Skip to content

ADR-0060 — Canonical replay I/O materialization profile v1

Status: ACCEPTED Date: 2026-10-04

Decision

The selected trade-v1 profile for new replay/research materialization is the existing behavior: Parquet v2.6, zstd compression, 65,536 rows per row group, all canonical trade-v1 columns, and sequential DataGateway.scan() reads with a caller-selected logical batch default of 65,536. The physical reader projects only canonical columns, uses one batch and fragment read-ahead, and disables reader threads so physical scheduling cannot alter order or multiply the logical batch-memory bound.

This profile changes no semantic record, canonical content hash, ordering, or DataGateway contract. Compression and row-group layout are physical artifact choices: they change the exact file hash and require normal publication/ certification evidence for a replaced artifact, while identical ordered rows retain their canonical content identity. Do not change these defaults for a large multi-year materialization without a new measured profile decision.

Evidence and scope

The focused layout test proves zstd/default row grouping and that the same ordered rows are returned through bounded scanning from a different physical layout. It is a fixture-level behavioral proof, not a throughput benchmark; hardware-dependent performance measurements remain an operational follow-up.

No semantic recertification is required here because no artifact writer or fixture artifact changes. A future profile change must create replacement artifacts and recertify their affected publication evidence.

Out of scope

Representation materialization, feature precomputation, replay semantics, source-acquired evidence rules, and changing existing artifacts.

Parent tracking: #232. Closes issue #235.