Skip to content

Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.

fix(git): align in-memory churn_percentile with SQL PERCENT_RANK and tie handling (#4159) (#4189)

Rishabh Mishra58a153e599b4
diff --git a/packages/core/src/repowise/core/ingestion/git_indexer/enrich.py b/packages/core/src/repowise/core/ingestion/git_indexer/enrich.py
index f45c1518f..9308d52c6 100644
--- a/packages/core/src/repowise/core/ingestion/git_indexer/enrich.py
+++ b/packages/core/src/repowise/core/ingestion/git_indexer/enrich.py
@@ -267,42 +267,65 @@ def compute_percentiles(metadata_list: list[dict]) -> None:
if not metadata_list:
return
- # Sort by temporal_hotspot_score (primary) then commit_count_90d (tiebreak)
- sorted_by_churn = sorted(
- range(len(metadata_list)),
- key=lambda i: (
- metadata_list[i].get("temporal_hotspot_score") or 0.0,
- metadata_list[i].get("commit_count_90d", 0),
- ),
- )
+ # Non-code history-tier files do not participate in code churn rankings.
+ code_metas = [m for m in metadata_list if not m.get("history_only")]
+ for meta in metadata_list:
+ if meta.get("history_only"):
+ meta["churn_percentile"] = 0.0
+ meta["is_hotspot"] = False
+ meta.setdefault("change_entropy_pct", 0.0)
+ meta.setdefault("co_change_scatter_pct", 0.0)
+ meta.setdefault("prior_defect_pct", 0.0)
+
+ if not code_metas:
+ return
- total = len(metadata_list)
- for rank, idx in enumerate(sorted_by_churn):
- meta = metadata_list[idx]
- churn_pct = rank / total if total > 0 else 0.0
- meta["churn_percentile"] = churn_pct
- # Hotspot: top 25% churn (churn_percentile >= 0.75) AND the absolute
- # activity floors — without the floors, a quiet repo's top quartile is
- # just "every file touched in 90 days" (issue #361). Folded into the
- # rank loop so the list is scanned once.
- if churn_pct >= 0.75 and meets_hotspot_floors(meta):
- meta["is_hotspot"] = True
+ total = len(code_metas)
+ if total < 2:
+ for meta in code_metas:
+ meta["churn_percentile"] = 0.0
+ else:
+ # Group keys to mirror SQL PERCENT_RANK() with tie-awareness:
+ # (RANK() - 1) / (total - 1)
+ keys = [
+ (
+ float(m.get("temporal_hotspot_score") or 0.0),
+ int(m.get("commit_count_90d") or 0),
+ )
+ for m in code_metas
+ ]
+ key_counts = Counter(keys)
+ rank_lookup: dict[tuple[float, int], float] = {}
+ running = 0
+ for key in sorted(key_counts):
+ rank_lookup[key] = running / (total - 1)
+ running += key_counts[key]
+
+ for meta, key in zip(code_metas, keys, strict=True):
+ churn_pct = min(rank_lookup[key], 1.0)
+ meta["churn_percentile"] = churn_pct
+ # Hotspot: top 25% churn (churn_percentile >= 0.75) AND the absolute
+ # activity floors — without the floors, a quiet repo's top quartile is
+ # just "every file touched in 90 days" (issue #361). Folded into the
+ # rank loop so the list is scanned once.
+ if churn_pct >= 0.75 and meets_hotspot_floors(meta):
+ meta["is_hotspot"] = True
# Files with zero entropy — every file on the ESSENTIAL tier, plus
# FULL-tier files that only ever changed alone — stay at 0.0 so the
# change_entropy biomarker is silent for them.
_rank_within_eligible(
- metadata_list, source_key="change_entropy", target_key="change_entropy_pct"
+ code_metas, source_key="change_entropy", target_key="change_entropy_pct"
)
# Ranked on the decayed partner mass, not the count: a count saturates at
# the storage cap and never retires, so it can only ever ratchet up.
_rank_within_eligible(
- metadata_list, source_key="co_change_mass", target_key="co_change_scatter_pct"
+ code_metas, source_key="co_change_mass", target_key="co_change_scatter_pct"
)
# Bug-fix history, ranked over every file: a file with no fixes in the
# window is a measured zero, not a missing signal.
_rank_over_population(
- metadata_list, source_key="prior_defect_count", target_key="prior_defect_pct"
+ code_metas, source_key="prior_defect_count", target_key="prior_defect_pct"
)
diff --git a/tests/unit/ingestion/test_git_history_tier.py b/tests/unit/ingestion/test_git_history_tier.py
index 7ecb820f2..8e3ee7b2d 100644
--- a/tests/unit/ingestion/test_git_history_tier.py
+++ b/tests/unit/ingestion/test_git_history_tier.py
@@ -117,7 +117,7 @@ async def test_non_code_files_get_history_from_the_shared_walk(tmp_path, monkeyp
assert meta["a.py"]["history_only"] is False
assert meta["a.py"]["commit_count_total"] == _git_log_count(tmp_path, "a.py")
assert meta["a.py"]["primary_owner_line_pct"] is not None
- assert {meta["a.py"]["churn_percentile"], meta["b.py"]["churn_percentile"]} <= {0.0, 0.5}
+ assert {meta["a.py"]["churn_percentile"], meta["b.py"]["churn_percentile"]} <= {0.0, 1.0}
async def test_update_writes_the_same_history_rows_as_a_full_index(tmp_path, monkeypatch) -> None:
@@ -206,3 +206,23 @@ def test_history_tier_keeps_files_in_reincluded_dir(tmp_path: Path) -> None:
assert kept_custom == {"src/build/README.md"}
+def test_two_code_files_plus_history_only_percentile_calculation():
+ from repowise.core.ingestion.git_indexer.enrich import compute_percentiles
+
+ metas = [
+ {"file_path": "a.py", "temporal_hotspot_score": 10.0, "commit_count_90d": 5},
+ {"file_path": "b.py", "temporal_hotspot_score": 2.0, "commit_count_90d": 1},
+ {"file_path": "README.md", "temporal_hotspot_score": 0.0, "commit_count_90d": 20, "history_only": True},
+ {"file_path": "config.json", "temporal_hotspot_score": 0.0, "commit_count_90d": 5, "history_only": True},
+ ]
+ compute_percentiles(metas)
+
+ # History-only files get 0.0 churn percentile and do not participate in code ranking
+ assert metas[2]["churn_percentile"] == 0.0
+ assert metas[3]["churn_percentile"] == 0.0
+
+ # Two code files: lower is 0.0, higher is 1.0
+ assert metas[0]["churn_percentile"] == 1.0
+ assert metas[1]["churn_percentile"] == 0.0
+
+
diff --git a/tests/unit/test_git_indexer.py b/tests/unit/test_git_indexer.py
index 068a090cf..6a25382e5 100644
--- a/tests/unit/test_git_indexer.py
+++ b/tests/unit/test_git_indexer.py
@@ -318,6 +318,48 @@ class TestHotspotAbsoluteFloors:
assets = next(m for m in metas if m["file_path"] == "assets.bin")
assert assets["is_hotspot"] is True
+ def test_percentiles_parity_with_sql_percent_rank(self) -> None:
+ """N=5 percentiles mirror SQL PERCENT_RANK() exactly: [0.0, 0.25, 0.50, 0.75, 1.00]."""
+ from repowise.core.ingestion.git_indexer.enrich import compute_percentiles
+
+ metas = [
+ {
+ "file_path": f"f{i}.py",
+ "temporal_hotspot_score": float(i * 10),
+ "commit_count_90d": 10,
+ "lines": 100,
+ }
+ for i in range(1, 6)
+ ]
+ compute_percentiles(metas)
+ pcts = [m["churn_percentile"] for m in metas]
+ assert pcts == [0.0, 0.25, 0.50, 0.75, 1.00]
+ # f4 and f5 are >= 0.75 and clear floors
+ assert metas[0].get("is_hotspot", False) is False
+ assert metas[1].get("is_hotspot", False) is False
+ assert metas[2].get("is_hotspot", False) is False
+ assert metas[3].get("is_hotspot", False) is True
+ assert metas[4].get("is_hotspot", False) is True
+
+ def test_percentiles_tie_breaking_shared_rank(self) -> None:
+ """Ties share the same rank and hotspot classification instead of being arbitrarily split."""
+ from repowise.core.ingestion.git_indexer.enrich import compute_percentiles
+
+ metas = [
+ {"file_path": "quiet.py", "temporal_hotspot_score": 1.0, "commit_count_90d": 2, "lines": 50},
+ {"file_path": "tied_a.py", "temporal_hotspot_score": 50.0, "commit_count_90d": 12, "lines": 100},
+ {"file_path": "tied_b.py", "temporal_hotspot_score": 50.0, "commit_count_90d": 12, "lines": 100},
+ {"file_path": "tied_c.py", "temporal_hotspot_score": 50.0, "commit_count_90d": 12, "lines": 100},
+ ]
+ compute_percentiles(metas)
+ assert metas[0]["churn_percentile"] == 0.0
+ # All three tied items have exactly 1 file scoring strictly lower out of 3 possible: 1 / 3 = 0.333...
+ assert metas[1]["churn_percentile"] == pytest.approx(1 / 3)
+ assert metas[2]["churn_percentile"] == pytest.approx(1 / 3)
+ assert metas[3]["churn_percentile"] == pytest.approx(1 / 3)
+ # Hotspot status is identical for all three
+ assert metas[1].get("is_hotspot", False) == metas[2].get("is_hotspot", False) == metas[3].get("is_hotspot", False)
+
class TestCountActiveContributors:
"""count_active_contributors derives the repo's 90-day team size from