Scaling learning workflows across institutions without rebuilding the core.
TopGrade is an education platform that evolved from independent tutoring to serving thousands of concurrent students across multiple universities and school networks.
The Context
TopGrade built an intuitive interface for lesson scheduling, coursework distribution, and digital evaluations. In the early days, single-server infrastructure comfortably supported thousands of daily interactions.
However, as university partnerships expanded, usage patterns shifted from steady browsing to extreme synchronized bursts: 15,000 students submitting time-capped exams within a 90-second window. The shared database connection pool seized, background grading jobs backed up, and teachers could not access student records.
The Engineering Problem
Noisy-neighbor locks during synchronized exam spikes.
When Institution A ran an engineering exam with 3,000 concurrent submissions, the database write throughput spiked. The relational database locked rows across shared question bank tables, causing queries for Institution B (who were simply trying to view attendance) to time out after 30 seconds.
The instinct was to upgrade to massive AWS database instances. But vertical scaling is an expensive bandage that fails when concurrency triples again during final exam week. The data architecture itself had to be isolated.
What Tenlyr Did
We redesigned the data access paths and decoupled real-time student ingestion from heavy analytical evaluation:
Schema-Level Institutional Partitioning
Separated data stores by institutional tenants so heavy write activity in one school never exhausts connection pools or triggers table locks for another.
Asynchronous Submission Ingestion
Transformed exam submission from a blocking database transaction into an atomic, durable queue ingestion. Submissions acknowledge in 32ms without waiting for evaluation calculations.
Syllabus & Question Caching
Placed static course syllabus trees and question banks behind a Redis edge cache with automated invalidation, offloading 75% of read queries from primary PostgreSQL instances.
Real-Time Concurrency Telemetry
Added custom Prometheus metrics on queue backlog latency and database pool saturation, giving institutional IT teams live visibility during high-stakes testing windows.
The Result
Predictable sub-50ms latency across 50,000+ student exam sessions.
TopGrade successfully onboarded multiple university networks with zero platform rebuilds. The system comfortably handles synchronized burst traffic without noisy-neighbor contention or exam submission data loss.
What We Learned: Educational traffic is non-linear.
Standard web products experience bell-curve traffic. EdTech experiences sudden vertical cliffs: 200 users at 9:59 AM, 18,000 users at 10:00 AM. Architecture that relies on slow autoscaling triggers will fail students every time. Decoupling ingress from compute and partitioning tenant data paths is the only sustainable way to survive exam spikes.
Scaling a product with bursty or unpredictable load?
Tell us what's happening. We'll identify where queries and queues are choking.