Module geo Future

Geo Module - Future Enhancements

Scope

• hardening and refinement of geospatial backend/index/query runtime behavior
• expansion of deterministic reliability under mixed CPU/GPU capability pressure
• stronger benchmark-backed guardrails for geo hot paths

Design Constraints

• geo contracts remain backward compatible within major release line.
• geometry validation and backend fallback behavior remain explicit and deterministic.
• advanced query execution remains bounded and observable.
• runtime backend transitions remain auditable.

Required Interfaces

Interface	Requirement
backend interfaces	deterministic CPU/GPU dispatch and fallback semantics
index/geometry interfaces	stable indexing and geometry-validation contracts
advanced query interfaces	bounded join/clustering/raster/temporal behavior
integration interfaces	explicit capability detection and bridge behavior

Implementation Notes (Q3 2026 Status Sync)

Geospatial GPU Backend - CUDA (Completed)

• baseline Haversine distance kernel is present (spherical earth computation, < 0.5% accuracy target); deterministic precision-mode parity hardening complete.
• Vincenty distance kernel is implemented (WGS-84 ellipsoid model, float-kilometre output precision); falls back to Haversine for nearly-antipodal points.
• baseline point-in-polygon containment kernel is present (ray-casting); fallback and diagnostics parity across degraded runtime scenarios remains open.
• device memory lifecycle and host↔device transfer behavior maintain explicit failure signaling and bounded fallback semantics.
• batched per-pair distance dispatch is supported via both Haversine and Vincenty kernels through the GeoDistanceFormula parameter.

Geospatial GPU Backend - HIP (In Progress Parity)

• HIP Haversine and point-in-polygon runtime paths are present; full parity with CUDA diagnostics and mixed-capability behavior remains open.
• hipMemcpy / hipMalloc allocation and transfer paths require continued reliability verification under sustained mixed-load scenarios.
• ROCm optimization remains constrained by deterministic behavior requirements before benchmark re-baselining.

Geospatial GPU Backend - OpenCL Path (Planned v1.5.0+)

• OpenCL kernels for broader GPU compatibility
• Portable kernel compilation pipeline
• Interop with existing CUDA/HIP dispatch layer

Phase-Aligned Hardening Plan (Phase 1-6)

• Phase 1 (Design/API): preserve geo contract and error taxonomy compatibility while tightening deterministic backend-dispatch and precision guarantees.
• Phase 2 (Core): close remaining backend fallback parity gaps across CPU/CUDA/HIP dispatch and advanced query boundedness.
• Phase 3 (Error/Edge): standardize fail-closed geometry validation and backend-switch diagnostics for degraded-capability incidents.
• Phase 4 (Tests): expand regressions for mixed backend + precision permutations and complex join/raster validation edges.
• Phase 5 (Performance): stabilize benchmark envelopes for CPU/GPU geo kernels, indexing, and join hot paths with reproducible p95/p99 guardrails.
• Phase 6 (Documentation/Acceptance): keep roadmap/future/evidence synchronization current and traceable for issue-driven closure.

Test Strategy

• unit and integration suites for backend, indexing, geometry, and advanced query paths.
• regressions for invalid geometry, unsupported features, and fallback transitions.
• deterministic stress runs for mixed CPU/GPU geospatial workloads.
• release-profile benchmark runs for mapped geo targets.

Performance Targets

• CPU/GPU geo hot paths remain within regression budgets.
• spatial index and join workloads remain stable at p95/p99 envelopes.
• benchmark manifests for mapped geo targets reach no-missing-case status.

Security / Reliability

• maintain strict geometry validation before geospatial execution.
• preserve explicit failure signaling for degraded or unsupported paths.
• enforce bounded behavior for advanced geo operations under pressure.
• keep diagnostics actionable for production geo incidents.

CUDA Geospatial Kernels

• Haversine distance kernel delivered (Q3 2026): batched per-pair spherical distance computation.
• Vincenty distance kernel delivered (Q3 2026): WGS-84 ellipsoid model, fallback for nearly-antipodal points.
• Point-in-polygon containment kernel delivered (Q3 2026): ray-casting, batched dispatch.
• Open: full diagnostics parity and mixed-capability fallback hardening (Q4 2026).

Boost.Geometry Integration

• CPU exact backend implemented via boost::geometry: supports union, difference, intersection for polygon inputs.
• Backend registered as boost_cpu_exact in the backend dispatch table.
• Open: expanded regression coverage for complex polygon edge cases (Q4 2026).

---

ThemisDB 1.9.0-beta · Home · Module-Index · GitHub · Issues