Summary of "Top 5 techniques for building the worst microservice system ever - William Brander - NDC London 2023"

Top 5 techniques for building the worst microservice system ever

(Summary of William Brander — NDC London 2023)

A humorous, experience-driven walkthrough of common microservice anti-patterns and how they degrade reliability, performance and maintainability. Examples are .NET-focused but the patterns apply broadly.

High-level theme

This talk uses real-world support experience (Particular Software / NServiceBus) to highlight microservice anti-patterns. The goal is to show how well-intentioned architectural choices can introduce latency, operational complexity, and brittle coupling.

Core concepts covered

Strangler‑fig migration: extracting functionality from a monolith can preserve original coupling while adding network latency and ops complexity.
.NET garbage collection and performance: Gen 0/1/2 GC behavior — Gen‑2 promotions (historically stop‑the‑world) can cause big latency spikes; long‑lived threads (e.g., waiting on network) increase Gen‑2 promotions.
Throughput vs distribution: adding network hops typically reduces throughput and raises contention compared to a monolith.
Big‑bang rewrite risks: loses incremental delivery, underestimates work, and creates long‑lived integration layers.
Messaging queues and reliability pitfalls: naive use introduces poison messages, retry storms, dead‑letter queues, and the need for throttling/backoff.
Not‑Invented‑Here (NIH): building homegrown frameworks wastes effort and increases maintenance compared to battle‑tested libraries.
Service boundary design: modeling services by nouns (product, order, customer) often creates chatty cross‑service calls when users invoke verbs that span domains.
Read models / CQRS & eventing: projections and pub/sub reduce synchronous coupling but don’t eliminate logical coupling — business rule changes still need propagation.
Engine pattern: a central runtime (engine) exposes interfaces; services implement interfaces and drop compiled plugins into the engine. This keeps logical ownership while hosting code elsewhere — useful selectively but introduces coupling and deployment tradeoffs.
Logical vs physical boundaries: moving code physically doesn’t change logical ownership; conflating them causes subtle coupling.

Five anti‑pattern techniques (and why they hurt)

Put HTTP/network calls in front of everything
- Adds network latency, increases thread lifetimes and GC promotions, lowers throughput, and creates topology coupling.
Attempt a big‑bang rewrite
- Stops incremental delivery, underestimates effort, and results in long‑living migration layers. Works only for tiny systems or personal CV projects.
Replace sync calls with naive queuing and reinvent reliability
- Poison messages can block queues. Naive retries cause backoff storms or DDOS‑like behavior on downstream systems. Teams often end up building custom frameworks instead of using proven tooling.
Build everything yourself (Not‑Invented‑Here)
- Reimplementing messaging, retry, UI, etc., for the joy of engineering stalls business delivery and increases maintenance burden.
Define services around nouns (objects) rather than verbs (behaviors)
- Leads to chatty, cross‑service requests, API gateway/orchestration sprawl, duplicated read models, and hidden coupling.

Practical notes and recommended alternatives

Watch for Gen‑2 GC pauses in .NET; distributing work can aggravate GC effects.
Use established libraries for messaging and resiliency rather than reimplementing:
- NServiceBus (speaker’s employer)
- Rebus, MassTransit (alternatives)
- Polly (transient‑fault handling / retries)
Use CQRS/read models and eventing to improve read‑side performance, but remember they do not remove logical coupling — you must still synchronize business rule changes.
Consider the engine pattern selectively to avoid duplicating projection logic across services: keep implementations in the logical owner’s codebase but execute them within a shared runtime. Evaluate coupling and deployment complexity before adopting.

Resources and references

Particular Software / NServiceBus (practical support experience behind the talk)
Libraries/tools mentioned: NServiceBus, Rebus, MassTransit, Polly (for retry/transient‑fault handling)
YouTube: a microservices explainer by “crisam” — recommended for describing the chatty‑services problem
Historical/illustrative reference: a Stack Overflow incident / blog post about GC pauses (used to illustrate Gen‑2 impact)