Redis Real-Time Decisioning Demo | Ad Server Revenue Optimization

Stage 1: The Architecture

StreamSight Media's Real-Time Decisioning Stack

Five tiers form a continuous loop: ingest, context, decide, act, learn. This is not just ad selection. It is revenue-aware decisioning under strict latency and policy constraints. Systems of record stay where they are, including Snowflake as the customer preference repository for likes, dislikes, and declared content interests. Redis Feature Form turns Snowflake preference data and live stream signals into online features. Redis RAM serves the hot ad-break working set. Redis Flex holds broader household history, preference vectors, and campaign context. Together, they make those preferences actionable during the ad break.

Data Sources

Ad Inventory Platform

Operational source

Campaign Management System

Operational source

CDP / Identity Graph

Household and device identity

Snowflake Preferences

Customer likes, dislikes, genres, and brand exclusions

Content Metadata Store

Content genre, ratings, and placement context

Measurement / Attribution APIs

Operational source

Kafka / Viewing Events

Operational source

Ingest Layer

Redis Data Integration (RDI)

Activates campaign, inventory, identity, and Snowflake preference data into Redis

Redis Feature Form

Online feature serving from Snowflake preferences, campaign data, and live viewing events with train-serve parity

Unified Context Layer

Redis RAM

Hot ad-break working set, live session state, active frequency caps, and real-time bid signals

Redis Flex

Warm household viewing history, preference vectors, campaign response patterns, and identity graph context

Feature Store

Brand fit, ad tolerance, conversion propensity, and revenue yield optimization features

Redis Context Retriever

Assembles the Campaign 360 — audience state, inventory context, and bid history — and exposes it as structured MCP tools for the decision engine

Decision Engine

Eligibility Rules

Policy, compliance, and inventory constraints

NBA Ranker

Contextual ranking weighted by value, risk, and fit

Policy Arbitration

One action selected for the current moment

Output Channels (Act)

CTV App

Customer-facing activation surface

Mobile Streaming App

Customer-facing activation surface

Web Player

Customer-facing activation surface

Advertiser Reporting

Customer-facing activation surface

↺

Learn: Decision logs and outcomes stream back into model training, policy updates, and the Redis context layer.

Decision latency

14.9 ms

Yield lift

+18%

Completion rate lift

+11 pts

Stage 2: Trigger Event

Jordan Kim Household creates a decision moment

A streaming session hits an ad break and the platform must decide which ad to serve now. Balance relevance, campaign constraints, household context, and revenue yield inside the ad slot budget.

Live Trigger

Jordan Kim Household

CTV session | family profile | premium FAST viewer | sports content

MEDIA

Eventad_break_start

Customer surfaceCTV app

Decision objectiveYield + relevance

Decision windowBefore the surface finishes rendering

Context requiredHistory + live signals + policy

Business stakes+18%

Why This Moment Matters

This is not just ad selection. It is revenue-aware decisioning under strict latency and policy constraints.

Without Redis: the application waits on siloed repositories and defaults to a generic next step.

With Redis: the decision surface opens with ranked, contextual action already staged for the moment.

Stage 3: Ingest

Industry repositories and streams flow into Redis

These are common repositories and streaming APIs for media decisioning. They remain the systems of record. Redis activates the working set needed to decide now.

Redis Data Integration (RDI)Redis Feature Form

Source Systems → Redis

FREE

FreeWheel / GAM

Common repository or streaming source for freewheel context

→

SNOW

Snowflake Customer Preferences

Likes, dislikes, genre affinity, brand exclusions, household-level preference history

→

CDP

Snowplow / CDP

Session, identity, consent, and viewing-event context

→

KAFK

Kafka / Redis Streams

Event streaming via Kafka or Redis Streams — no separate broker required with Redis Streams built-in

→

CONT

Content metadata CMS

Common repository or streaming source for content context

→

ATTR

Attribution APIs

Common repository or streaming source for attribution context

→

MEAS

Measurement data warehouse

Common repository or streaming source for measurement context

→

Ingest Pipeline Status

CDC modeReal-time

Streaming lag<100 ms

Feature parityTrain = serve

Custom sync code0 lines

Serving roleOperational context layer

Stage 4: Context Assembly

History and live state converge into one working view

Redis assembles household profile history, Snowflake likes and dislikes, live viewing intent, campaign eligibility, and policy constraints in one low-latency lookup path. What looked like a simple ad-break event becomes a much richer decision moment here.

Redis RAMRedis FlexRedis Context Retriever

Historical Context

Customer value bandhigh

Tenure / relationship depthEstablished

Snowflake likesAuto, home improvement, premium sports

Snowflake dislikesTravel offers, repeated QSR ads

Prior interaction patternKnown baseline available

Eligibility stateResolved in memory

Policy constraintsCurrent and active

Action historyFrequency caps enforced

Live Context

Current intentYield + relevance

Streaming signal stateFresh for this session

Capacity / inventoryAvailable

Risk / complianceWithin active rules

Surface readinessCTV app

Decision scopePremium auto campaign with retail overlay

Key insight: Redis Context Retriever assembles the Campaign 360 — audience state, inventory context, and bid history — so the decision engine has exactly what it needs. The winning action becomes obvious only when historical context and live signals appear in the same response path.

Stage 5: Feature Serving

Online features hydrate the ranker in milliseconds

Redis Feature Form serves model-ready signals from Redis with train-serve parity. No fan-out calls to downstream systems during decisioning. No stale fallbacks.

Redis Feature FormRedis Feature Store

snowflake_preference_vector

Likes, dislikes, and declared brand preferences activated from Snowflake into Redis.

auto +0.92 / travel -0.710.3 ms

frequency_cap_status

Real-time feature served from Redis with train-serve parity.

clear0.4 ms

campaign_priority

Real-time feature served from Redis with train-serve parity.

tier 10.5 ms

content_adjacency_score

Model or ranking signal pulled online at decision time.

high0.6 ms

predicted_completion

Real-time feature served from Redis with train-serve parity.

0.870.7 ms

yield_per_impression

Real-time feature served from Redis with train-serve parity.

$0.0340.8 ms

Feature Serving Performance

Features Hydrated

186

P99 Lookup Latency

<20ms

Train / Serve Parity

100%

Stage 6: Ranking

Candidate actions are ranked for this moment

Rules, policies, vector similarity, and business weighting combine to rank the best action now. The winner is selected because it fits the moment, not because it is the easiest generic fallback.

Redis SearchNBA RankerRules Engine

#1 Winner

PRIMARY ACTION

Premium auto campaign with retail overlay

Highest revenue after cap, Snowflake likes/dislikes, affinity, and adjacency filters

NBA score0.94

#2 Candidate

ALTERNATIVE

Generic brand spot

Safe but lower yield and lower completion probability

NBA score0.79

Suppressed

POLICY

Competing travel ad

Suppressed by Snowflake dislike signal, brand exclusion, and frequency cap

Suppressed0.24

Stage 7: Business Impact

The value of choosing the right action now

Fast decisions matter, but the real value is choosing the right action while the moment is still open. Redis improves both latency and outcome quality.

Decision Economics

Decision latency14.9 ms

Yield lift+18%

Completion rate lift+11 pts

The key insight: value compounds when the decision surface uses live operational context instead of generic fallback logic.

Moment Outcome

Generic

Siloed systems
lower relevance

→

Redis

Context-aware
higher impact

Stage 8: Outcome

Same surface. Different decision layer.

On the left is a generic or delayed path. On the right is the Redis-powered experience with ranked action already staged. Same customer moment. Different outcome.

Generic Experience

Jordan Kim Household

Media workflow

What the system knows

Partial profile, delayed retrieval, and limited live context.

Action shown

Generic brand spot

Outcome risk

The action fits broadly, but not the full moment.

Lower

relevance

Higher

fallback use

Siloed

context

Redis-Powered Experience

Jordan Kim Household

Decision-ready profile

#1 Next Best Action

Premium auto campaign with retail overlay

Highest revenue after cap, affinity, and adjacency filters

Why it wins

Combines history, live signals, business value, and active policy constraints in the same low-latency response.

Visible impact

Yield lift: +18%.

High

relevance

Low

fallback use

Ready

before render

Stage 9: The Architecture, Proven

Architecture first. Architecture last. Outcome in the middle.

The same five tiers you saw at the start now tie directly to a measurable decision outcome. Common repositories and streaming APIs stay in place, including Snowflake as the preference repository for likes and dislikes. Redis Feature Form serves the online features, Redis RAM serves the hot ad-break state, and Redis Flex preserves the broader household, vector, and campaign context. Redis remains the operational context layer that activates Snowflake preference signals at ad-decision speed.

Data Sources

Ad Inventory Platform

Operational source

Campaign Management System

Operational source

CDP / Identity Graph

Household and device identity

Snowflake Preferences

Customer likes, dislikes, genres, and brand exclusions

Content Metadata Store

Content genre, ratings, and placement context

Measurement / Attribution APIs

Operational source

Kafka / Viewing Events

Operational source

Ingest Layer

RDI

CDC and activation from systems of record

Streams + Features

Live event hydration and model parity

Unified Context Layer

Redis RAM

Hot ad-break working set, live session state, active frequency caps, and real-time bid signals

Redis Flex

Warm household viewing history, preference vectors, campaign response patterns, and identity graph context

Feature Store

Brand fit, ad tolerance, conversion propensity, and revenue yield optimization features

Redis Context Retriever

Assembles the Campaign 360 — audience state, inventory context, and bid history — and exposes it as structured MCP tools for the decision engine

Decision Engine

Rules

Eligibility, compliance, and policy

Ranker

Propensity, value, and fit

Arbitration

One action selected now

Output Channels

CTV App

Customer-facing activation surface

Mobile Streaming App

Customer-facing activation surface

Web Player

Customer-facing activation surface

Advertiser Reporting

Customer-facing activation surface

↺

Learn: Outcomes feed retraining, policy updates, and future decisions without changing the systems of record.

Decision latency

14.9 ms

Yield lift

+18%

Completion rate lift

+11 pts