santosomar/szz-bug-identifier

SZZ Bug Identifier

SZZ (Śliwerski, Zimmermann, Zeller, 2005) answers: given a bug-fix commit, which earlier commit introduced the bug? It works by blaming the lines the fix touched.

The algorithm — classic SZZ

1. Identify the fix commit. Usually via a link to an issue tracker (Fixes #1234, BUG-567) or a message keyword (fix, bug, patch).
2. Extract the deleted/modified lines. In the fix's diff, every line removed or changed is a line that was (potentially) buggy.
3. Blame each of those lines. git blame <fix>^ -- <file> on each modified line to find the commit that last touched it before the fix.
4. Those blamed commits are bug-introducing candidates.

That's the whole algorithm. The rest is noise filtering.

Noise filters — classic SZZ's known weaknesses

| Noise source | Why it's wrong | Filter | | --------------------------------- | -------------------------------------------------------- | ------------------------------------------------------ | | Whitespace / formatting changes | Blame hits a prettier run, not the real introducer | git blame -w; ignore commits that only touch formatting | | Comment-only changes | The fix edited a comment too — that line is not the bug | Strip comment lines before blaming | | Large refactor commits | Every line blames to the Great Refactor of 2019 | git blame --ignore-rev with a curated ignore-list | | The line was added by the fix | No blame target — added lines didn't exist before | Only blame deleted/modified lines, not added | | Bug predates the repo | Blame hits the initial import commit | Flag — can't attribute | | Moved file | Blame stops at the git mv | git blame -C -M to follow moves/copies | | Blamed commit is newer than bug report | The bug existed before that commit; blame is wrong | Discard candidates with commit-date > bug-report-date |

Worked example

Fix commit: c4a9f1b — Fix: null check in getUserEmail (closes #892)

  public String getUserEmail(long id) {
      User u = repo.find(id);
-     return u.getEmail();
+     if (u == null) return null;
+     return u.getEmail();
  }

Step 2: Modified line is return u.getEmail(); (the old version).

Step 3: git blame c4a9f1b^ -- UserService.java at that line → a17d3e0 — Add UserService (Jane, 2021-03-04).

Step 4: Candidate = a17d3e0.

Filters:

• Whitespace? No, it's a real code line. ✓
• Comment? No. ✓
• Refactor in ignore-list? No. ✓
• Added by the fix? No — this is the old version of a modified line. ✓
• Newer than bug report? Issue #892 opened 2023-11-01; a17d3e0 is 2021. ✓

Verdict: a17d3e0 introduced the bug. The null-check was never there.

Edge cases

• Multi-line fix where lines blame to different commits: Each blamed commit is a candidate. Usually one is the real introducer and the rest are incidental; → semantic-szz-analyzer to distinguish.
• Fix is a one-line addition (no deletion): Classic SZZ has nothing to blame. Heuristic: blame the surrounding lines (the ones the added line sits between). Low confidence.
• The fix reverts an earlier commit entirely: git revert — the reverted commit IS the bug-introducer, definitionally. Shortcut: check if the fix is a revert before running SZZ.
• Cherry-picked commits: The blame may point at the cherry-pick, not the original. Check for (cherry picked from commit …) trailers and follow the chain.

Do not

• Do not treat SZZ output as ground truth. Published precision is 40–70% without filters. It's a candidate generator for human review.
• Do not run SZZ on a repo without a curated .git-blame-ignore-revs. Without it, every result blames the last formatting pass.
• Do not use commit-message keyword matching (fix, bug) as your only fix-identification signal. False-positive rate is brutal. Prefer issue-tracker links.
• Do not run SZZ at bug-fix time for a single bug — that's what → regression-root-cause-analyzer is for (bisect is more precise). SZZ is for batch mining.

Output format

fix: <sha> — <subject>
  candidates:
    <sha> — <subject>  (<date>, <author>)
      blamed from: <file>:<line>
      filters passed: whitespace ✓  comment ✓  date ✓
      confidence: <high|medium|low>
    ...