Expand description
Arrow-native decision-batch storage (System One P0.3).
Decision model outputs are stored as opaque Arrow IPC stream payloads keyed
by the producing decision-schema fingerprint, so eval/training reads back
the exact RecordBatch the model emitted — no row-wise SQL materialization
and no lossy re-typing. Each row also carries the (schema, model, calib)
version triple for joins against calibration_params.
Structs§
- Decision
Batch Meta - Row metadata for one stored decision batch (everything but the payload).
- Decision
Batch Store - Decision-batch store over any
StorageBackend: the write/read API for Arrow-native model output (eval harness, distillation corpora, re-fit input).
Constants§
- DECISION_
BATCH_ TABLE - Decision batch table name.
- PAYLOAD_
FORMAT_ IPC_ STREAM - Payload format tag for Arrow IPC stream bytes.
Functions§
- decision_
batch_ schema - Arrow schema of the
decision_batchestable. - decode_
batch_ ipc - Decode an Arrow IPC stream payload back into a single batch (multi-batch streams are concatenated).
- encode_
batch_ ipc - Serialize a batch as a self-describing Arrow IPC stream.
- mint_
batch_ id - Mint a new batch id.