Pulse
Calculates live personalization features inside a 10ms deadline, sitting directly in the critical path of every public page render.
Engineering highlights
The core architectural breakthroughs.
High-scale engineering demand distilled into four verified production milestones.
Zero-delay critical path injection into Next.js SSR page renders
Real-time personalization across 4 worldwide cloud regions
Active-Active Redis Enterprise clusters eliminate cross-region drift
Sub-second user action feedback loop boosts subscriber retention
Executive overview
The engineering challenge & solution.
Serving real-time recommendations to 45 million subscribers. To keep recommendations compelling, user behavior (clicks, skips, search terms) had to influence the next screen render within seconds across North America, Europe, and Asia. Legacy feature pipelines took 45ms to hydrate user profiles, causing unacceptable web page render lag.
We built an ultra-low-latency distributed feature platform in Go utilizing Redis Enterprise multi-region active-active replication, Kafka event streams, and PgBouncer connection multiplexing. Feature definitions are declared as code in a central registry that compiles to both real-time Go structs and Snowflake SQL transforms.
Engineering governance
The three non-negotiables.
These are the architectural constraints that shaped every boundary — no trade-offs or compromises permitted.
7ms Hard p99 Read Deadline
Feature extraction must return in 7ms p99 to avoid delaying the Next.js server-side page render.
Strict Online/Offline Feature Parity
Features used by offline training models must match the online production feature store with 100% mathematical parity.
Multi-Region Active-Active Replication
Features synchronized across 4 worldwide cloud regions without global locking overhead.
Eliminating Online-Offline Feature Drift Across 4 Global Cloud Regions
Data scientists trained recommendation models using historical data warehouses, but differences in time-zone aggregation and data types between Snowflake and production Redis caused model predictions to degrade by 22% in production.
We built a unified feature compiler in Go and Python. Feature computation logic is defined once in PyArrow/Polars and compiled directly into WASM modules for the Go online serving engine and SQL transformations for the offline data lake. A continuous shadow validator queries both engines and flags any divergence greater than 0.001%.
Request lifecycle
How data moves through Pulse.
Production verification
Measured outcomes in production.
Every metric below was captured under real production traffic and audited against historical baseline data.
Operational resilience
Automated safety guardrails.
Systems fail in production. Here is how Pulse survives network partitions, upstream timeouts, and anomalous inputs without human intervention.
Local Memory Cache Fallback
If Redis cluster connectivity lags, daemon serves verified local L1 memory cache.
Automated Drift Monitor
Alerts telemetry if online feature distributions diverge from training baselines.
Graceful Feature Degradation
Drops low-priority auxiliary features if upstream network budget exceeds 5ms.
Next step
Building a system with this level of demand?
Bring us the constraint that keeps your engineering leadership up at night — the latency ceiling, the compliance perimeter, or the unyielding reliability requirement.