name: type-set-u128
kind: type
description: 16-byte multi-select bitmask over an inline dictionary of up to 128 labels.
type: reference
applies_to: inspect, predict
Bytes
Fixed-width: 16 bytes per record, little-endian word order — words[0] holds bits 0..63 and occupies the lowest 8 bytes. Not bit-packed (the value is consumed atomically as a bitmask). Stride contributes 16 bytes. Bit i corresponds to dictionary entry i — bit set means “label i selected”. Too wide for the uint64 scalar value API: encoding.ReadFieldValue / WriteFieldValue REFUSE it with ENCODING_TYPE_MISMATCH rather than silently truncate to the low 64 bits. FieldType.IsWideSet() names that split for callers that must branch; wide masks travel as SetMask.
Range
Up to MaxSetEntries() = 128 distinct labels in the inline dictionary. Empty mask is a valid value meaning “no labels selected” — it is NOT null. Overflow on import → PULSE_IMPORT_SET_OVERFLOW. Inference picks the SMALLEST rung that fits the observed dictionary, so a 65-option column lands here and a 129-option one promotes to set_u256. Width is paid by every record: a 70-label column costs 16 bytes per row whether or not any row sets a high bit.
Null
Orthogonal. Nullable: true participates in the per-record null bitmap; null is signalled there alone. Empty mask is distinct from null. In-band sentinels are explicitly disallowed for set_* types — no mask value is reserved.
Dictionary
Present (HasDictionary() = true). Inline schema block, shared across all records; MaxDictEntries() = 128. Shard cohesion: union-merge under the 128-label cap; width overflow → PULSE_SHARD_DICT_WIDTH_OVERFLOW. The motivating source is an SPSS multiple-DICHOTOMY response set with more than 64 constituents, imported beside its constituent columns (cohort-schema-design, SPSS import).
See
- Skill:
cohort-schema-design(Field-type matrix, Nullability + per-record bitmap, Width overflow). - Cross-link:
type-set-u64(the rung below),type-set-u256,op-agg-set-frequency,op-attr-set-has.