name: type-f32
kind: type
description: IEEE-754 single-precision float; ~7 significant digits.
type: reference
applies_to: inspect, predict
Bytes
Fixed-width: 4 bytes per record, little-endian IEEE-754 binary32. Not bit-packed. Stride contributes 4 bytes; reader uses math.Float32frombits(binary.LittleEndian.Uint32(...)).
Range
±3.4e38 with ~7 decimal significant digits. Right-sized for measurement-grade signals: sensor readings, percentages, ratios, normalized scores. Reach for f64 when you need wider dynamic range or cumulative sums where rounding error compounds.
Null
Orthogonal. Nullable: true participates in the per-record null bitmap. There is no in-band sentinel — NaN is the IEEE NaN value, not null; 0.0 is the float zero, not null. IsNumeric() and IsNumericForAnalytics() are both true; aggregators apply null-skip semantics.
Dictionary
Absent. Floats are never dictionary-encoded — distinct cardinality is effectively unbounded.
See
- Skill:
cohort-schema-design(Field-type matrix, Selection heuristics). - Cross-link:
type-f64for higher precision;type-decimal128for exact money.