2.3 KiB
ADR-0005: Component storage before a general ECS
Status
Adopted for proof as bounded experiment; production ECS deferred to M1 measurements.
Context
Composition is locked; taxonomy inheritance is prohibited. The proof needs identity/state, not dynamic queries.
Decision
Use sorted component-composed rows with monotonic non-reused u64 IDs, capped at100,000. Public values expose no third-party handles. lower_bound lookup is O(log n); erase is O(n).
Alternatives Considered
EnTT is MIT, header-only C++ with documented uses. Flecs offers C/C++, relationships, queries/scheduling under MIT. Relative workload performance is unmeasured.
Benefits
Small lifecycle/replay surface, stable order, easy diagnostics.
Costs
Less flexible than archetypes or independent pools; erase moves rows.
Risks
Prototype grows into homemade ECS, ID exhaustion, view invalidation, structural-order drift.
Consequences
Compare churn/query/memory/order/diagnostics on sparse/dense workloads before adoption. Defer structural edits to stable phases. Reject exhausted/stale IDs; generational reuse needs separate design. EnTT, Flecs.
M1.5 measurement update (2026-09-19)
ADR-0021 adds an isolated EnTT4 comparison, with full byte-equivalence, closed-form arithmetic, retired-ID and restore tests. See the protocol and results.
Windows and Linux evidence supports retaining sorted rows for the bounded proof and upcoming scheduler contract, not adopting a general ECS yet. At 10000 entities, EnTT substantially reduces individual removal/replacement cost; the row adapter uses less live container memory and has cheaper updates/export in the Windows workloads. Linux sparse updates favor EnTT (~2.4x), while dense updates/export still favor rows. These are tradeoffs, not a universal ranking. Both compiler families' results and sanitizer evidence are recorded with the experiment; native container growth differs between standard libraries.
The production decision remains deferred until variable component sizes, capability add/remove, high churn over long lifetimes, interacting systems and bounded external-ID mapping are measured. Compare EnTT groups and batched row structural updates then. No public EnTT handles or dependency enter libevo.