PLATFORM ENGINEERING · .NET
Feature Flags
Service
A self-hosted feature flag service and .NET SDK. I built it to evaluate flags quickly, cache definitions and send updates to connected clients.
INDEPENDENT PROOF OF CONCEPT Runs with self-hosted infrastructure; no public live API.
Release build · API host-pinned to two cores · 30s warm-up + 2m measurement
dotnet test · all projects
01 / WHY I BUILT IT
Why I built
the service.
Feature flag checks sit on request paths. I wanted to build a self-hosted .NET service that keeps flag definitions close to the API, evaluates targeting rules for each request and sends changes to connected clients.
It gave me a way to work through caching and invalidation, deterministic rollouts, a typed client SDK, automated tests and load testing in one project.
02 / ARCHITECTURE
How the pieces
fit together.
The API and application code hold the service logic; infrastructure handles PostgreSQL, Redis and caching.
Conceptual view of implemented components; it does not describe a live production deployment. Redis also carries the FusionCache invalidation backplane and SignalR scale-out. *SDK caching is optional.
03 / UNDER THE HOOD
A few decisions
behind the code.
Compute evaluations; cache definitions
The active v1.1 path reads a cached flag definition, then computes the user result on every request. Percentage rollouts use deterministic XxHash32 bucketing; span-based list parsing avoids splitting strings into temporary arrays. ADR 0002 records the move away from per-user result caching.
EvaluationService.cs·ADR 0002Cache the thing that changes less
CachedFeatureFlagRepository decorates the repository. FusionCache uses process memory as L1 and Redis as L2, with fail-safe and eager refresh options. Writes update cached values and remove related entries; its Redis backplane distributes invalidations between instances. PostgreSQL remains the persisted source of truth.
Cached repository·Cache registrationA typed client with opt-in ergonomics
The .NET SDK provides typed access to evaluation and management endpoints. Conditional ETag caching and the DI package retry handler are available for callers to configure; they are not implicit behavior of every client. The included Blazor sample demonstrates flag evaluation and SignalR updates locally.
Client implementation·Optional retry handlerMeasure the API and its dependencies
The API exposes Prometheus metrics; dashboards in Grafana include request and cache signals. k6 drives the documented local load test. Tests cover service and client behavior, while the benchmark is evidence for one machine and one workload—not a general capacity promise.
Observability setup·Evaluation endpoint04 / PERFORMANCE
Benchmarked on
my machine.
The v1.1 run was recorded on 20 February 2026.
Windows 11 · Ryzen 9 9950X3D · 64 GB RAM. API Release process pinned to two cores; PostgreSQL, Redis, Prometheus and Grafana in Docker; k6 ran on the host over localhost. The flag-definition cache was warm: 30-second warm-up, then two minutes of measurement (75 preallocated / 150 max virtual users). This is a single-machine local result, and does not establish behavior under internet latency, multi-node deployment or other workloads.
05 / SCOPE & NEXT STEPS
A useful PoC,
with room to grow.
The repository explicitly describes this as a proof of concept, not a production-ready service. Its README lists rate limiting as unimplemented. The published 50k screenshots predate the v1.1 computed-evaluation run, and the distributed benchmark documentation describes a test procedure rather than published distributed results.
Infrastructure documentation calls out dedicated cluster testing as future work; the local benchmark notes that dedicated infrastructure would give more representative results. I would extend validation across nodes and failure scenarios before making production capacity claims.
Core service and SDK paths have automated tests, but the current project should still be treated as an engineering demonstration. The metrics here do not imply external users, availability commitments or a hosted service.
06 / GO DEEPER