1) The document discusses challenges with big data analysis including ensuring complete data coverage from all relevant sources like devices, platforms and browser configurations.
2) It also discusses the challenge of effective monitoring to detect issues that could corrupt alerting data, giving examples of how the company Forter addresses these challenges through techniques like API monitoring and machine learning anomaly detection.
3) The key takeaways are to understand all parts of the data pipeline, log errors from both client and server, and flag any incidents affecting input data for data scientists.
2. About myself - Erez Koren
In the computer business since 2nd grade
Love building products and hacking stuff
Currently in my 3rd startup adventure
Working at Forter from before day one
2017
3. About Forter
We catch Fraudsters &
protect E-commerce merchants
Founded 3.5 years ago
~80 employees worldwide
Backed by
2017
4.
5. We detect fraud, give a real-time decision (approve/decline) every time and
guarantee it (chargeback protection). Covering the whole customer lifecycle
We collect data from browsers (JS) and mobile apps (through SDK )
We also receive order/account data S2S into our API and reply with our
decision in real-time
Our stack:
Forter - What & How We Do It
2017
Compliance:
And more...
9. Today’s internet is a jungle.
There are thousands of devices, platforms, browsers and configurations.
Are you sure you are collecting data from all / most of the relevant sources?
2017
The COVERAGE Challenge
16. 16
EXAMPLE FOR UNEXPECTED DATA IN REAL WORLD
Chorme
Safari
Mobile Safari
Firefox
IE
Android Browser
Edge
Chrome WebView
PhantomJS
undefined
Opera
WebKit
17. 17
Detect exceptions that occurs on client side
Browsers (JS), Mobile SDKs and any other client integrations
CLIENT SIDE CODE MONITORING
18. 18
JS SCRIPT TIMEOUTS
Merchant checked the website with a
browser that is not supporting javascript
Detect gaps between script request from server and script events received
19. Compare the data segments of the
general population versus the data
segment spread in your data
Test it as if you were a real user
Even if everything is working now, in
the future it will not
Takeaways
2017
21. 2017
The MONITORING Challenge
Is “measuring everything” good enough?
How often are you checking the graphs?
Do you have enough alerts or too many?
There are always technical issues that can
corrupt the alerting data
26. 26
1. Making sure we don’t slow the site down, or impact checkout funnel via
automated Selenium tests (with & without our script, multiple browsers)
2. Incremental deployment support for
JS SCRIPT MONITORING
27. 27
ML FEATURES ANOMALY DETECTION
Monitoring system’s health
by measuring our Machine
Learning features
distribution over time
30. 2017
Takeaways
Make sure every alert can be drilled down into a graph and
relate to the raw metric
Know how to investigate - leave breadcrumbs to raw data (even
when the data is aggregated)
Differentiate between critical alerts and other alerts
(that can be fixed the next morning)
Measure low values as well as the high ones - alerts for low
values (e.g. CPU) is something that most systems are
missing
31. 2017
Takeaways
Understand the pipes and filters make sure
there are no hidden blockages in the data
pipelines
Log errors both from client side and server
side when possible and analyze together
Make sure incidents that affect input data are
shared with your data scientists
by using “dirty” or “partial” flag