ADR 0063: independent native stage limits

• Status: Accepted
• Date: 2026-09-13
• Scope: ClickHouse-to-MSSQL bounded native DDA
• Contract: Stage 2 specification

Context

The encoding process pool and BCP import thread pool have different bottlenecks. Both currently inherit one count. Merely adding optional dataclass fields would also change persisted resource-limit records and closed v1 benchmark admission.

Decision

Add independent keyword-only optional encoding/import counts, each falling back to the existing parallelism. Keep one retained-work map, with capacity max(effective_encoding_parallelism, effective_import_parallelism) + max_pending. The payload bound remains one additional frame beyond that capacity, multiplied by max_bytes. Preserve retry slots, validation and publication authority from ADR 0057. Import concurrency bounds file import/verification tasks, not every independently opened target connection.

The limits model owns a finite canonical serializer. Legacy-effective policies serialize the exact eight historical fields. Other policies include both resolved counts. The fallback parallelism remains significant. Journal equality stays strict; reads do not rewrite old records. New run envelopes use v2 for extended policies; v1 limits are frozen explicitly. Existing correctness receipts, observations, campaigns and comparisons retain their schemas and authority.

Comparisons still require identical full configurations. Independent-policy tuning is a separate diagnostic experiment with separate hashes, not permission to produce a cross-policy comparison PASS. Never normalize historical evidence to manufacture equivalence. Old readers reject new run versions.

Consequences and alternatives

Users can tune CPU and target work separately without changing extraction or creating unbounded queues. Max-based capacity does not guarantee simultaneous saturation of both pools. Sum-based capacity and independent queues were rejected because they enlarge retention and break legacy resource expectations. Automatic tuning, distributed scheduler abstractions and cross-policy benchmark semantics are deferred. Upgrade readers before adopting extended reports; finish or settle extended-policy invocations before downgrading. Keep their original policy for recovery. Unit, spawned-worker and approved Docker evidence must prove bounds, legacy journal/report compatibility and typed delivery before separate release.