Microservices vs Monolith: The Honest Take

Microservices were the religion of the 2010s. The 2020s have been a quiet correction. Modular monoliths are back, and there are good reasons.

The Cost of Technical Debt, Realistically

Technical debt is real, but not every piece of it costs the same. The debt that hurts is the debt in code you touch often. Debt in code nobody touches is just code nobody touches.

So the actually useful question isn't "how much debt do we have?", it's "what debt is in our hot paths?" Fix the hot-path debt. Leave the cold-path debt. Life gets better.

The Architecture Decision That Stays With You

Architectural decisions have a half-life of about three years before they either compound into a solid system or become the thing everyone's working around. The ones that compound well share a few traits: they match the team's skills, they leave room for the most likely changes, and they don't require rare talent to operate.

The ones that age poorly usually chased a trend, optimised for a scale that never arrived, or assumed a team composition that didn't materialise.

Signs You've Over-Engineered

If onboarding a new engineer takes more than a few days, something's off. If a local dev environment needs Docker Compose and six services running to test a simple change, something's off. If every deploy requires a ritual that only two people understand, something's off.

Not all of these are urgent fixes. But they accumulate as tax on every future change. Budget occasional time to pay the tax down.

The Testing Strategy That Matches Your Stage

Testing strategy varies by company stage and nobody tells you this. At pre-PMF, extensive test coverage can actually slow you down because the code you're testing is likely to get thrown away. At PMF and beyond, inadequate testing is a pure liability that will bite you repeatedly.

Our rough rule: pre-PMF, test the two or three things whose failure would embarrass you (auth, payments, data integrity). Post-PMF, grow coverage deliberately in the areas with the most change or the most risk. Don't chase coverage percentages as a vanity metric. Chase the confidence to ship without anxiety. Related read: our post on legacy modernisation covers the flip side of this.

Signs You've Under-Engineered

The opposite side: production outages that could have been caught by tests nobody wrote. Customer data in states that shouldn't be possible because there's no validation layer. A rewrite conversation every eighteen months because the codebase has calcified.

Under-engineering and over-engineering are both real. The sweet spot moves with team size and maturity. Check in on it every quarter.

How We Make These Calls

On every project, we do a written architecture decision record for anything that's hard to change. What we're deciding. What options we considered. Why we picked this one. What we'd reconsider if conditions changed.

It takes thirty minutes per decision and saves literal months of future second-guessing. A year in, when someone asks "why did we do it this way?", the ADR has the answer. Fewer religious debates, more work.

Documentation That Actually Gets Read

Most developer documentation goes unread. The problem isn't usually that developers are lazy — it's that the documentation isn't targeted at what they need. Long architecture documents nobody reads. Outdated READMEs. Tribal knowledge that never makes it to a wiki.

The docs that actually get read are task-focused and short. How do I run this locally? How do I deploy? How do I debug a failing test? Answer those questions well and you've solved 80% of the documentation problem. Leave the comprehensive architecture docs for when someone has already decided to learn the system deeply.

The State of Software Engineering in 2026

Software engineering has been through a few shifts in the last few years that are worth naming. AI-assisted development is now mainstream — most professional engineers use tools like GitHub Copilot, Cursor, or Claude for daily work. The productivity gains are real but less uniform than marketing suggests. Strong engineers get more out of AI than weak ones, which is widening the gap between the best and the rest.

Architecturally, the pendulum has swung back toward simplicity. Distributed systems are having a moment of re-evaluation. Microservices fever has cooled. Modular monoliths are respected again. Kubernetes is being rightly identified as overkill for most teams. These corrections are healthy and long overdue.

The craft of shipping software, though, remains unchanged in its essentials. Clear thinking, testability, observability, incremental improvement, and good relationships with the people you're building for. No framework or AI tool replaces those. The teams that win are the ones who've stayed disciplined about the fundamentals while adapting to new tools.

Myths That Slow Teams Down

One: 'we'll clean this up later'. Technical debt in hot paths compounds. Debt you don't fix soon becomes debt you'll never fix. Decide early whether something is worth cleaning up now versus leaving forever.

Two: 'we need to rewrite this from scratch'. Almost always wrong. Rewrites take longer than estimated, pause feature work, and often produce a worse version of the original. Refactor incrementally in nearly all cases.

Three: 'more testing is always better'. Diminishing returns are real. Testing infrastructure and test suites take time to build and maintain. Match testing investment to actual risk, not abstract coverage targets.

A Rewrite Decision We're Glad We Made

A SaaS client we work with had a core scheduling engine that had been rewritten from the original Python to Go in 2023 to handle scale. The Go rewrite worked for about eighteen months, then started showing architectural pain as new requirements landed.

The discussion in early 2025 was whether to rewrite again, to Rust this time, for better concurrency. We talked them out of it. The Go system was about 60% of ideal, which is fine. The pain points were concentrated in three specific modules, not the whole system.

We refactored those three modules instead. Six weeks of focused work, no Rust migration, no new team skills to hire for. Performance targets met, pain points resolved, roadmap preserved. This is usually the right answer. Rewrites feel clean. They're almost never the highest-ROI move on a running system.

The Short Checklist

If you take nothing else from this post, take this checklist. It's what we'd hand to someone just starting out in this area. None of it is revolutionary. All of it is worth doing. The compound effect of consistently doing these things, even without any other clever moves, is meaningful over a year or two. We'd rather see a team do the checklist competently than chase the latest trend while skipping the fundamentals.

1. Write architecture decision records for anything hard to change.
2. Pay down tech debt in your hot paths. Ignore debt in cold paths.
3. Have a 20-minute local setup for any new engineer. Longer is a signal.
4. Review your production error rate weekly. Every error above floor is a message.
5. Keep deployments boring. Excitement during deploy is a process failure.
6. Run retrospectives regularly. Act on two items per retro, not twelve.
7. Monitor p95 and p99 latency, not just averages. Your worst experiences hurt most.
8. Keep a running architecture decision log. Future-you will thank you.
9. Review error rates weekly. Not monthly. Weekly.
10. Run regular game-day exercises for critical failure modes.

If you've read this far, you probably care about getting this right. That's more than most people do, and it matters. Our offer stands: if you're stuck or want a sanity check on something you're planning, drop us a line. We've been through most versions of this particular problem and we're happy to share what worked. The best conversations we have are usually with people who've already done the reading and want to sharpen their thinking rather than start from zero. Either way, good luck with whatever you're building. The fact that you're thinking about this carefully gives you a meaningful head start on the people who aren't.

Closing thought: everything we've written here comes out of actual client work, not theory. If you try something based on this post and it doesn't work the way we described, that's a useful data point — both for you and for us. We'd be curious to hear what happened, either way.