Minimal failure reproduction

Minimal Reproduction & Delta Debugging

When to Use

Use once an initial reproduction (or candidate failure path) is found, or when narrowing down a complex multi-step workflow, large payload, or multi-variable environment to the smallest set of conditions that still triggers the exact original bug.

Andreas Zeller's Delta Debugging (ddmin) Workflow

Never change five variables at once when narrowing a bug. Follow a disciplined Hypothesis $\rightarrow$ Single-Variable Experiment $\rightarrow$ Observation Log loop:

1. Partition & Binary-Search the Input / Step Space

Given a failing scenario with $N$ steps, fields, or configuration flags:

1. Cut half of the non-essential setup steps, payload properties, or query filters.
2. Run the check and verify whether the exact same Failure Signature still occurs:
   • Same failure occurs ($\text{FAIL}_{\text{same}}$): The removed half is irrelevant; discard it permanently and bisect the remaining half.
   • Bug disappears ($\text{PASS}$): The removed half contains a required trigger condition; restore it and bisect the other half (or bisect inside the restored half).
   • Different error occurs ($\text{FAIL}_{\text{other}}$ / invalid input): The cut broke a structural precondition (e.g., missing required foreign key or syntax error); restore only the minimal structural dependency and continue.

2. Minimize Across Four Orthogonal Axes

• Action Sequence: Reduce a 15-click UI journey or 10-call API sequence to the minimum 1–3 state transitions required.
• Data Payload: Shrink a 5 MB JSON/CSV/file or 50-column database seed down to the smallest record count and minimal string/number values that still fail.
• Component Scope: If a bug inside a pure parser, validator, or service method can be triggered directly without spinning up the full browser + HTTP server, provide both the direct unit/integration reproduction and note its equivalence to the full flow.
• Configuration Delta: Strip custom settings until only the specific flag or environment difference that flips PASS $\rightarrow$ FAIL remains.

3. Maintain a Concise Experiment Log

Record every variable tested in a brief table (Attempt # | Changed Condition | Observed Result) so dead ends are never repeated and non-reproducible boundaries are documented with proof.

4. Guard Faithfulness ("Minimal $\neq$ Distorted")

After reaching the minimal case, verify Symptom Identity:

• Confirm the minimal reproduction produces the exact same error code, stack trace, or wrong state as the original report—never simplify away the real trigger and substitute a trivial validation error.