Evaluate technical debt through delivery friction, operational risk, and maintenance cost. Make room for a bounded improvement and check whether it helps.
Evaluate technical debt through delivery friction, operational risk, and maintenance cost. Make room for a bounded improvement and check whether it helps.
A roadmap can look achievable until the same deployment workaround, brittle integration, or shared module consumes another week. The feature estimate receives scrutiny. The recurring work around the system becomes background noise.
Then someone proposes a technical-debt sprint. Someone else asks which customer commitment should move. Without evidence about the constraint, both positions are difficult to evaluate.
The useful management decision is specific: how much capacity should be spent removing this constraint now, what work moves to make room, and what result would justify the investment?
I treat technical debt here as a decision about present and future engineering capacity. That framing connects implementation concerns to delivery commitments while leaving room for risks that cannot sensibly be reduced to hours saved.
For this discussion, a technical-debt item is an existing design or implementation choice that creates avoidable effort or restricts a needed change. The evidence should identify both the technical cause and the consequence.
Consider three descriptions:
The third description gives a manager something to investigate. It identifies the change being made, the additional work, and a possible failure mechanism. It still needs validation, but it is a stronger starting point than a label on a backlog ticket.
The original metaphor is useful at this level. In Ward Cunningham's 1992 experience report, incremental delivery creates a need for consolidation, and continuing to work around an unsuitable implementation consumes effort. That is the recurring cost this article examines.
An old component is not automatically a priority. Neither is unfamiliar code, a missing feature, or a design that differs from a preferred pattern. Routine operations also need capacity whether or not the underlying system has a debt problem. A useful proposal explains the additional burden associated with the specific constraint.
This distinction protects the discussion from becoming an argument about engineering taste. A small, rarely changed module can be an acceptable compromise. A small module on every critical delivery path can deserve attention immediately.
Three questions help describe the item without compressing everything into a score.
Recurring friction: What extra work happens when the affected path is used? Examples include repeated manual reconciliation, unnecessary coordination, duplicated changes, and investigations caused by misleading diagnostics. Record the frequency as well as the effort per occurrence.
Removal effort: What would a bounded improvement require? Include investigation, tests, rollout, migration, documentation, and removal of the old path. A refactor estimate that stops when the new abstraction compiles leaves substantial work unaccounted for.
Risk and future exposure: What could happen if the constraint remains? Identify the affected capability, the consequence, and the conditions that make the exposure relevant. A restore procedure that has never been exercised may deserve attention even if nobody currently spends time using it.
Some exposure needs prompt action under an existing reliability or security policy. It should not have to win a contest for the shortest effort payback. Other items can remain accepted compromises with an owner and a review trigger.
Technical debt can also emerge as understanding changes. Martin Fowler's debt quadrant distinguishes deliberate and inadvertent debt, as well as prudent and reckless choices. That distinction helps keep a review focused on the current decision instead of blaming the person who made an earlier one.
Capacity planning already involves deductions: leave, recurring commitments, support, and operational work. Technical friction becomes misleading when it is either omitted or deducted twice.
Consider a fictional five-engineer backend team planning a ten-working-day cycle. All numbers in this example are illustrative. Engineer-days represent a coarse effort budget, not interchangeable people or guaranteed delivery dates.
| Capacity category | Engineer-days | Treatment |
|---|---|---|
| Gross availability | 50 | Five engineers across ten working days |
| Leave and recurring commitments | 10 | Deduct once |
| Routine support and mandatory operations | 8 | Deduct once, excluding the friction below |
| Initiative capacity remaining | 32 | Budget available for features and improvements |
| Recurring friction within initiative work | 4 | Included within the 32 days |
| Direct feature progress with no repair | 28 | The remaining effort in this simplified scenario |
The four friction days are already part of the 32-day initiative budget. They are not another team or an extra allocation on top. If feature estimates already include those workarounds, adding a separate four-day deduction would count the same burden twice.
Likewise, if incident response is included in the eight operations days, the same incident hours must not appear again under friction. Use mutually exclusive categories, even when the first estimates are rough.
This model is a way to expose a trade-off. Work has dependencies, specialists have different constraints, and calendar time includes waiting. A six-engineer-day task may take several calendar weeks if it requires a particular reviewer or an external migration window. Capacity accounting cannot replace that dependency analysis.
Return to the fictional export service. Suppose a brief review of recent changes attributes roughly four engineer-days per cycle to duplicated rule edits, manual comparisons, and rework. The team proposes a six-engineer-day improvement.
The scope is deliberately narrow: put the shared rule behind one tested boundary, migrate the three handlers, and remove the duplicate paths. The proposal does not include rewriting the export engine or adopting a new framework.
For planning, assume the existing friction persists throughout the implementation cycle. That is a conservative simplification, not a claim about when benefits actually begin.
| Scenario | Initiative budget | Improvement effort | Friction inside feature work | Direct feature effort |
|---|---|---|---|---|
| Keep the existing path | 32 | 0 | 4 | 28 |
| Fund the repair this cycle | 32 | 6 | 4 | 22 |
| Later comparable cycle, if the repair succeeds | 32 | 0 | 1 | 31 |
The immediate trade-off is six fewer engineer-days of direct feature effort during the implementation cycle. The remaining 26 planned feature days still include four days of friction, leaving 22 in this model. A manager needs to identify which feature scope or commitment changes to absorb that allocation.
If later friction falls from four days to one, the proposed saving is three engineer-days per comparable cycle. Six days of effort divided by three days saved gives a simple recovery period of two comparable cycles after completion.
That arithmetic is a hypothesis check. It excludes the value of the displaced feature, changing demand, transition surprises, and the possibility that the repair creates new work. It is not a financial return or a promise to recover a delivery date.
Test a less favourable scenario before committing. If the improvement takes nine days and saves only one day per cycle, simple effort recovery takes nine comparable cycles. If that service will be retired in three cycles, an effort-saving case alone looks weak. If a critical release depends on removing the duplication, the relevant justification may instead be enabling that release safely.
The assumptions deserve as much attention as the division. Is another export change actually planned? Does the duplication cause the rework, or is the rule itself unclear? A short discovery task can be more valuable than a confident estimate built on the wrong cause.
A useful record can be short enough to read alongside the roadmap. It should answer the questions needed to allocate capacity.
| Field | Example for the fictional export service |
|---|---|
| Constraint | One rule is implemented separately in three handlers |
| Evidence | Recent changes required repeated edits, manual comparison, and correction |
| Current effect | About four engineer-days of attributed friction per comparable cycle, with uncertainty |
| Proposed boundary | One tested rule implementation, three migrated callers, old copies removed |
| Effort and dependencies | Six engineer-days proposed; shared reviewer and rollout window required |
| Displaced work | Name the feature slice moved out of this cycle during planning |
| Success check | Reduced comparison/rework effort while preserving export correctness |
| Owner and review | Named engineering owner; review after two comparable change cycles |
Add rollout and rollback expectations when the change affects a production path. For this example, that could mean representative rule tests, comparing outputs before switching callers, and keeping a recoverable release boundary. A shared rule also creates a common dependency, so reviewers should check that the handlers truly need identical behaviour.
Keep fact, estimate, and assumption distinct. A review comment is evidence of a concern. It is not evidence that the concern costs four days. A ticket label is an inventory mechanism. It is not an assessment of business impact.
The Software Engineering Institute's discussion of technical-debt consequences also covers architectural constraints and how their significance depends on future system changes. That is why this record names the affected work and the planning horizon.
The right delivery route depends on the shape of the constraint.
Include it in a feature when the feature already touches the boundary. Consolidating the rule while implementing the next export change may share tests and rollout work. Keep the improvement explicit in the estimate and acceptance criteria so it does not become invisible extra scope.
Protect a small allocation when friction is recurring and improvements are independent. A team may reserve capacity for a short list of evidenced items. Define what happens if incidents consume that allocation. Record the displacement and revise the plan instead of assuming the improvement will somehow fit later in the same cycle.
Create a dedicated initiative when migration spans boundaries. A service split, data migration, or replacement of a shared protocol may need coordinated ownership, compatibility work, and a staged cutover. Give it milestones that retire specific constraints, including the point at which the old path disappears.
A percentage can support a local planning policy, but it cannot select the work or establish that an improvement is worthwhile. There is no percentage in this article that applies to every team. A newly launched service, a mature platform, and a system approaching retirement have different needs.
Capacity also needs the right skills and attention. Reserving six days on a spreadsheet does little if the only engineer able to review the change is already committed elsewhere. Agree on the actual participants and limit concurrent work enough for the allocation to be usable.
Leaving debt in place can be reasonable. A constraint may affect a low-value path, the repair may be riskier than continued operation, or another commitment may have greater immediate importance.
Record the accepted consequence and the condition that would change the decision. For the export example, a deferral might read:
Keep the duplicated rule through the current release. Reconsider before adding another export format, or if another change misses a handler. The service owner will review the item at the next planning session.
This is an illustrative decision, not a recommendation to tolerate a known production failure. The important parts are a named boundary, an owner, and a trigger that someone will actually review.
Be equally explicit when accepting a new shortcut. State why it is useful now, what future change it makes harder, and whether a follow-up is already funded. A ticket with no owner or revisit condition offers little protection against indefinite deferral.
Deferral should not transfer the burden into evenings or unpaid recovery work. If capacity is unavailable, adjust the scope or commitment. An improvement plan that relies on engineers privately absorbing the work does not represent the team's actual capacity.
When the improvement ships, compare the observed result with the initial hypothesis. For the export service, examine a small set of comparable rule changes: were coordinated edits removed, did manual comparison decrease, and did correctness remain acceptable?
Use team-level evidence that can be collected with little overhead. Existing change records, support notes, and a short retrospective may be enough. Detailed individual timesheets would introduce a measurement burden that this example does not require.
Demand may change, and the next task may be harder than the previous one. If the affected path has not been exercised again, the recurring benefit is still unverified. Mark that uncertainty and schedule an appropriate check rather than declaring the saving achieved when the code merges.
If friction remains, revisit the cause. Perhaps the real constraint was an unclear business rule, a slow approval process, or a shared test environment. Further refactoring might leave that constraint untouched.
Close the initiative when the agreed boundary is complete and the result has been assessed. There will usually be another possible improvement nearby. Evaluate it as another decision instead of expanding the original allocation indefinitely.
Take one repeated source of engineering friction into the next planning conversation. Describe the constraint, attach evidence, propose the smallest useful intervention, and identify the work that moves to fund it. Give the result an owner and a review point.
For the fictional team, that produces a clear commitment: reserve six engineer-days to consolidate the export rule, reduce feature scope accordingly, and check the proposed reduction in recurring work after two comparable cycles. If the assumptions change, revise the decision openly.
That is the capacity decision: choosing how much delivery effort to invest now, what exposure to accept in the meantime, and what evidence will show whether the investment helped.
Related reading: explore engineering practices and system-design articles or all articles.
Scala, JVM, AI, and backend systems. I send practical articles when there is something worth reading.
Join the newsletterNecessary cookies keep the site working. With your permission, I use analytics cookies to understand which articles are useful.