Tacnode™

Recommendations that lag the session

A delivery app keeps recommending burgers after a customer searches “vegan,” because the ranking reads a profile a feature job rebuilds minutes later.

Photo: Elevate / Unsplash
→→At0:12
The story

Leah searches “vegan” in her delivery app.

  • She’s hosting a vegan friend tonight. Most weeks she orders burgers or pizza.
  • The app sends her search as an event, through Segment onto a Kafka stream, like every tap.
  • Two vegan places nearby come back: Green Fork and Root & Leaf.
What Leah sees
What’s happening underneath
readsApp eventsSegmentsearch sentLeahOn the app tonightsearched “vegan”Event streamKafkanormal loadFeature jobSpark job · Airflowruns every few minutesCustomer profileFeature storeprofile: burgers, pizzaHome feedPicked for youburgers · pizza · wingsRecommendationsRanking serviceranking“vegan” + 2 taps
1 / 10

Illustrative scenario. Timings show the shape of the problem, not a measurement of any one company.

Your stack today, and with the Tacnode Context Lake™

Today

App events
Segment
Event stream
Kafka
Feature job
Spark job · Airflow
Her search waits for the next run
Customer profile
Feature store
Recommendations
Ranking service

Most recommendation stacks are built around profiles. App events flow into a stream, a job turns them into features every few minutes or every night, and the ranking service reads the result from a feature store. That works for long-term taste. It misses what a customer wants in this session, which can change within a minute, and the session is where the order is placed or lost.

With the Tacnode Context Lake™

App events
Segment
Tacnode Context Lake™
Search and taps arrive as they happen
Recommendations
Ranking service
  • Her search counted right awaySession events arrived from the app’s own event stream, with no feature job in between.
  • One snapshotTonight’s session and her order history were read as of the same moment.
  • What was shown was savedThe next ranking knows what she has already seen.

Questions to ask your team

  1. 1When a customer searches for something new, how long until your home feed reflects it?
  2. 2Do your recommendations read session signals directly, or from features a job rebuilds on a schedule?
  3. 3How often do sessions with a search end with no order, and what does the feed show after that search?