Skip to content

:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.

fix(replay): drop surfacing features that overflow float32 before scoring (#124935)

Tue Haulund65834f51b124
diff --git a/posthog/temporal/session_replay/surfacing_scoring_sweep/features.py b/posthog/temporal/session_replay/surfacing_scoring_sweep/features.py
index 2ed284d98b81..efe107822f8b 100644
--- a/posthog/temporal/session_replay/surfacing_scoring_sweep/features.py
+++ b/posthog/temporal/session_replay/surfacing_scoring_sweep/features.py
@@ -54,6 +54,9 @@ ID_COLUMNS: tuple[str, ...] = ("team_id", "session_id", "distinct_id", "min_firs
_NUMERIC = "iuf"
_FLOAT = "f"
+# XGBoost casts features to float32, so a finite float64 above this still reaches it as inf.
+_FLOAT32_MAX = float(np.finfo(np.float32).max)
+
@dataclass(frozen=True)
class FeatureSpec:
@@ -244,10 +247,10 @@ def validate_features(df: pd.DataFrame, *, feature_names: tuple[str, ...]) -> No
def out_of_contract_row_mask(df: pd.DataFrame, *, feature_names: tuple[str, ...]) -> pd.Series:
"""Boolean mask of rows whose feature values break the trained contract.
- Flags +/-inf (a feature-engineering bug — division returning inf) and
- values outside `FEATURE_RANGES` bounds. NaN passes — XGBoost handles it
- natively, and our SQL deliberately produces NULL (→ NaN) when
- denominators are zero.
+ Flags +/-inf (a feature-engineering bug — division returning inf), values
+ that overflow float32, and values outside `FEATURE_RANGES` bounds. NaN
+ passes — XGBoost handles it natively, and our SQL deliberately produces
+ NULL (→ NaN) when denominators are zero.
Replay events are client-controlled, so a single crafted session can
produce an out-of-range derived feature. Flagging per-row lets callers
@@ -262,8 +265,8 @@ def out_of_contract_row_mask(df: pd.DataFrame, *, feature_names: tuple[str, ...]
series: pd.Series = df.loc[:, name]
spec = FEATURE_RANGES[name]
if series.dtype.kind == "f":
- arr = series.to_numpy()
- mask |= ~(np.isfinite(arr) | np.isnan(arr))
+ # NaN compares False, so it passes; +/-inf compares True.
+ mask |= np.abs(series.to_numpy()) > _FLOAT32_MAX
if spec.min_value is not None:
mask |= series.lt(spec.min_value).fillna(False)
if spec.max_value is not None:
diff --git a/posthog/temporal/tests/session_replay/surfacing_scoring_sweep/test_features.py b/posthog/temporal/tests/session_replay/surfacing_scoring_sweep/test_features.py
index 3aa9b3a721d9..e124786b85cd 100644
--- a/posthog/temporal/tests/session_replay/surfacing_scoring_sweep/test_features.py
+++ b/posthog/temporal/tests/session_replay/surfacing_scoring_sweep/test_features.py
@@ -107,6 +107,9 @@ class TestOutOfContractRowMask:
("selection_copy_count", -1),
("event_rate", float("inf")),
("event_rate", float("-inf")),
+ # Finite in float64 but inf once XGBoost casts to float32.
+ ("event_rate", 1e39),
+ ("mouse_mean_x", -1e39),
],
)
def test_out_of_contract_value_flags_only_its_row(
@@ -118,6 +121,8 @@ class TestOutOfContractRowMask:
) -> None:
good_row = feature_frame.iloc[0].copy()
df = pd.DataFrame([feature_frame.iloc[0], good_row]).reset_index(drop=True)
+ # ClickHouse returns Float64; the float32 fixture would turn 1e39 into inf before the mask sees it.
+ df[column] = df[column].astype(np.float64)
df.loc[0, column] = bad_value
mask = out_of_contract_row_mask(df, feature_names=feature_names_for_tests)