Case Studies

30 adtech architectures, annotated.

One practitioner's read on the real decisions — technology choices, trade-offs, and what I'd do differently. Filter by topic below.

30 case studies

Klaviyo ClickHouse

Klaviyo — Real-Time Segmentation

Klaviyo rebuilt real-time segmentation on a 192-node ClickHouse cluster using bi-level sharding, handling tens of billions of segment membership changes per day.

Streaming OLAP Why Not Snowflake
Priority
Braze ClickHouse

Braze — Ads Analytics Pipeline Rebuild

Braze replaced a dual MongoDB + Snowflake pipeline with ClickHouse Cloud, achieving 8x performance improvement while handling 1B+ events per hour.

Streaming OLAP Why Not Snowflake
LifeStreet ClickHouse

LifeStreet — Programmatic RTB at 5PB

LifeStreet, one of the first US adtech companies on ClickHouse (2016), runs 5PB of RTB data, campaign budgets, and raw logs on a single geo-distributed cluster.

Streaming OLAP Why Not Snowflake
Priority
Reddit Apache Druid

Reddit — Ad Budget Pacing in 30ms

Reddit solved sub-30ms ad budget pacing with Apache Druid, achieving 100% accuracy even during Kafka and Flink outages — replacing failed batch and streaming-to-batch approaches.

Streaming OLAP Why Not Snowflake Kafka & Flink
LiquidM Apache Druid

LiquidM — DSP with 50+ Targeting Dimensions

LiquidM runs a DSP handling millions of requests/second, filtering on 50+ targeting dimensions simultaneously within 100ms using Apache Druid.

Streaming OLAP Why Not Snowflake
Priority
Moloco Apache Pinot

Moloco — DSP: 10 Minutes → Milliseconds

Moloco cut query latency from 10 minutes to milliseconds by migrating its $1B+ ad platform to Apache Pinot + StarTree, now serving 6M queries/second.

Streaming OLAP Why Not Snowflake
UberEats Apache Pinot

UberEats — Merchant-Facing Ad Analytics

UberEats uses Apache Pinot's upsert capability to serve accurate real-time ad analytics to merchants, updating impressions, clicks, and conversions from a Kafka changelog.

Streaming OLAP
Sovrn Apache Pinot

Sovrn — Publisher Ad Reporting

Sovrn went from 24–48 hour reporting lag to second-level real-time data for publisher ad reporting using StarTree managed Apache Pinot.

Streaming OLAP
Netflix Apache Druid

Netflix — In-House Ads Event Processing Pipeline

Netflix rebuilt its entire ad infrastructure in-house in January 2024, using Kafka, Flink, and Apache Druid with a sessionization pattern to collapse raw events into structured Ad Sessions.

Streaming OLAP Kafka & Flink
Multiple (Simon Data, Tealium, Amplitude) Snowflake

Snowflake + First-Party Data Stack (CDPs + The Trade Desk)

Seven clients are live with a Simon Data/Tealium/Amplitude → Snowflake → The Trade Desk architecture, using Snowpipe for continuous ingestion with no direct API builds between CDPs and DSPs.

Composable CDP Clean Rooms
LiveRamp Snowflake

LiveRamp Clean Room — Architecture at Terabyte Scale

LiveRamp's clean room architecture runs at terabyte scale on Kubernetes with dynamic Spark workloads, using differential privacy to enable cross-publisher data matching without raw list sharing.

Clean Rooms Composable CDP
Deep Sync Databricks

Deep Sync — Identity Spine for 98% of U.S. Households

Deep Sync built an identity spine covering 98%+ of U.S. households on Databricks Lakehouse, providing the data foundation for large-scale audience modeling and lookalike targeting.

Composable CDP Clean Rooms Databricks vs Snowflake
Priority
Walmart Retail Media

Walmart — Scintilla + Vizio = CTV-to-Store Attribution Loop

Walmart's Scintilla platform + Vizio acquisition ($2.3B) created the industry's most complete cross-channel attribution loop, linking mobile, in-store Scan & Go, and CTV data for $4.4B in ad revenue.

Retail Media CTV Measurement
Industry Retail Media

The RMN Fragmentation Problem

With 200+ RMN networks using incompatible attribution windows, measurement fragmentation is the defining problem of retail media — creating a clear opportunity for warehouse-native normalization.

Retail Media MMM & Attribution
Innovid CTV Measurement

Innovid — Census-Level CTV Measurement at 95M Homes

Innovid built census-level CTV measurement across 95M+ U.S. homes using a privacy-compliant identity resolution infrastructure, before being acquired by Mediaocean in November 2024.

CTV Measurement
Industry MMM

The MMM Renaissance Narrative

Cookie deprecation killed deterministic MTA. Meta's Robyn (2021) and Google's Meridian (2024) catalyzed a return to MMM — now open-source, continuous, and the only viable cross-channel measurement in a cookieless world.

MMM & Attribution
Industry Snowflake

The Composable CDP Architecture (Definitive Stack)

The composable CDP stack (ELT → Warehouse → dbt → Reverse ETL → Destinations) is now the standard for data warehouse-native activation, but its warehouse round-trip makes it unsuitable for in-session personalization.

Composable CDP
Priority
Industry MMM

The MMM + Incrementality Flywheel

The MMM + Incrementality Flywheel — geo holdout → iROAS → Bayesian priors → MMM calibration — converts incrementality from a one-time audit into a continuously improving measurement system built on the warehouse.

MMM & Attribution