Real-Time Pricing Engine Case Study

The challenge

A large-scale ordering system needed a high-performance real-time pricing and promotion engine capable of processing 100K+ transactions daily. The existing architecture had been built for a much lower transaction volume and was showing multiple failure modes under current load.

⚠️ Pricing latency exceeding 50ms. The existing system was performing synchronous database lookups for every pricing rule evaluation - resulting in price and promotion calculation times consistently over 50ms. At 100K+ daily transactions during peak promotional periods, this latency directly impacted cart abandonment rates and checkout conversion.

⚠️ Inefficient cross-service communication. The pricing service was a synchronous dependency for the order service - every price check blocked an API thread. During high concurrency, the order service thread pool would back up waiting for pricing responses, causing cascading latency across unrelated order operations.

⚠️ No AI-driven pricing intelligence. The pricing engine was purely rule-based with no dynamic optimisation. Promotion personalisation, demand-responsive pricing, and customer-segment-specific offers required manual configuration changes - and the business team couldn't iterate fast enough to take advantage of real-time market signals.

⚠️ Limited scalability and reliability. The monolithic pricing service had a single point of failure. During high-traffic promotional events, pricing service degradation had caused order failures - a direct revenue impact. The architecture needed a fundamental redesign to handle 10× burst traffic during sales events.

Performance architecture

Redesigned Pricing Pipeline - After OKRUTI

<50ms

Average pricing latency

100K+

Transactions per day

99.9%

Uptime across peak load

Order Request → Kafka Topic → AI Pricing Engine → GridGain Cache Lookup → Promotion Validator → Price Published (<50ms)

Latency improvement

Before - synchronous DB lookup per rule >50ms avg

After - GridGain in-memory + Kafka async pipeline <50ms achieved

Peak burst - promotional events (10× normal traffic) Maintained SLA

Our solution

OKRUTI redesigned the entire pricing pipeline from synchronous and database-dependent to event-driven and in-memory - with AI pricing intelligence layered on top. The new architecture is fundamentally different in its approach to data access: instead of reading pricing rules from a database on every request, rules live in GridGain's in-memory distributed grid and are evaluated in microseconds.

→ Apache Kafka Event-Driven Pipeline: Redesigned the pricing service as an event consumer. Order events are published to a Kafka topic; the pricing engine consumes them, evaluates rules, and publishes the priced order - decoupling the pricing calculation from the order service entirely. The order service no longer blocks on a pricing API call; it reacts to a pricing event.

→ GridGain In-Memory Computing: All active pricing rules, promotion definitions, customer segment mappings, and product pricing matrices are loaded into GridGain's distributed in-memory grid. Rule evaluation - which previously required database round trips - now runs against in-memory data structures with microsecond access latency, eliminating the core bottleneck.

→ AI Pricing Intelligence: Integrated ML models for dynamic price optimisation, customer-segment personalisation, and promotion relevance scoring. The AI layer evaluates each transaction's context - customer history, cart composition, real-time demand signals - and adjusts promotion presentation and pricing recommendations accordingly, without any rule configuration change required.

→ Microservices Decomposition: Split the monolithic pricing service into four independent services - Price Calculator, Promotion Engine, AI Recommendation Layer, and Price Publisher - each horizontally scalable. During promotional events, only the Promotion Engine and AI Recommendation Layer need additional instances; the Price Calculator scales independently.

→ Elimination of N+1 Query Patterns: Profiled all data access in the original pricing code and identified 12 distinct N+1 query patterns - loops that generated a database query for each iteration. Rewrote all of these using batch-fetch patterns with in-memory lookup tables, eliminating hundreds of unnecessary database round trips per pricing calculation.

Results achieved

<50ms

Average pricing latency - down from over 50ms to well below the real-time threshold

100K+

Daily transaction capacity - handles peak promotional events without performance degradation

99.9%

Platform uptime - sustained across high-traffic promotional events and seasonal peaks

Dynamic personalisation - AI pricing intelligence live on every transaction, driving conversion lift

High-throughput pricing or promotion engine?

Whether you're rebuilding a legacy pricing system or engineering a new real-time engine from scratch, our performance engineering team can design the right architecture for your transaction volumes and latency targets.

Book Free Architecture Review → All fintech solutions

Real-time Pricing & Promotion
Engine at 100K+ Transactions/Day

High-throughput pricing or promotion engine?

Let's talk about your platform.

Real-time Pricing & PromotionEngine at 100K+ Transactions/Day

High-throughput pricing or promotion engine?

Related case studies

Let's talk about your platform.

Real-time Pricing & Promotion
Engine at 100K+ Transactions/Day