top of page

How Intelligent Infrastructure Turns Adtech's Peak Moments into Revenue

4 hours ago
4 min read

By Litan Shamir, Director of Infrastructure, Rise


Black Friday doesn't start on Black Friday.


At least, not for the infrastructure behind programmatic advertising.


Long before shopping activity peaks, advertising platforms must be ready to process surges in demand and make billions of decisions in milliseconds. Every request needs to be checked, enriched, priced, and routed to buyers before an auction's deadline expires.


At Rise, infrastructure isn't simply a cost center supporting the product. It is part of the product itself. It determines how quickly we respond to opportunities, how intelligently we allocate resources, and how efficiently we turn demand into revenue.


We've built our infrastructure around one principle: it should think, not just scale. Peak events like Black Friday put that principle to the test, but the engineering behind it happens throughout the year.


Built for the Burst


Every day at midnight US Eastern, advertising budgets reset, triggering a surge in demand. During Black Friday and other major shopping events, these fluctuations become even more pronounced.


Rise operates on a fully cloud-native, multi-region infrastructure designed to respond to these changes in real time. Services along the auction path scale horizontally based on live traffic signals, with progressive rollouts and disruption budgets built into their deployment. Computing capacity expands within seconds when demand rises and contracts when it subsides.


But handling the surge requires more than elasticity.


Before reaching the core auction, every request passes through a high-performance filtration layer that performs more than a dozen quality and eligibility checks, alongside partner-specific traffic shaping, in microseconds. This prevents expensive auction resources from being consumed by traffic that cannot generate value.


Inside the auction, strict per-request deadlines determine the time available for each buyer, while model execution is limited to tens of milliseconds. Circuit breakers isolate slow partners to protect overall performance.


During daily demand surges, this architecture is designed to keep timeouts stable, throttling near zero, and buyers responding at full rate.


The objective isn't simply to process more requests. It's to preserve auction performance and efficiency as demand increases.


Scale Without Proportional Cost


Cloud elasticity is easy to buy, but much harder to use efficiently.

A large majority of Rise's auction fleet operates on discounted spare cloud capacity, helping us absorb growth without a proportional increase in infrastructure spending.


Workloads are scheduled according to their computational requirements, using processor architectures that offer the best balance of performance and cost. Rather than manually optimizing individual servers, we define policies that guide these allocation decisions.


Our proprietary FinOps Hub connects cloud spending with business activity, measuring cost per million advertising opportunities, cost per dollar of revenue, and cost by region.


This allows us to evaluate infrastructure efficiency against the business value it supports, not simply the computing resources it consumes.


Data Architecture Built for Milliseconds

An auction is only as intelligent as the information available when it runs. But identity resolution, real-time auction signals, and historical analytics have fundamentally different performance requirements.


That's why Rise uses specialized data engines, each optimized for its workload.


Identity at scale. Our identity graph runs on a massively scalable wide-column database, handling well over 100,000 reads and writes per second with single-digit millisecond tail latency. We migrated this workload from a legacy in-memory database without downtime, using fallback reads and automatic data migration to maintain continuity.


High-speed, in-memory processing. Time-sensitive user signals and win tracking operate on distributed in-memory key-value clusters spanning hundreds of shards, supporting millions of operations per second with sub-millisecond latency.


Purpose-built automation. Our proprietary infrastructure operators automate cluster provisioning, scaling, failover, and persistence, while managing experimentation configurations as versioned platform resources.

The principle is to use the right storage engine for each workload: wide-column databases for large-scale keyed access, in-memory systems for ultra-low-latency operations, and a data warehouse for analytics.


This separation allows us to optimize performance and cost without forcing every workload into the same technical constraints.


Intelligence Inside the Auction


Infrastructure becomes a competitive advantage when it supports better decisions within the auction's strict time limits.


At Rise, machine learning operates directly within the auction path. Our pricing models run on a GPU-grade inference layer capable of serving tens of thousands of predictions per second across regions, with pre-bid and post-bid models operating within tightly controlled execution windows.


Price floors are continuously optimized using multi-armed-bandit models that learn from fresh auction outcomes across dozens of supply segments, allowing pricing decisions to adapt to changing market conditions throughout the day.

We also use large language models to analyze and classify page content, enriching inventory with more specific contextual categories that help buyers identify relevant advertising environments.


But not every AI workload belongs inside the auction.


Smaller, high-performance models handle real-time decisions, while larger language models support contextual enrichment outside the auction's critical execution path.


The challenge isn't simply adding AI. It's deploying the right intelligence where it creates value without compromising speed or economics.


The Infrastructure Behind Better Outcomes


Peak shopping events expose the difference between infrastructure built to handle average traffic and infrastructure engineered for extreme conditions.

As programmatic advertising becomes more data-intensive and automated, the challenge is balancing performance, intelligence, and cost efficiency at scale.

For publishers, that means evaluating more revenue opportunities efficiently. For buyers, it means faster, more informed auction decisions. And for Rise, it means supporting growth without allowing operational complexity and costs to grow at the same rate.


That's what AI-native infrastructure means to us: intelligent decisions at every layer, from computing resources and data architecture to auction optimization.


Because when the biggest moments arrive, the infrastructure behind them should already be ready.


By Litan Shamir, Director of Infrastructure, Rise


bottom of page