Practical .NET and Software Architecture Tips | Microsoft MVP

Hi there! 👋

I'm a senior software engineer, primarily working in the .NET ecosystem.

Things I talk about on LinkedIn:
- C# and .NET ❤
- Software engineering and architecture 
- Distributed systems and how to build them
- Databases and optimization techniques
- Career and personal growth

One of my goals (i̶s̶) was to become a Microsoft MVP in Developer Technologies.
I became an MVP in March of 2023.

If this sounds interesting, let's connect!

Milan Jovanović

You don’t need a Dockerfile anymore.

Yes, really.

Modern .NET can publish your app directly as a container image using the SDK. No scaffolding. No boilerplate. No hand-written Dockerfiles.

Here's what you'll learn:
- How dotnet publish builds a Linux container
- How to run it locally with Docker
- How to push it to GitHub Container Registry
- Why this makes self-hosting .NET apps on a VPS much simpler

If you’ve ever copied a Dockerfile you didn’t fully understand, this is for you.

Check it out: https://lnkd.in/dUFzz3du

This is also the first step in a new series I’m working on around self-hosting .NET apps without overengineering.

undefined undefined

AI in .NET is way easier than people think.

And you can start for free.

If you want a practical first project, build vector search:
- Turn text into embeddings (vectors)
- Store them in your vector store (pg_vector)
- Turn the user’s query into a vector
- Compare with cosine similarity to find the closest matches

That’s why it feels like “search by meaning” instead of keywords.

To do this cleanly in .NET, start with MEAI (Microsoft Extensions AI), the official library for LLM + embeddings.

Start here: https://lnkd.in/dkpadmVc

---

Do you want to simplify your development process? Grab my Clean Architecture template here and save 7 days of development time: https://lnkd.in/dfGYcyFC

I think this makes me an Azure core developer now

/s

Not all Load Balancing strategies are created equal.

If you are using YARP (Yet Another Reverse Proxy) in .NET, you have 5 built-in ways to distribute traffic. But which one should you choose?

Here is the breakdown:

1. RoundRobin: The classic. It cycles through servers evenly. Simple, but it assumes all servers are equally fast.

2. LeastRequests: The smart choice. It sends traffic to the server currently doing the least work.

3. Random: It picks a server at random. Surprisingly effective for high-volume, uniform workloads.

4. PowerOfTwoChoices: The balanced approach. It picks two random servers and chooses the one with fewer requests. It avoids the overhead of checking every server.

5. FirstAlphabetical: The "sticky" one. It always picks the first available server. Good for failover setups, bad for load distribution.

Most people stick with RoundRobin, but switching to LeastRequests can significantly improve tail latency in systems where request times vary.

I dive deeper into how to scale horizontally using YARP here: https://lnkd.in/daDNMpmB

Which strategy is your default choice?

---
Sign up for the .NET Weekly with 75K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dcvJmi7E

Stop calling DateTime.UtcNow in business logic.

This line looks harmless... and it’s the reason your tests are a mess.

Here’s the trap:

- DateTime.UtcNow > user.SubscriptionEnd

Seems harmless, right?

But the moment you try to test this logic, you realize the problem:

- You can’t control DateTime.UtcNow
- Your tests become flaky or complicated
- Business rules are tied to infrastructure

There are better ways:

1. Inject DateTime as a parameter — simplest, great for small methods

2. Introduce an IClock interface — more scalable, and test-friendly

By inverting the dependency on time, you make your domain logic:

✅ Easier to test
✅ More deterministic
✅ Separation of concerns

Next time you write a time-based condition, stop and ask:

“Will I need to test this?”

If the answer is yes, extract the clock.

---
Do you want to simplify your development process? Grab my Clean Architecture template here and save 7 days of development time: https://lnkd.in/dfGYcyFC

How did we make a simple idea so complicated?

CQRS is the simplest pattern out there.

First, there was CQS: Command Query Separation.

What does this mean? We split an object's methods into write (command) and read (query) methods.

CQRS is the evolution of CQS. CQRS works on the architectural level. At the same time, CQS works on the method (or class) level.

The CQRS pattern uses separate models for reading and updating data.

This separation can be logical OR physical.

My preferred approach is starting with logical CQRS.

One database + separate flows for writes and reads.

You can design a data model fine-tuned for each operation.

A lot of developers stay away from CQRS because they think it's complex.

The reality is you're probably already applying CQRS principles without knowing it.

Here's a simple implementation: https://lnkd.in/d-qUdpn3

What do you think about the CQRS pattern?

---
Sign up for the .NET Weekly with 75K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dnecq4zh

Should you stop using random UUIDs in .NET?

Traditional UUIDs (Guid.NewGuid()) are randomized. Great for uniqueness, but terrible for database performance.

They cause:

❌ Expensive index rebalancing
❌ Poor insert performance
❌ High fragmentation

That’s why many developers switched to ULIDs. It's 128-bit, sortable, and still unique.

They preserve temporal order, which makes them more index-friendly.

But now, there’s something new:

✅ Guid.CreateVersion7() in .NET 9+

UUID V7 is time-ordered and compatible with ULIDs, without needing a 3rd-party package.

You get the benefits of ULID, but natively supported in .NET.

Would you switch from ULID to V7 UUIDs?

---

Do you want to simplify your development process? Grab my Clean Architecture template here and save 7 days of development time: https://lnkd.in/dfGYcyFC

Here are 9 excellent libraries I use in my projects:

1. YARP
2. Polly
3. EF Core
4. Refit
5. Testcontainers
6. OpenTelemetry
7. FluentValidation
8. xUnit
9. Dapper

Which library would you add to the list?

---
Sign up for the .NET Weekly with 76K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dbU3m97Q

Stop trying to force Domain-Driven Design into every feature.

Sometimes, the best architecture is the simplest one: The Transaction Script.

It’s a classic pattern from Martin Fowler, and it fits 80% of CRUD scenarios perfectly.

Instead of building complex aggregates and value objects, you just write a simple, procedural script that handles a single request:

1. 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗲 the input data.
2. 𝗖𝗮𝗹𝗰𝘂𝗹𝗮𝘁𝗲 the result (business logic).
3. 𝗣𝗲𝗿𝘀𝗶𝘀𝘁 the changes to the database.
4. 𝗥𝗲𝘁𝘂𝗿𝗻 the response.

It is fast, readable, and easy to test.

I wrote an article showing a side-by-side comparison of this vs. a Rich Domain Model.

See the code here: https://lnkd.in/d9dwgpdq

👉 When to switch: Start with a Transaction Script. Only refactor to a Domain Model when the business logic becomes "chatty", meaning you have too many if/else blocks or complicated state transitions that are hard to track in a single method.

---
Sign up for the .NET Weekly with 76K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dbU3m97Q

The best HttpClient config is the one you don’t repeat in every service.

In .NET, you set it up once with IHttpClientFactory (and Aspire if you’re using it):
- Discovery for service-to-service calls
- Resilience defaults so transient failures don’t hurt you

Then you just inject the client and write business code.

Want the full walkthrough?

Start here: https://lnkd.in/dQNxjHix

---

Do you want to simplify your development process? Grab my Clean Architecture template here and save 7 days of development time: https://lnkd.in/dfGYcyFC

In .NET, you usually have to choose between Speed and Consistency.

In-Memory Caching (IMemoryCache):
✅ Blazing fast (nanoseconds).
❌ Data is stuck on one server. If you have 10 nodes, you have 10 caches.

Distributed Caching (IDistributedCache / Redis):
✅ Shared across all servers. Consistent data.
❌ Network latency. Serialization costs. It's slower.

The Solution: HybridCache.

Don't choose one. Use both.

A Hybrid Cache creates a tiered architecture:

- L1 (Local): The app checks RAM first. If it's there, return instantly.
- L2 (Distributed): If missing locally, check Redis.
- Source: If missing in Redis, fetch from DB, then populate both L1 and L2.

While Microsoft is working on a native implementation, I still swear by FusionCache as the most robust, battle-tested library for this.

I wrote a guide on how to set this up properly: https://lnkd.in/d6NKgMVp

Are you currently using just Redis, or a multi-layer cache?

---
Sign up for the .NET Weekly with 75K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dCMpeCNe

Most OAuth demos are useless in the real world.

They skip the part that matters: how public clients should actually get tokens.

I just published a step-by-step video on running Keycloak in Docker + OAuth 2.0 Authorization Code flow with PKCE.

What you’ll see:
- Keycloak (latest) running locally with Docker/Compose
- Realm + user + public client setup
- Using .well-known/openid-configuration so you don’t guess endpoints
- Full PKCE flow in Postman (authorize → code → token exchange)
- Quick JWT inspection + the “why is this 401/403?” scope moment

If you’re building SPAs or mobile apps, PKCE isn’t optional. It’s the safe default.

Watch it here: https://lnkd.in/deZWiEbi

The worst kind of bug is the one that explodes 3 days after deployment.

This is the hidden danger of the standard Options Pattern in .NET.

You might bind your appsettings.json to a class, inject it, and everything looks fine. The app starts up perfectly.

But if a required API Key is missing? You won't know until a user triggers that specific feature, and the app crashes.

This is why "Fail Fast" is crucial for configuration.

If your config is invalid, your application should refuse to start. Period.

You can achieve this by extending the Options Pattern with IValidateOptions.

Instead of just binding data, you attach validation logic:

1. Define rules (e.g., "ApiKey cannot be null", "RetryCount must be > 0").
2. Register the validator.
3. If the config breaks the rules, the DI container throws an exception immediately at startup.

I took this a step further and integrated FluentValidation to make the rules cleaner and more expressive.

Here is the complete implementation code: https://lnkd.in/djz5Yfux

---
Sign up for the .NET Weekly with 75K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dcvJmi7E

Hidden danger of Options Pattern in .NET

I failed a .NET interview question that looked easy.

“Add caching to an async API.”

The tricky part isn’t caching.
It’s what happens when 50 requests hit the same cold cache at once:
- race conditions
- cache stampedes (thundering herd)
- why ConcurrentDictionary still isn’t enough
- why you can’t await inside a lock
- the SemaphoreSlim async-lock pattern (with try/finally, double-check, timeout)
- scaling it with per-key locks to reduce contention

I break it down step-by-step using a currency conversion example, and show the exact refactor path from “works locally” to “works under load.”

Watch it here: https://lnkd.in/dyN7_acX

If you need to run a task "Every Friday at 5 PM," how do you implement it?

A While(true) loop with a Thread.Sleep() is not the answer.

You need a robust scheduler that understands Cron Expressions.

A Cron expression allows you to define complex recurring schedules with a compact string.

For example:

• 15 10 * * * = 10:15 AM every day.
• 15 10 * * MON-FRI = 10:15 AM, but only on weekdays.

In the .NET ecosystem, the heavyweight champion for this is Quartz.

It parses these expressions for you and ensures your background services fire exactly when they are supposed to.

I broke down how to configure your first job and trigger in this article: https://lnkd.in/dfy-U7jp

What library do you currently use for background jobs?

---
Sign up for the .NET Weekly with 75K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dJGs8Uzv

Stop wasting the first 3 days of every project setting up boilerplate.

I built a free, production-ready Clean Architecture template to handle the plumbing for you.

It's already used by 40,000+ developers.

Here is what's inside the latest version:

- 𝗨𝗽𝗱𝗮𝘁𝗲𝗱 𝘁𝗼 .𝗡𝗘𝗧 𝟭𝟬
- 𝗡𝗲𝘄 .𝘀𝗹𝗻𝘅 𝗦𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝗙𝗼𝗿𝗺𝗮𝘁
- CQRS (No MediatR dependency)
- Structured Logging & Validation
- JWT Authentication & RBAC
- Docker Compose support

It saves you hours of setup time.

Download it for free here: https://lnkd.in/dbU3m97Q

💡 Why the shift to .slnx? The new Solution format in .NET is a game-changer. It replaces the old bloated syntax with a clean XML syntax. It makes merge conflicts in your solution file a thing of the past.

AI won’t replace you in 2026.

But a developer who knows how to engineer with AI absolutely will.

I recorded a new video on how I’m using AI tooling for .NET development (mainly Cursor), and the point isn’t Cursor.

It’s the workflow.

Here’s what I do every time:
- Plan first (write the steps before touching code)
- Give AI tight constraints (what to change, what NOT to change)
- Review like it’s a PR (DI lifetimes, packages, SQL correctness, connection handling)
- Verify (build + tests) so you don’t ship “confidently wrong” code

In the video, I walk through a real refactor: moving query handlers from EF Core to raw SQL/Dapper, and show the exact spots where AI gets you 90% there... and where that last 10% can break things.

You can watch it here: https://lnkd.in/dzZiKR66

If you’re trying to “stay relevant” in the AI era, the skill isn’t prompting.

It’s owning correctness.

No alternative text description for this image

Data Annotations are great for simple things.

[Required]? Fine.
[MaxLength(50)]? Sure.

But what happens when you need to validate something against a database?
Or validate based on a configuration setting in appsettings.json?

Data Annotations hit a wall. You can't easily inject dependencies into an Attribute.

This is where FluentValidation shines.

Because your validators are just standard classes registered in the DI container, you can inject anything you need.

Want to check if a CouponCode is valid by calling the IPricingService? Just inject it into the OrderValidator constructor.

It turns validation from a static check into a dynamic, intelligent logic layer.

I explain how to wire this up seamlessly in ASPNET Core here: https://lnkd.in/d9nVR8fu

It is a game changer for complex business rules.

---
Sign up for the .NET Weekly with 75K+ other engineers, and get a free Clean Architecture template: https://lnkd.in/dCMpeCNe

Software Engineer Manager at Juniper By IMC

Software Engineer Manager with over 15 years of experience in software development and three years of experience in management. Proficient in various programming languages, including [C#, Node, Javascript, and Python]. 

Experienced in leading cross-functional teams to deliver complex software projects on time and within budget. 

Strong background in agile development methodologies and project management.

Extensive experience in the entire life cycle of the software design process, including requirements definition, prototyping, design, implementation, testing, and maintenance.

Raul Junco

I treat system design like a checklist.

Here’s my go-to mental TODO list when building systems:

Fundamentals
- What are the read patterns?
- What are the write patterns?
- Who owns the source of truth?
- Is consistency or availability more critical?
- Single writer or multiple writers?

Architecture
- Synchronous or async?
- Do I need a queue, or is a cron job enough?
- Can I separate the compute from storage?
- Stateless or stateful services?
- Contracts versioned?

Reliability
- What happens when this fails?
- Where’s the retry logic, and is it idempotent?
- Are we alerting to symptoms or root causes?
- Timeouts configured?

Scaling
- How do reads scale? 
- How do writes scale?
- Will this design hold up at 10x traffic?
- What’s the hot path, and how do we optimize it?
 
Observability
- Do we log what we need to debug in production?
- Can we trace a request across services?
- What metrics define “healthy”?
- Debuggable without redeploy?

I don’t always follow it linearly, but if something breaks or feels off, this list helps me pinpoint what I may have overlooked.

Add something for the next engineer 👇

Every developer should know that tenant isolation is not a database problem. 

It’s a blast-radius problem.

I learned this the hard way.
One missing tenant filter. 
That’s all it takes to turn a normal deploy into a security incident.

Every multi-tenant system eventually picks one of three isolation levels.
Each one trades safety, cost, and operational pain in different ways.


1. Database per tenant

This is the strongest isolation you can get.
Each tenant lives in its own database. No shared tables. No shared state.

The upside is obvious.
A bug in one tenant cannot leak data from another. Audits are simpler. Compliance conversations are shorter. When something breaks, the blast radius stays small.

The downside shows up later.
Operational overhead grows fast. You manage hundreds or thousands of databases. Migrations become orchestration problems. Costs scale with tenant count, not usage.

This model works when tenants are large, regulated, or high-risk. It breaks down when you try to apply it blindly to long-tail customers.


2. Schema per tenant

This is the middle ground most teams underestimate.

All tenants share a database, but each one gets a separate schema. Tables stay isolated, but infrastructure stays manageable.

You get clearer boundaries than row-level isolation. You avoid the explosion of databases. Audits remain reasonable. Most accidental data leaks disappear.

But complexity still creeps in.
Migrations must run across many schemas. Cross-tenant reporting becomes awkward. Automation is not optional anymore. Without it, this model collapses under its own weight.

This approach works well when tenants vary in size and you want isolation without full separation.


3. Row-level isolation

This is the cheapest and most dangerous option.
All tenants share the same tables. Isolation lives in a tenant_id column and your queries. Infrastructure stays simple. Costs stay low. Scaling is easy.

The risk is brutal.
One missing filter equals a data leak. One refactor can break isolation. One rushed hotfix can expose everything. Security depends on every layer doing the right thing every time.

This model only works when you add heavy guardrails: strict query scoping, database policies, service-level enforcement, and tests that actively try to cross tenant boundaries.

Without those, you’re betting the company on discipline.

Tenant isolation is not a storage choice.
It’s a trust decision.

Learn this, it's a classic Interview question.

Not all latency problems are solved with cache.
Many are solved with better data layout.

Every option in this question sounds reasonable at first.

A. Covering indexes
Feels boring. No new system.
Just fewer bytes read per query.

B. Denormalized read table
Smaller rows. Cleaner reads.
Until async updates introduce stale profiles and user-facing bugs.

C. Columnar store
Columnar means fast scans. Great for analytics.
Bad fit for low-latency, strongly consistent lookups.

D. Redis cache
Instant reads. Simple story.
Also, cache invalidation, stampedes, and dual-write headaches.

Each option optimizes something.
But what's the simplest one that optimizes the hot path itself?
For me, the answer is A.

Why covering indexes win

1. They remove heap fetches from the hot path.
2. They keep reads shard-local with no extra network hops.
3. They preserve strong consistency without async pipelines.
4. They optimize data layout instead of adding systems.
5. They scale predictably with traffic, not complexity.

At the core, this is not a scaling or a redesign problem.
It’s a data access problem.

Only 45% picked the right answer.
That’s why we need to keep practicing simple decisions.

Your laptop is the worst place to learn AI agents.

Because it teaches the wrong habits.

- Agents with full access to everything
- API keys sitting in configs
- Workflows you can’t reproduce
- Automation you can’t audit

This feels fast, then you move to real systems, and everything breaks.

There’s a better way to learn.

Instead of starting on your laptop, you can start in a real, isolated development environment where:

- The environment is reproducible
- Agents run alongside your code, not on top of your machine
- Tasks are explicit and repeatable
- You can see what automation actually does

That is why Coder Community Edition is a strong option for developers learning AI-assisted development.

1. It is open source.
2. It is free.
3. You keep your IDE. (VS Code, JetBrains, ...)
4. You can run it in your browser or self-host it.

Just a safer place to learn how AI-assisted development really works.

Learning the right model early costs less than fixing the wrong one later.

No more “works on my machine.”
Instead 👉 https://fandf.co/4pKavqY


Thanks to the team at Coder for collaborating on this post!

Number one skill to break into AI or Software Dev in 2026!
=> Mastering Python for real-world data handling.

AI models need good data, period. And we have been saving messy data for decades. 

Wrong formats.
Missing values.
Inconsistent types.
Silent assumptions.

If you can’t read, clean, validate, and transform data correctly, no AI system will work reliably.

That’s why Python matters.

If I were starting today, this is the roadmap I’d follow:

Step 1: Build strong fundamentals
Control flow, functions, and core data types.

My go-to course:
Python Programming Fundamentals - https://lnkd.in/eRC_EqCw

Step 2: Practice real data handling
Transform data safely.
Handle bad inputs.
Make assumptions explicit.

For this, you need to practice with messy, imperfect, real datasets:
- Kaggle
- Data[.]gov
- World Bank Open Data

This is where you’ll build the muscle memory that sets top developers apart.

Step 3: Think like a 10X developer
Go deeper into data structures, reusable patterns, and reading real code.
A structured program is a huge time-saver here.

My course recommendation:
Associate Python Developer Track - https://lnkd.in/eg5hDjFk

Frameworks come later.
They always do.

FastAPI -> need APIs and backends and data validation and contracts matter
Pandas -> for structured data manipulation
NumPy -> for numerical data and performance
scikit-learn -> once data is clean and usable
PyTorch -> when models enter the picture

The bottom line: if you can wrangle data with Python, you’ll always be in demand in AI and software.
Build your foundation now and you’ll be ready for whatever tech comes next.

Thanks to the team at DataCamp for supporting this post.

taplio

The Battle

Milan Jovanović

Raul Junco

Followers

Posts

Engagements

Likes

Performance ScoreMilan Jovanović's LinkedIn Performance

Performance ScoreRaul Junco's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Milan Jovanović

Raul Junco

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Milan Jovanović's LinkedIn Performance

Performance ScoreRaul Junco's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreMilan Jovanović's LinkedIn Performance