StaticComputeSHA-256 of the canonical JSON of each feature column's name, encoding kind and source column, in column order.
StaticComputeThe snapshot's content hash: SHA-256 of the canonical JSON of the matrix and row-manifest file hashes, the feature schema hash, PredictionUnitID, LabelVersion, FeatureSpecVersion, EvaluationProtocolID and DataCutoff. The matrix file holds the header, so the encodings, incumbent and slice columns and decision spec are covered through its hash.
The provenance file is deliberately outside the hash: it records when each value was computed (aggregates, series and missing values carry the assembly time), so including it would give identical data a new hash on every assembly, and identical data must resolve to one snapshot. Its own hash is stored and verified on read.
IDs are normalized, so the hash does not depend on the database's UUID casing. DataCutoff must be a
YYYY-MM-DD day; any other spelling of the same cutoff would hash differently, so it is refused.
StaticParseThe manifest file's rows, each checked with StoreSchemas.SnapshotRowKey (also read alone, without the matrix).
StaticParseThe three files back to content. Every line is checked with its store schema — a stored file is text that a hash
check proves unchanged, not well-formed — and so is the structure across the files: every matrix row is as wide
as the header's columns, the matrix rows, manifest lines and provenance lines are equally many, and provenance
line i carries manifest line i's rowId. A violation throws SnapshotFileFormatError naming the file and line.
StaticSerializeThe content as the three files' bytes. Refuses content whose rows, manifest and provenance do not align, a row whose width differs from the header's columns, and a NaN or infinite numeric cell.
The Training Snapshot file format: serialize, parse and fingerprint.