Tracking All the Things Chris Powers Engineering Manager at Groupon

Where We Were (Tech) Tracking Chaos! 20-30 tracking pixels per page Tracking servers on the brink of falling over Multiple tracking systems, schemas across platforms Unable to collect rich data

Where We Were (Org) Organizational Chaos! No central ownership, no oversight Little communication/collaboration between trackers Multiple tracking systems across organization Different groups with different data requirements

Data Consumers CEO: Key Business Metrics Marketing: Attribution, SEM Spend PM/Developers: Feature Metrics, Lift Data Scientists: Data Spelunking

Where do we track?

API Desktop Browsers Mobile Browsers iOS Android Web App Touch App Backend Service Backend Service Email

Browsers Mobile API Services Email Transform Data Warehouse Real-time Analytics Ad Hoc Queries Raw Store UAT, Staging, Production

What do we track?

Tracking Dichotomies Centralized vs. Decentralized (Core vs. Ad Hoc) Explicit vs. Implicit Fat vs. Skinny Client-side vs. Server-side Logic

Message Overhead Client/Server Event IDs Client/Server Timestamps Client Message Index Schema ID App ID - where did the message come from? Message Encoding - To JSON, or not to JSON?

TIME User Session Session Page Page Page Page Page Page Page TRACKING EVENTS

User/Device Data User Identifier User Agent (parsed or raw) Metadata: Number visits, kind of user, etc. Logged in / logged out Leave tracking cookie after logout?

Sessionization Session ID (Timestamp + User ID) Session Expiry Logic Set cookie w/ wildcard domain, time limit (not session) Session ID will be used as a key for client data Referral Data

Page View Data What constitutes a “page view”? AJAX? Mobile? Page ID (Session ID + Timestamp) Type (what kind of a page) URL (parsed or not?) Country/Locale (can this change in a session?)

Page View Data, Cont. App Specific Metadata Referring Page ID, Referring Click ID Other app specific attribution logic Tracking Library Version Beware the Back Button, Reload & Browser Cache

How do we track?

Explicit Tracking TrackingHub - simple one liner tracking TrackingHub.add(“msgName”, {some: “data”})

Implicit Tracking Bloodhound - Easy impression/click tracking

Importance of Visualizations Make the (nearly) invisible visible! Improve adoption from non-tech team members Improve adoption from developers too! Provide visibility outside of engineering/product org

Delivering Client Messages Batch messages to reduce overhead, server load Persist message cache across page loads Use retry logic if delivery fails Options for encoding batches Origin switches are a pain point (http/https, subdomain)

Persistence Tips Version all messages Create migration functions to upgrade data to newest version Isolate top-level localStorage keys to reduce churn Abstract away storage engine (cookie, localStorage, db) Track storage usage, fire warning messages Purge old data (beware of ghosts)

Data Verification Surprisingly hard to verify correctness Define success metrics and loss thresholds up front You will have data loss - how will you measure it? Identify points throughout pipe where validation can occur Consider a tracking pixel as a point of comparison Use message indexes to identify dropped messages

Testing the Tracking Unit Test the tracking components and all “core” tracking Use a crawler to fire messages, look for them in data stores Keep realtime alerts looking for missing/malformed keys Develop process when quality is questioned, don’t panic!

Tracking Security Nothing from the client can be fully trusted. Bots will run JS, and they will do weird things. Be prepared and able to throw out data by session/user/ip. Know your users’ patterns, and identify outliers. Be prepared to block IPs from the tracking endpoint.

Questions? @chrisjpowers