Module distributed tensor Roadmap

Navigation: Home > Modules

Distributed Tensor Module Roadmap

Current Status

EPIC 3 distributed tensor contract ownership and core implementation are complete for sub-issues 3.1–3.7. Phase 3 runtime failure/degraded-mode semantics are implemented and verified. Phase 4 broadened contract and fault-path coverage has been delivered: test_phase4_contract_coverage.cpp adds 58 tests spanning all 7 sub-issues (lifecycle, subclasses, manifest, placement, integrity, recovery, planner, infrastructure). Benchmark evidence and final acceptance documentation remain pending.

Test Infrastructure (2026-07-28 Update): Focused test targets are now properly configured in tests/epic3_distributed_tensor/CMakeLists.txt with:

• 11 unit test targets: module_epic3_distributed_tensor_*_focused
  • distributed_planner_test_focused (contract)
  • manifest_store_phase_a_focused (advisory-only store)
  • lifecycle_staleness_management_focused (lifecycle FSM)
  • tensor_delta_log_focused (Phase A/B)
  • tensor_rebuild_fallback_focused (Phase B)
  • phase3_failure_semantics_focused (error handling)
  • phase4_contract_coverage_focused (broad coverage, 58 tests)
  • tensor_storage_strategy_focused (strategy validation)
  • tensor_training_coordinator_focused (training)
  • tensor_update_worker_focused (update orchestration)
  • integrity_verification_test_focused (verification)
• 1 benchmark target: module_epic3_distributed_tensor_integrity_verification_bench_focused

Tensor-Update Rollout Track: dynamic tensor-update infrastructure is staged across four phases (A–D) in ai_working/HYBRID_RETRIEVAL_ROLLOUT_PLAN.md (issue #5468). Phase A (manifest schema + advisory-only policy) is complete. Phase B (delta log + partial refit) targets Q3 2026. Phase C (shard summary coordination) targets Q4 2026. Phase D (optional acceleration) targets 2027+.

Module readiness: implementation and Phase 3/4 validation are complete for the EPIC 3 core artifact path; benchmark evidence, acceptance documentation, and the follow-on tensor-update track remain gated. Focused test targets now properly configured for CI/evidence collection.

In Progress

• [~] EPIC 3 documentation-governance alignment across roadmap/future/audit files (Target: Q3 2026)
• [~] phase-gate acceptance criteria definition for distributed test/benchmark readiness (Target: Q3 2026)
• [~] Phase 5: Adapter-Distribution & Sharding-Kopplung — LoRA artifact distribution interfaces, RAID/Merkle proof engine, distribution receipts, shard snapshots, and recovery ordering (Issue #5418, Target: Q3 2026)
• [~] tensor artifact manifest schema: ArtifactManifest + ManifestStore wiring (Target: Q3 2026)
• derived-artifact lifecycle and staleness management: LifecycleState FSM, ArtifactClassifier, manifest fields, invalidation manager (issue #5442, Target: Q3 2026)
• phase-4 distributed contract and fault-path coverage expansion beyond the focused regression suite (Target: Q4 2026)

Planned Features

Tensor-Update Infrastructure Track (issue #5468)

Phase A — Exact-First Infrastructure (Q3 2026, start)

• tensor delta log schema (DeltaLogEntry, DeltaWindow, TensorDeltaLog in src/distributed_tensor/include/tensor_delta_log.h; 309 LOC)
• manifest schema committed with advisory-only artifact policy documented (Target: Q3 2026)
• Prometheus freshness metrics wired (tensor_freshness_age_seconds, tensor_delta_log_entries_total) (Target: Q3 2026)
• snapshot-based rebuild worker (src/distributed_tensor/src/snapshot_update_worker.cc; 524 LOC; patch/refit/rebuild thresholds)
• exact graph fallback guarantee documented and tested (src/distributed_tensor/include/exact_graph_fallback.h; tests/epic3_distributed_tensor/test_tensor_rebuild_fallback.cpp)

Phase B — Local Tensor Maintenance (Q3 2026, weeks 6–9)

• patch path for small tensor deltas (bounded delta window) (patch_threshold_pct_ in snapshot_update_worker.cc)
• partial refit for bounded windows (refit_threshold_pct_ in snapshot_update_worker.cc)
• rebuild fallback for unstable or large update sets (fallback rebuild in snapshot_update_worker.cc:139-146)
• planner freshness gates (tensor_freshness_age_seconds threshold checks) (Prometheus metric wired)
• ctest gate: test_tensor_delta_log (tests/epic3_distributed_tensor/test_tensor_delta_log.cpp)
• ctest gate: test_tensor_rebuild_fallback (tests/epic3_distributed_tensor/test_tensor_rebuild_fallback.cpp)
• ctest gate: test_phase_b_edge_cases (tests/epic3_distributed_tensor/test_phase_b_edge_cases.cpp) — Phase B comprehensive edge case testing
• benchmark gate: bench_tensor_partial_refit (benchmarks/epic3_distributed_tensor/bench_tensor_partial_refit.cc) — Performance targets: p99 <= 30ms for bounded windows

Phase B Implementation (2026-08-17 Completion):

• Enhanced patch path with delta window bounds checking (sequence continuity, age validation, structural mutation detection)
• Partial refit with state machine transitions (PARTIAL_REFITTED state, rank cap enforcement, residual thresholds)
• Rebuild fallback triggering on:
  • Delta log overflow detection (95% threshold on max entries)
  • Instability detection heuristics (high mutation frequency, residual thresholds)
  • Fail-closed behavior on manifest validation failure
• Comprehensive edge case test suite (PBE-01..24): 24 tests covering patch bounds, overflow, instability, state transitions
• Full benchmark suite with performance targets:
  • Patch path: p99 <= 5ms (10-50 entry windows)
  • Partial refit: p99 <= 30ms (100-3200 entry windows)
  • Rebuild fallback: p99 <= 100ms
  • Decision logic: < 100µs per call
• Production-ready error handling: manifest validation, rank cap checks, residual enforcement
• State machine transitions for all three paths (PATCHED, PARTIAL_REFITTED, REBUILT)

Phase C — Distributed Tensor Artifact Coordination (Q4 2026) ✅ COMPLETE

• shard summary refresh with freshness timestamps (✅ Implemented 2026-08-17)
• summary-first routing with escalation to exact-on-demand (✅ Implemented 2026-08-17)
• exact-on-demand tensor fetch (replaces summary when accuracy required) (✅ Implemented 2026-08-17)
• multi-shard freshness consensus (✅ Implemented 2026-08-17)
• ctest gate: module_epic3_distributed_tensor_tensor_shard_summary_focused (✅ 41 tests 2026-08-17)
• benchmark gate: module_epic3_distributed_tensor_bench_summary_first_focused (✅ 12 benchmarks 2026-08-17)

Phase D — Optional Acceleration (2027+)

• [~] CPU/GPU break-even validation for tensor operations (Target: 2027)
• [~] bounded GPU refinement after break-even proven (Target: 2027)
• no unconditional GPU dependency on any tensor path (Target: 2027)

Short-term (3-6 months)

• implement phase-3 runtime failure and degraded-mode semantics (Target: Q4 2026)
• deliver phase-4 distributed contract and fault-path tests (Target: Q4 2026)
• establish phase-5 distributed benchmark and hardening baselines (Wave D: bench_distributed_tensor_dedicated_gates.cpp — DT-BM-01..04 delivered 2026-09-16)

Mid-term (6-12 months)

• complete phase-6 acceptance documentation from measured evidence (Target: Q1 2027)
• complete phase-7 default workflow integration after gates pass (Target: Q1 2027)

Implementation Phases

Phase 1: Design / API Contract

• EPIC 3.1 (Artifact Classes) contract ownership mapped and documented
• EPIC 3.2 (Manifest Schema) contract ownership mapped and documented
• EPIC 3.3 (Shard Placement) contract ownership mapped and documented
• EPIC 3.4 (Integrity Model) contract ownership mapped and documented
• tensor-update rollout track staged (Phases A–D) in ai_working/HYBRID_RETRIEVAL_ROLLOUT_PLAN.md
• advisory-only artifact policy documented with invariant: tensor artifacts never replace graph-verified results (Target: Q3 2026)
• derived-artifact lifecycle model: LifecycleState (READY/STALE/INVALIDATED/REBUILDING/FAILED), TruthSemantic, ArtifactClass, RebuildState, UpdateMode, InvalidationReason defined in artifact_manifest.h (issue #5442, Target: Q3 2026)
• manifest lifecycle fields defined: source_seq_start, source_seq_end, delta_lag, artifact_age_ms, residual, rank_cap, lifecycle_state, invalidation_reason (issue #5442, Target: Q3 2026)
• advisory-only enforcement: ArtifactClassifier::isValidCombination prevents DERIVED/EPHEMERAL from acquiring GROUND_TRUTH semantic (issue #5442, Target: Q3 2026)

Phase 2: Core Implementation

• production translation units and public headers implemented
• themis_distributed_tensor wired into the active community build graph
• integrity translation units created for EPIC 3.4
• deterministic JSON hashing and manifest-facing proof/receipt serialization implemented
• cached-receipt versus fresh-verification rules documented for query-path consumers
• 3.4 runtime distributed failure semantics completed across recovery/planner integration
• lifecycle transition helpers (transitionToRebuilding, transitionToReadyAfterRebuild, transitionToFailed) and planner freshness gate helper (shouldRejectForPlanner) implemented in ArtifactInvalidationManager (issue #5442, Target: Q3 2026)
• Phase A: tensor delta log + manifest store wired (Target: Q3 2026) ✅ TDL-01..18 in tests/epic3_distributed_tensor/test_tensor_delta_log.cpp
• Phase A: snapshot-based rebuild worker implemented (Target: Q3 2026) ✅ RFB-01..10 in tests/epic3_distributed_tensor/test_tensor_rebuild_fallback.cpp
• Phase B: patch path + partial refit + rebuild fallback implemented (Target: Q3 2026) ✅ 2026-08-17 complete: tests/epic3_distributed_tensor/test_phase_b_edge_cases.cpp (24 tests), benchmarks/epic3_distributed_tensor/bench_tensor_partial_refit.cc
• Phase C: shard summary refresh + summary-first routing + exact-on-demand fetch implemented (Target: Q4 2026) ✅ 2026-08-17 complete: tests/epic3_distributed_tensor/test_tensor_shard_summary.cpp (41 tests), benchmarks/epic3_distributed_tensor/bench_tensor_summary_first.cc (12 benchmarks)

Phase 3: Error Handling and Edge Cases

• runtime distributed failure semantics implemented and verified
• corruption and tampering detection implemented
• partial-receipt and stale-receipt edge case handling implemented
• recovery coordination hooks implemented (EPIC 3.5 integration ready)
• Phase B: rebuild fallback triggered on delta log overflow and instability (Target: Q3 2026) ✅ 2026-08-17: delta log overflow detection (95% threshold), instability detection heuristics, fail-closed behavior
• Phase C: exact-on-demand fallback when summary freshness threshold exceeded (Target: Q4 2026) ✅ 2026-08-17: multi-shard freshness consensus, exact-on-demand escalation logic

Phase 4: Tests

• focused Phase 3 regression suite implemented in tests/epic3_distributed_tensor/
• broaden contract and fault-path coverage across all EPIC 3 components — test_phase4_contract_coverage.cpp (58 tests, sub-issues 3.1–3.7)
• unit tests for SHA-256 computation and validation (40+ test cases)
• Merkle proof verification tests (happy path, invalid, tampering)
• receipt chain verification tests (genesis, appends, tampering)
• integration tests for full verification workflow
• integrity verification unit and benchmark sources added for follow-on coverage expansion
• lifecycle & staleness management unit tests — test_lifecycle_staleness_management.cpp (31 tests, LSM-01..LSM-31, issue #5442)
• test_tensor_delta_log — Phase A ctest gate ✅ Done; TDL-01..TDL-18, TDL-A1..A3 (tests/epic3_distributed_tensor/test_tensor_delta_log.cpp)
• test_tensor_rebuild_fallback — Phase A/B ctest gate ✅ Done; RFB-01..RFB-10, RFB-A1..A5 (tests/epic3_distributed_tensor/test_tensor_rebuild_fallback.cpp)
• test_phase_b_edge_cases — Phase B edge case ctest gate ✅ Done; PBE-01..PBE-24 (24 tests, tests/epic3_distributed_tensor/test_phase_b_edge_cases.cpp)
• test_tensor_shard_summary — Phase C ctest gate ✅ Done; 41 tests (tests/epic3_distributed_tensor/test_tensor_shard_summary.cpp), 2026-08-17 completion

Phase 5: Performance and Hardening

• [~] Adapter-Distribution & Sharding-Kopplung: LoRA artifact distribution APIs, RAID/Merkle proof engine, distribution receipts, shard snapshots, recovery ordering (Issue #5418)
• benchmark suite implemented in benchmarks/epic3_distributed_tensor/
• benchmark suite implemented in tests/epic3_distributed_tensor/integrity_verification_bench.cc
• [~] performance optimization with hardware acceleration (SHA-NI, AVX-512)
• alignment with graph provenance performance targets
• bench_tensor_partial_refit — Phase B benchmark gate ✅ Done; 8 benchmarks with performance targets: p99 <= 30ms for 100-3200 entry windows, 2026-08-17 completion
• bench_tensor_summary_first — Phase C benchmark gate ✅ Done; 12 benchmarks validating summary-first routing and exact-on-demand escalation, 2026-08-17 completion

Phase 6: Documentation and Acceptance

• core distributed tensor docs aligned to source-verifiable implementation state
• tensor-update rollout track documented in ai_working/HYBRID_RETRIEVAL_ROLLOUT_PLAN.md (issue #5468)
• [~] acceptance docs tied to measured runtime behavior ⏳ Phase 6 Acceptance Checklist updated 2026-08-17; benchmark evidence collection pending

Phase 7: Production Integration

• default pipeline integration enabled after gates are met
• [~] Phase A production enable: manifest schema + advisory policy validation gates pending (Target: Q3 2026)
• [~] Phase B production enable: rebuild fallback & instability detection gates pending (Target: Q3 2026)
• [~] Phase C production enable: shard summary coordination gates pending (after sharding Phase C gates at 70% ready) (Target: Q4 2026)

Production Readiness Checklist

• contract ownership and scope boundaries documented
• security and performance expectations documented for scaffold phase
• tensor-update infrastructure rollout track defined (issue #5468)
• runtime resilience hardening completed for Phase 3 safety gates
• test gates implemented and passing for Phase 3 and Phase 4 regression suites
• [~] benchmark gates implemented (sources complete); evidence collection pending (Target: Q4 2026)
• [~] Phase A advisory-only invariant enforced and verified; production enable gate pending (Target: Q3 2026)
• [~] Phase B ctest gates (test_tensor_delta_log, test_tensor_rebuild_fallback, test_phase_b_edge_cases) passing; production enable gate pending (Target: Q3 2026)
• [~] Phase C ctest gates (test_tensor_shard_summary) passing; production enable gate pending (Target: Q4 2026)
• [~] Phase C benchmark gates (bench_tensor_summary_first) implemented; evidence collection pending (Target: Q4 2026)

Test Evidence & Acceptance (2026-07-28)

Configured Focused Test Targets

All test sources in tests/epic3_distributed_tensor/ are now properly registered as focused test targets in CMakeLists.txt with explicit MODULE=epic3_distributed_tensor namespace:

Unit Tests (TIER=unit, TIMEOUT=120s)

• ✅ module_epic3_distributed_tensor_distributed_planner_test_focused — Planner contract (placement, routing)
• ✅ module_epic3_distributed_tensor_manifest_store_phase_a_focused — ManifestStore (MS-01..12)
• ✅ module_epic3_distributed_tensor_lifecycle_staleness_management_focused — Lifecycle FSM (LSM-01..31)
• ✅ module_epic3_distributed_tensor_tensor_delta_log_focused — DeltaLog schema (TDL-01..18)
• ✅ module_epic3_distributed_tensor_tensor_rebuild_fallback_focused — Rebuild fallback (RFB-01..10)
• ✅ module_epic3_distributed_tensor_phase3_failure_semantics_focused — Failure handling
• ✅ module_epic3_distributed_tensor_phase4_contract_coverage_focused — Contract coverage (58 tests)
• ✅ module_epic3_distributed_tensor_tensor_storage_strategy_focused — Storage strategy
• ✅ module_epic3_distributed_tensor_tensor_training_coordinator_focused — Training coordination
• ✅ module_epic3_distributed_tensor_tensor_update_worker_focused — Update worker orchestration
• ✅ module_epic3_distributed_tensor_integrity_verification_test_focused — Verification contract

Benchmarks (TIER=benchmark, TIMEOUT=300s)

• ✅ module_epic3_distributed_tensor_integrity_verification_bench_focused — Integrity benchmark (IV-BENCH-01..08)

Evidence Collection Status

• Test targets properly configured in tests/epic3_distributed_tensor/CMakeLists.txt (2026-07-28)
• Test registration via themis_register_module_focused_test with MODULE=epic3_distributed_tensor
• Build verification: Tests compile on windows-release preset
• Execution verification: All focused tests execute and pass
• Performance evidence: Benchmark baseline captured
• Coverage verification: All sub-issues (3.1–3.7) exercise in tests

Closure Path (for issue #5640)

1. ✅ Roadmap priorities validated against ROADMAP.md source
2. ✅ Future focus points validated against FUTURE_ENHANCEMENTS.md source
3. ✅ Test configuration refined: focused targets now properly discoverable
4. ⏳ Evidence collection: Pending CI run on windows-release preset
5. ⏳ Acceptance criteria verification: Pending test execution results
6. ⏳ Issue closure: Pending evidence documentation

Known Issues & Limitations

• Benchmark evidence and final acceptance documentation are still pending before default pipeline integration can be enabled.
• Phase A/B/C tensor-update rollout items remain incomplete and must preserve the advisory-only artifact invariant until their gates pass.
• Phase C depends on sharding reaching 70% gap reduction — currently 35% ready (🔴).
• Phase D (optional GPU) requires gpu reaching 85% error-handling gap reduction — currently 35% ready (🔴).
• Advisory artifact invariant: tensor summaries and shard summaries are never authoritative truth; Graph Truth Layer remains CPU-first and exact in all phases.

Breaking Changes

• No roadmap-level breaking change planned. Phase transitions are additive and gated.
• Any change to the ArtifactManifest schema must be backward-compatible or versioned.

Program Execution Model — Wave Context

This module is a contributing module in the program-level Wave A → B → C → D execution model. It does not own a primary wave deliverable but must remain release_critical-green throughout all waves and must deliver Wave D operability improvements in Q1 2027. See ../../ROADMAP.md for the full wave model and exit criteria.

Wave D Contribution for distributed_tensor

• Deliver or validate distributed tracing, high-cardinality stress coverage, exporter reliability, and operator remediation hints as applicable to this module (Wave D: test_distributed_tensor_highcardinality_stress.cpp — 3 cases + RUNBOOK_DISTRIBUTED_TENSOR.md delivered 2026-09-16)
• Contribute to or validate long-duration soak test coverage for this module's primary paths (Wave D: test_distributed_tensor_soak.cpp — DistTensorSoak_ShardedOpThroughput, DistTensorSoak_SyncStability, DistTensorSoak_PartitionReliability delivered 2026-09-16)
• Ensure runbook coverage for operator-critical scenarios in this module (Wave D: RUNBOOK_DISTRIBUTED_TENSOR.md — 5 scenarios with log patterns delivered 2026-09-16)

Cross-Wave Requirements

• release_critical CI must remain green on develop throughout all waves (Target: ongoing)
• p95/p99 benchmarks must be refreshed on representative hardware before Wave D sign-off (Target: Q1 2027)
• No behavioral regression may be introduced into modules in Wave A/B/C scope from changes in this module.

Program-Level Success Criteria (contribution)

• This module's distributed/acceleration paths fail closed (Target: Q1 2027)
• Benchmark-backed p95/p99 baselines exist on representative hardware (Target: Q1 2027)
• Operator-critical paths have diagnostics, alerts, and runbooks (Wave D: RUNBOOK_DISTRIBUTED_TENSOR.md delivered 2026-09-16)

---

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