ADR-0008 — Bounded sampling for large and backed matrices

• Status: Accepted
• Date: 2026-07 (as embodied in the 0.2.0b1 codebase)

Context

Full densification or exhaustive scans of large matrices are unsafe for memory and runtime. Backed AnnData must fail gracefully.

Decision

• Never densify a full sparse matrix merely to validate it.
• Numeric / AI-readiness checks use bounded row sampling (sample_rows, default 5000) for large or backed inputs.
• Sparse in-memory paths may inspect .data without densifying.
• Unsupported modes skip with an explicit check execution record rather than crashing the run.

Consequences

• Rare defects outside the sample may be missed (documented limitation).
• Large public profiles remain practical to validate.

Alternatives considered

• Always full-matrix scan — rejected for scalability.
• Skip numeric checks on backed data silently — rejected; skips are recorded.

References

• src/cp_anndata_validator/sampling.py, loading.py
• Matrix / AI-readiness checks and their tests
• Large or backed AnnData