Cognitiv — ML Feature Store for Real-Time Bidding
Cognitiv uses ClickHouse as an ML feature store for real-time bidding, computing and serving features within the 100ms bid window constraint.
Scale
Massive datasets for ML feature computation within <100ms bid windows
Before
—
After
ClickHouse as ML feature store for programmatic advertising — features computed and served at bid latency
Key Insight
Feature stores for RTB are latency-critical — features must be computed and served within the bid window (<100ms). This shapes every architectural decision.
In a Snowflake Conversation
Feature stores for RTB are latency-critical — features must be computed and served within the bid window (<100ms). ClickHouse's read performance at this latency profile is what makes it viable.
My Read
Practitioner commentary coming soon.
Relevant Conversations