Insights · Architecture · 8 min read
From monolith to modular — without the regret
"Composable regret" is now a named phenomenon, and a major vendor has publicly walked away from MACH. Both failures repeat the same mistake the monolith made.
Ask an operator why they are still on a platform they have outgrown and you will rarely hear that they like it. You will hear some version of: we know what replacing it costs, and we cannot face it. For years the industry's prescribed answer was composable — decompose everything, best-of-breed every layer, MACH all the way down.
The last two years have complicated that story, usefully.
What the market just learned
"Composable regret" has become a term of art. Enterprises that decomposed aggressively found themselves acting as their own system integrators — stitching APIs across a dozen vendors, debugging data flows that no single supplier owned, and watching engineering backlogs multiply while marketing waited weeks for changes that used to take days. Total cost of ownership analyzes have reported migrations landing at two to three times the cost of the original build, because a full decomposition is not a port — it is a rebuild of the architecture pattern itself, plus an integration bill that renews annually.
The clearest signal came from inside the movement: in 2025 VTEX publicly left the MACH Alliance, its co-CEO describing the dogmatic version of the journey as "paved with hidden costs, operational nightmares, unfulfilled promises." Retail technology leaders followed with a franker assessment than the previous decade had permitted — that below a certain scale and engineering depth, the fully composed stack costs more than it returns, and plenty of businesses doing the math are deliberately keeping a suite at the core.
It would be easy to read all that as vindication for the monolith. It is not — the monolith's problems did not go anywhere. Twenty-year-old platforms still block every initiative that touches them, still concentrate risk in systems nobody fully understands, and still cannot move at the speed the business needs. The correct reading is narrower and more useful: a big-bang rebuild into many fine-grained parts carries the same concentrated risk as a big-bang rebuild into one, plus permanent integration overhead. The failure was never really about the target architecture. It was about the size of the bet.
Dogmatic composable failed for the same reason big-bang replatforms fail. The unit of success was the whole program.
The version that works
The migrations that succeed share a shape regardless of destination. They pick the module where the pain is sharpest — commissions, commerce, the back office — and replace that, while everything else keeps running. The strangler pattern, in its practical form: put a seam around one capability, build the replacement behind the seam, run old and new in parallel until the numbers reconcile, cut over, and only then choose the next piece. Value lands in steps. Every step is reversible. Nobody bets the company on a cutover weekend, and — just as important — the organization learns the new architecture one module at a time instead of all at once.
The discipline that makes it work is at the seams. Each module needs a clean contract with the rest of the system — orders in, events out — and the data flows across those contracts have to be owned, monitored and reconciled continuously, because the parallel-run period is where the confidence comes from. Teams that skip the reconciliation to save time are re-importing the big-bang risk through the side door.
In channels where the systems touch people's earnings, this matters doubly, because the riskiest systems are the ones a salesforce feels. Replace a commission engine inside a big-bang program and payout timing becomes one line in someone else's risk register. Replace it as its own step, dual-run against the old engine, and you can show the field that the numbers match before anything changes for them. The sequencing is not just risk management — it is trust management.
The test to apply
When evaluating any platform, one question does most of the work: can I adopt the part I need without adopting the parts I do not?
If the answer is no, you are not being sold a platform. You are being sold a project — and the market has now run that experiment at scale, in both directions, with the same result. The monolith concentrated risk in one big decision. Dogmatic composable scattered it across thirty small vendors and called that progress. The survivable path is the boring one in the middle: modular where modularity pays, incremental always, and reversible at every step.