Multi-Tenant SaaS Platform

Overview

The multi-tenant SaaS platform I lead at Mediana: a YARP API gateway fronts a fleet of .NET 10 microservices — every request is authenticated against the identity service first, then routed to the owning service. Service-to-service calls ride gRPC, each service keeps its own Redis cache, and Apache Kafka absorbs the high-throughput ingress and chains the downstream processing stages — hot endpoints hold a p99 of 60ms at roughly 2,000 requests per second.

The platform runs its own wallet service for balances and transactions. Everything deploys to Kubernetes with high availability; SQL Server serves the operational data while ClickHouse archives cold history at column-store compression to cut storage costs. OpenTelemetry traces every processing step into SigNoz for latency analysis, and errors ship to Splunk.

Private company repositories — the diagram shows the production topology, not a public file inventory.

• Mediana (private)company GitLab

Tech stack

Design patterns

• Auth-first gateway

  Every request authenticates against the identity service before YARP routes it — services trust the gateway boundary, never the raw caller.

• gRPC inside, HTTP outside

  Public traffic speaks HTTP at the gateway; between services it is typed gRPC contracts for low-latency internal calls.

• Kafka-buffered processing

  High-throughput ingress lands on Kafka and fans out through processing stages, so bursts never back-pressure the request path — that is what holds p99 at 60ms.

• Cache per service

  Each microservice owns its Redis cache — hot reads stay local and no service couples to another through shared cache state.

• Hot/cold storage split

  Operational data lives in SQL Server; history archives into ClickHouse, whose columnar compression is a large part of the storage-cost reduction.

• Traced end to end

  OpenTelemetry spans every hop into SigNoz — per-stage timings and latency come from traces, not guesses — while error streams feed Splunk.

System design