Amey-Thakur Cyclomatic complexity
Cyclomatic complexity counts the independent paths through a function: one plus
each branch point (if, for, while, case, and every && or ||). Paths
multiply the tests a function needs and the states a reader must hold at once,
and past roughly ten paths a function stops fitting in one head.
Method
- Measure before you judge. Run a real tool:
radon cc -s file.py, the ESLintcomplexityrule,gocyclo, or SonarQube. Rank functions by score and start with the worst. A number ends the argument about whether a function merely "feels" complex. - Set a threshold, treat crossings as a signal not a law. Flag functions over 10 for a look and over 15 for a rewrite plan. The score is a smoke alarm: it says where to look, not that the code is wrong. Some parsers and state machines earn their number.
- Extract cohesive branches into named functions. Each block pulled out drops the parent's count and gives the logic a name and a seam to test. A 40-path function that becomes a 6-path coordinator calling five 8-path helpers is far more testable, even though the total branches barely move.
- Replace nesting with early returns. Guard invalid and edge cases at the
top and return, so the main path steps out from under the arrow of nested
ifblocks. Every level removed cuts the states a reader tracks through the rest of the function. - Turn a long switch on a value into table dispatch. Map keys to handlers:
handlers[event.type](event). A twelve-case switch scoring 12 becomes a lookup scoring 1, and adding a case is a map entry rather than a new branch. - Collapse compound conditions into named predicates.
if (a && b && !c)hides several paths in one line. ExtractisEligible(...)returning a boolean: the branches move into a small, separately testable function whose name states what it decides.
Checks
- Does the tool's score actually drop after the change, not just relocate to a helper you forgot to re-measure?
- Can you still name every path the function takes, or did extraction only hide them behind a call?
- Is the deepest nesting two levels or fewer once the guards are in place?
Boundaries
The score is one signal: a low number with terrible names is still unreadable, and a slightly high number in a genuinely branchy state machine can be the clearest form available. Do not extract single-use helpers with vague names just to move a metric. Defer length questions to function-size and profiling to the performance skills.