ntroduction of Signal such as sinosoidal signals, definition of signals
 
'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014
'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014
'Metrics That Matter': Gabrielle Benefield @ Colombo Agile Con 2014
 
WECC_CIPC_CIPv5_SurveyReport_JBB_12152015
WECC_CIPC_CIPv5_SurveyReport_JBB_12152015WECC_CIPC_CIPv5_SurveyReport_JBB_12152015
WECC_CIPC_CIPv5_SurveyReport_JBB_12152015
 
Value Stream Mapping
Value Stream MappingValue Stream Mapping
Value Stream Mapping
 
Teaching with Tableau
Teaching with TableauTeaching with Tableau
Teaching with Tableau
 
Value Stream Mapping – Stories From the Trenches
Value Stream Mapping – Stories From the TrenchesValue Stream Mapping – Stories From the Trenches
Value Stream Mapping – Stories From the Trenches
 
Estimating and planning Agile projects
Estimating and planning Agile projectsEstimating and planning Agile projects
Estimating and planning Agile projects
 
Improve phase lean six sigma tollgate template
Improve phase   lean six sigma tollgate templateImprove phase   lean six sigma tollgate template
Improve phase lean six sigma tollgate template
 
Improve phase lean six sigma tollgate template
Improve phase   lean six sigma tollgate templateImprove phase   lean six sigma tollgate template
Improve phase lean six sigma tollgate template
 

Recently uploaded

FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdfFIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance
 
Free Complete Python - A step towards Data Science
Free Complete Python - A step towards Data ScienceFree Complete Python - A step towards Data Science
Free Complete Python - A step towards Data Science
RinaMondal9
 
Secstrike : Reverse Engineering & Pwnable tools for CTF.pptx
Secstrike : Reverse Engineering & Pwnable tools for CTF.pptxSecstrike : Reverse Engineering & Pwnable tools for CTF.pptx
Secstrike : Reverse Engineering & Pwnable tools for CTF.pptx
nkrafacyberclub
 
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
SOFTTECHHUB
 
Video Streaming: Then, Now, and in the Future
Video Streaming: Then, Now, and in the FutureVideo Streaming: Then, Now, and in the Future
Video Streaming: Then, Now, and in the Future
Alpen-Adria-Universität
 
Removing Uninteresting Bytes in Software Fuzzing
Removing Uninteresting Bytes in Software FuzzingRemoving Uninteresting Bytes in Software Fuzzing
Removing Uninteresting Bytes in Software Fuzzing
Aftab Hussain
 
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdfSmart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
91mobiles
 
PCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase TeamPCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase Team
ControlCase
 
The Future of Platform Engineering
The Future of Platform EngineeringThe Future of Platform Engineering
The Future of Platform Engineering
Jemma Hussein Allen
 
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
DanBrown980551
 
Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !
KatiaHIMEUR1
 
GraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge GraphGraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge Graph
Guy Korland
 
Encryption in Microsoft 365 - ExpertsLive Netherlands 2024
Encryption in Microsoft 365 - ExpertsLive Netherlands 2024Encryption in Microsoft 365 - ExpertsLive Netherlands 2024
Encryption in Microsoft 365 - ExpertsLive Netherlands 2024
Albert Hoitingh
 
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
James Anderson
 
DevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA ConnectDevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA Connect
Kari Kakkonen
 
Microsoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdfMicrosoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdf
Uni Systems S.M.S.A.
 
Introduction to CHERI technology - Cybersecurity
Introduction to CHERI technology - CybersecurityIntroduction to CHERI technology - Cybersecurity
Introduction to CHERI technology - Cybersecurity
mikeeftimakis1
 
Climate Impact of Software Testing at Nordic Testing Days
Climate Impact of Software Testing at Nordic Testing DaysClimate Impact of Software Testing at Nordic Testing Days
Climate Impact of Software Testing at Nordic Testing Days
Kari Kakkonen
 
Communications Mining Series - Zero to Hero - Session 1
Communications Mining Series - Zero to Hero - Session 1Communications Mining Series - Zero to Hero - Session 1
Communications Mining Series - Zero to Hero - Session 1
DianaGray10
 
20240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 202420240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 2024
Matthew Sinclair
 

Recently uploaded (20)

FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdfFIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
FIDO Alliance Osaka Seminar: Passkeys and the Road Ahead.pdf
 
Free Complete Python - A step towards Data Science
Free Complete Python - A step towards Data ScienceFree Complete Python - A step towards Data Science
Free Complete Python - A step towards Data Science
 
Secstrike : Reverse Engineering & Pwnable tools for CTF.pptx
Secstrike : Reverse Engineering & Pwnable tools for CTF.pptxSecstrike : Reverse Engineering & Pwnable tools for CTF.pptx
Secstrike : Reverse Engineering & Pwnable tools for CTF.pptx
 
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
Why You Should Replace Windows 11 with Nitrux Linux 3.5.0 for enhanced perfor...
 
Video Streaming: Then, Now, and in the Future
Video Streaming: Then, Now, and in the FutureVideo Streaming: Then, Now, and in the Future
Video Streaming: Then, Now, and in the Future
 
Removing Uninteresting Bytes in Software Fuzzing
Removing Uninteresting Bytes in Software FuzzingRemoving Uninteresting Bytes in Software Fuzzing
Removing Uninteresting Bytes in Software Fuzzing
 
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdfSmart TV Buyer Insights Survey 2024 by 91mobiles.pdf
Smart TV Buyer Insights Survey 2024 by 91mobiles.pdf
 
PCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase TeamPCI PIN Basics Webinar from the Controlcase Team
PCI PIN Basics Webinar from the Controlcase Team
 
The Future of Platform Engineering
The Future of Platform EngineeringThe Future of Platform Engineering
The Future of Platform Engineering
 
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
LF Energy Webinar: Electrical Grid Modelling and Simulation Through PowSyBl -...
 
Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !Securing your Kubernetes cluster_ a step-by-step guide to success !
Securing your Kubernetes cluster_ a step-by-step guide to success !
 
GraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge GraphGraphRAG is All You need? LLM & Knowledge Graph
GraphRAG is All You need? LLM & Knowledge Graph
 
Encryption in Microsoft 365 - ExpertsLive Netherlands 2024
Encryption in Microsoft 365 - ExpertsLive Netherlands 2024Encryption in Microsoft 365 - ExpertsLive Netherlands 2024
Encryption in Microsoft 365 - ExpertsLive Netherlands 2024
 
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
Alt. GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using ...
 
DevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA ConnectDevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA Connect
 
Microsoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdfMicrosoft - Power Platform_G.Aspiotis.pdf
Microsoft - Power Platform_G.Aspiotis.pdf
 
Introduction to CHERI technology - Cybersecurity
Introduction to CHERI technology - CybersecurityIntroduction to CHERI technology - Cybersecurity
Introduction to CHERI technology - Cybersecurity
 
Climate Impact of Software Testing at Nordic Testing Days
Climate Impact of Software Testing at Nordic Testing DaysClimate Impact of Software Testing at Nordic Testing Days
Climate Impact of Software Testing at Nordic Testing Days
 
Communications Mining Series - Zero to Hero - Session 1
Communications Mining Series - Zero to Hero - Session 1Communications Mining Series - Zero to Hero - Session 1
Communications Mining Series - Zero to Hero - Session 1
 
20240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 202420240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 2024
 

Mitigating User Experience from 'Breaking Bad': The Twitter Approach [Velocity New York 2014]

  • 1. Mitigating User Experience from 'Breaking Bad’ @Twitter | PK, AK 2014 1 The Twitter Approach Piyush Kumar, Arun Kejariwal (@pi_kumar, @arun_kejariwal) Performance and Capacity Engineering @ Twitter September 2014
  • 2. @Twitter | PK, AK 2014 2 Internet Usage • Increasing User Base and Data Traffic  >4 Billion Internet users by 2018, up from 2.5 B in 2013 [1]  9.2 B mobile subscriptions by 2019 [2]  10x growth in mobile data traffic between 2013 and 2019 [2]  Sales of wearable devices to reach 111.9 million units by 2018 [3] • Organic growth  Over 271 M monthly active users*  Over 173 B Timeline Views* [1] http://www.cisco.com/c/en/us/solutions/service-provider/visual-networking-index-vni/index.html#~overview (June 2014) [2] http://www.ericsson.com/res/docs/2014/ericsson-mobility-report-june-2014.pdf [3] http://online.wsj.com/articles/apple-plans-multiple-designs-for-smartwatch-1403245062 * https://investor.twitterinc.com/releasedetail.cfm?ReleaseID=862505
  • 3. @Twitter | PK, AK 2014 3 Internet Usage (contd.) • Real-time consumption [2] [1] [1] http://socialtimes.com/twitter-succeeded-espn-failed_b199578 [2] http://online.wsj.com/articles/airlines-take-to-twitter-to-keep-world-informed-1406125253
  • 4. Systematic Capacity Planning: Why Bother? • Evolving product landscape @Twitter | PK, AK 2014 4  Agile development  New products  Experimentation  New features  A/B Testing • Organic user growth • Growing user engagement
  • 5. @Twitter | PK, AK 2014 5 Factors Impacting Capacity Planning • Breakout(s)  Unexpected ramp up
  • 6. Factors Impacting Capacity Planning (contd.) • Breakout(s) @Twitter | PK, AK 2014 6  Unexpected mean shift
  • 7. @Twitter | PK, AK 2014 7 Breakout Detection: How? • Visual  Large number of services  Not scalable  @ Twitter – Algorithmic approach  Detect breakout(s) via statistical learning
  • 8. @Twitter | PK, AK 2014 8 Breakout Detection: Why Bother? • Detect cold and hot services  Cold  Optimization  Hot  Higher traffic  Operational  Efficiency  Minimize performance impact • Detect Performance Regressions • Characterize user engagement  Assess an A/B Test  Roll out for a new platform
  • 9. @Twitter | PK, AK 2014 9 Breakout Detection (contd.) • Key Questions  How do we characterize a breakout?  Input o Time Series  Magnitude o Sustained increase or decrease  Direction o Positive – Rising services o Negative – Cold services  Account for characteristics of time series  Local spikes o Removed using rolling median • Commonly used in computer vision for noise reduction, called edge thinning  Seasonality o May induce false positive(s)
  • 10. @Twitter | PK, AK 2014 10 Existing Approaches • Support and Resistance  Donchian Channels • Limitations  Establishing support and resistance zones
  • 11. @Twitter | PK, AK 2014 11 Existing Approaches (contd.) • MACD (Moving Average Convergence Divergence)  Finance  MACD: Moving Average Convergence Divergence. Used in finance. o Difference between fast (12 day) and slow (26 day) EMA (Exponential Moving Average)  Signal o 9 day EMA of MACD  Histogram o MACD - Signal
  • 12. @Twitter | PK, AK 2014 12 Existing Approaches (contd.) • MACD (Moving Average Convergence Divergence) • Limitations  How to determine the appropriate period of MACD?  Inaccurate
  • 13. @Twitter | PK, AK 2014 13 Our Approach • Applying the change point detection algorithm  Rolling Median of the raw time series  Raw time series  Very susceptible to the noise or outliers  Moving average  SMA or EMA  Mean is susceptible to outliers o SMA and EMA are sensitive to noise for a small window size [1] [1] http: //www.mathworks.com/help/signal/examples/signal-smoothing.html Useful Not useful Not useful
  • 14. @Twitter | PK, AK 2014 14 Our Approach (contd.) • Step 1: Compute rolling median  Definition  Take the median of a given window size, use it as a single point. That window slides over every point in the time series.  Properties of rolling median  Robust to anomalies o Better capture the underlying trend o Strong noise attenuation property o Produces a consistent and unbiased estimate of input distribution [1] [1] Non Linear Signal Processing – Ganzalo R. Arce, Page 84
  • 15. @Twitter | PK, AK 2014 15 Our Approach (contd.) • Hypothesis testing To determine if a significant change point exists or not.  H0 – Null Hypothesis  No change point exists  H1 – Alternative Hypothesis  One significant change point exists  Test Statistic  Test statistic is a quantity calculated from our sample of data.  Its value is used to decide whether or not the null hypothesis should be rejected in our hypothesis test  Critical Value/Threshold  Threshold to which test statistic is compared to determine if the null hypothesis should be rejected.
  • 16. @Twitter | PK, AK 2014 16 Our Approach (contd.) • Step 2: Change-point detection  Hypothesis testing [Change point Detection - David Hinkley, 1970]  Maximize Log-likelihood Estimation (MLE)  Incoming data is represented by data, y1:n ∈ (y1, …, yn )  For a model with a change point at τ1where τ1 ∈ {1, 2, ….n-1}, the maximum likelihood is  p(.) is the probability density function, Θ – MLE of parameter  Inference o Test Statistic H1 – Alternative Hypothesis o If λ > c, then we reject the null hypothesis, where ‘c’ is the threshold. o Value of c is still an open question in statistical research
  • 17. @Twitter | PK, AK 2014 17 Our Approach (contd.) Input time series • Illustration – Detecting mean shift Anomaly, not mean shift Rolling median, input to change-point algorithm Input Time series
  • 18. @Twitter | PK, AK 2014 18 Comparison evaluation Time Series Input Change point Algorithm vs. MACD and Donchian Channels Donchian Channels MACD Change point Algorithm
  • 19. @Twitter | PK, AK 2014 19 Our Approach (contd.) • Detecting positive/negative breakouts  Hot services (initial mean < final mean)  Cold services (initial mean > final mean) • Minimizing false positives  Define mean shift threshold  Measuring mean shift in the presence of anomalous data
  • 20. @Twitter | PK, AK 2014 20 Our Approach: Evaluation (contd.) • Ramp up
  • 21. @Twitter | PK, AK 2014 21 Our Approach: Evaluation (contd.) • Mean shift
  • 22. @Twitter | PK, AK 2014 22 Our Approach: Evaluation (contd.) • Detecting Breakout in presence of anomalies Anomaly Mean shift happening here, not visible to the human eye
  • 23. @Twitter | PK, AK 2014 23 Our Approach: Evaluation (contd.) • How things look like after anomaly removal ~17% jump!
  • 24. @Twitter | PK, AK 2014 24 Our Approach: Evaluation (contd.) • Robust in the presence of anomalies Input time series with lots of spikes Rolling median of the input time series
  • 25. @Twitter | PK, AK 2014 25 Our Approach: Evaluation (contd.) • Other Data – Stocks
  • 26. @Twitter | PK, AK 2014 26 Our Approach: Evaluation (contd.) • Gold Price
  • 27. @Twitter | PK, AK 2014 27 Our Approach (contd.) • Boundary conditions and Limitations  Address sensitivity during the training period  Minimize false positives  Threshold change  Non normally distributed data
  • 28. @Twitter | PK, AK 2014 28 Future Work • Heuristic  Length of breakout segment >10% of total length • Learn the threshold over time
  • 29. @Twitter | PK, AK 2014 29 Acknowledgments • Capacity Engineering Team at Twitter
  • 30. @Twitter | PK, AK 2014 30