Skip to content

Forward roadmap — genuinely deferred

This is an internal, contributor-facing page, not a headline doc. The authoritative, operation-by-operation status lives in the capability matrix; every shipped capability has a deep-dive under Query Surfaces or Analytics & Distribution, and the authoritative CONCEPT:EG-* definitions are in concepts. The historical "Universal-DB Program" backlog (EG-045..345 — SQL/SPARQL/OWL/Cypher/GraphQL parity, multi-wire adapters, broker, observability, GIS, tensors, streams, KV-cache, agent-memory, the LTAP lakehouse, real ANN pushdown, GDS, PL/pgSQL, SQLite .db I/O, raster pyramids, the numeric kernel + Surface-B analytics UDFs, Calvin, ROS2/DDS) is shipped — see the capability matrix and CHANGELOG.md.

The North Star: Seamless page tracks the cross-modal seam backlog specifically — every write→read path that crosses modalities, per surface (RPC, SQL wires, SPARQL, GraphQL). The in-transaction cross-modal seam is done over RPC (EG-359..363) and pgwire (EG-KG.txn.isolation-ryow-begin-set); mysql/mssql wire parity, GraphQL cross-modal, and in-txn tsdb read-your-own-writes are the open rows there.

What remains is a short list of genuinely-deferred items. Each is folded, as a note, into the deep-dive that owns it.

Item Status Owning deep-dive
Admin console UI — browser surface for tenants/shards/RBAC/backup-PITR (the engine exposes the APIs; the UI is unbuilt) 🗺 Operations Runbook
Live dashboards UI — a Grafana-style front-end over the shipped PromQL/logs/traces query APIs (the query side ships; the UI does not) 🗺 Observability
GPU offload beyond distance/elementwise — reasoning / ANN-build kernels on the GPU (the distance + elementwise CUDA kernels ship and auto-validate on any GPU host) 🔶 Distribution / Robotics / GPU
Calvin multi-node epoch routing — the RESTART-EPOCH ROUTING LOGIC ships (CrossShardCoordinator::acquire_ollp_with_restart_epoch + epoch_for_restart/epoch_packed_seq/EpochFanInRegistry, src/raft/cross_shard_txn.rs, gated calvin/harness; commit 88c9f34): a restarted OLLP txn is routed into a specific, deterministic epoch of the multi-node fan-in (attempt N → base_epoch + (N-1)), packed into a cross-epoch-comparable GlobalSeq, so every node re-deriving the same exchanged per-epoch batches lands the restart at the identical epoch + seq. Proven by pure-logic tests (eg349_*) + the live multi-node harness calvin_ollp_epoch_routing_restart_agrees_across_nodes. Remaining (genuine deferral, not the routing itself): the real cross-node GOSSIP/REPLICATION TRANSPORT that populates each node's EpochFanInRegistry with every peer's per-epoch batch before an epoch closes — participants are local groups on the coordinator node today; a real cross-host soak of that transport is the open step. 🔶 (routing ✅; cross-node transport deferred) Distribution / Robotics / GPU
Numeric Surface-B — graph/timeseries unification — bringing graph-algo + timeseries ops under the eg-numeric kernel via native Method surfaces (the vector/stat/linalg UDFs + cross-modal analytics ship) 🗺 Analytics Program · Numeric Kernel
Numeric migration P2/P3/P5 — the agent-utilities-side swap of its 598 numpy sites to the xp shim and the eventual numpy/scipy drop (the kernel wheel now publishes to PyPI, so this is a downstream dependency swap) 🗺 Numeric Kernel

Legend: 🔶 in-progress · 🗺 designed, not started. | Item | Status | Notes | |------|:------:|-------| | ROS2 real DDS/RTPS wire | ✅ | The rosbridge-WebSocket JSON bridge ships (EG-KG.domains.robotics-gpu-distribution — engine CDC ↔ ROS2 topics via rosbridge_server, no DDS C stack). EG-KG.ingest.dds-transport adds the transport SEAM (src/server/dds.rs, the DdsTransport trait unifying the WS bridge and TWO native DDS legs behind ONE interface): a real, CI-buildable native DDS/RTPS leg behind the ros2-dds feature via the pure-Rust rustdds crate (native Rust DDS + RTPS — mio/pnet/speedy/cdr-encoding, no CycloneDDS/rmw/C toolchain), and — S5 — the CycloneDDS-C-backed rmw leg behind the ros2-rmw feature via the safe cyclonedds Rust crate (vendored C sources, cmake-built, no libclang/network at build time), the REAL rmw_cyclonedds stack for genuine zero-config live-ros2 interop. Both legs are exercised by a real RTPS loopback pub/sub test and share the SAME rmw topic-name/type-name mangling (EG-KG.ingest.rmw-topic-prefix). Kept OUT of the main build — only the opt-in full-extras layer (the default/full build links no rustdds/cyclonedds). Residual: the ROS 2 Iron/Jazzy type-hash (RIHS01_…) typesupport descriptor is deferred on both legs — Humble matches by mangled type name only, which both satisfy. | | Admin console UI | 🗺 | A browser admin surface (tenants, shards, RBAC, backup/PITR). The engine exposes the APIs; the UI is unbuilt. | | Live dashboards UI | 🗺 | A Grafana-style dashboard front-end over the PromQL/logs/traces query APIs (EG-172/302/162/163). The query side ships; the UI does not. | | Python LMCache connector (driver) | ✅ | Shipped: epistemic_graph.kvcache (CONCEPT:EG-KG.backend.shipped-pip-installable-python) — a pip-installable vLLM/LMCache remote-backend driver for the EG-KG.backend.is-configured-so-co KV-cache endpoint. RemoteKVConnector (get/put/contains/exists/stats over /kv/<hash> + /kv/stats, bearer-token) + RemoteKVL2Connector (LMCache native_plugin L2 adapter). Stdlib-only import; httpx behind the [lmcache] extra. | | CUDA distance/tensor kernel — live GPU validation | 🔶 | The GPU dispatch seam + real cudarc CUDA backend ship (EG-KG.compute.gpu-distance-seam/3.6): the CUDA-C kernels compile via NVRTC and launch on a device, with a CPU fallback. The kernels are correctness-matched to the CPU ground truth but await validation on real GPU hardware (none in CI); reasoning/ANN-build GPU offload beyond distance/elementwise remains open. | | Full Iceberg Avro manifest | ✅ | The LTAP lakehouse tier (EG-KG.storage.lsn-as-snapshot-returns) ships a real spec-compliant Iceberg v2 Avro manifest + manifest-list writer (EG-KG.storage.eg-iceberg-avro-manifest/EG-KG.storage.iceberg-manifest-list, crates/eg-lake/src/iceberg_avro.rs, behind lake via pure-Rust apache-avro) — a committed snapshot's metadata.json now references real Avro that Spark/Trino/DuckDB follow. Per-column stats (column_sizes/value_counts/null_value_counts/nan_value_counts/lower_bounds/upper_bounds, keyed by field-id) are gathered at materialize time and emitted so external readers do predicate pushdown / file skipping (EG-KG.storage.iceberg-avro-manifest-carries). Partition field_summary is null by design — the spec is unpartitioned (zero partition fields). | | CUDA distance/tensor kernel — live GPU validation | ✅ | The GPU dispatch seam + real cudarc CUDA backend ship (EG-KG.compute.gpu-distance-seam/EG-KG.backend.real-cuda-tensor-backend): the CUDA-C kernels compile via NVRTC and launch on a device, with a CPU fallback. EG-KG.compute.tensor-gpu-distance closes the validation gap with a GPU-gated parity test in each crate (eg-ann distance + eg-tensor elementwise) that detects a device at runtime and, when one is present, asserts the real CUDA kernel == the CPU ground truth (bitwise-close — f32 relative tolerance for the accumulating distance kernel, exact for the single-op f64 elementwise kernel), else SKIPS cleanly. So the kernels now auto-validate wherever a GPU exists (e.g. the GB10 box) while the test remains a no-op in GPU-less CI. Documented remaining: reasoning/ANN-build GPU offload beyond distance/elementwise. | | Full Iceberg Avro manifest | ✅ | The LTAP lakehouse tier (EG-KG.storage.lsn-as-snapshot-returns) ships a real spec-compliant Iceberg v2 Avro manifest + manifest-list writer (EG-KG.storage.eg-iceberg-avro-manifest/EG-KG.storage.iceberg-manifest-list, crates/eg-lake/src/iceberg_avro.rs, behind lake via pure-Rust apache-avro) — a committed snapshot's metadata.json now references real Avro that Spark/Trino/DuckDB follow. Documented omission: per-column stats + partition field-summary bounds (eg-lake tracks per-file counts only, unpartitioned spec). | | Raster tile pyramids | 🗺 | Raster tile pyramids beyond the shipped vector-tile (MVT) + Shapefile/KML/GeoParquet GIS I/O (EG-KG.domains.map-tiles/306). | | PL/pgSQL procedural bodies | ✅ | CREATE FUNCTION … LANGUAGE plpgsql bodies execute (EG-KG.query.eg-validate-procedural-body/EG-KG.query.concept-7, crates/eg-query/src/sql/plpgsql.rs, folded into the sql feature — pure interpreter, no new deps). A hand-written statement interpreter runs DECLARE vars, BEGIN..END, :=, IF/ELSIF/ELSE, LOOP/WHILE/FOR … LOOP (integer range), EXIT/CONTINUE, RETURN, RAISE, and SELECT … INTO var over a variable environment; a bare SELECT fn(args) / CALL proc(args) triggers it and every embedded SQL/expression runs back through the existing read path. Documented out of scope: set-returning RETURN NEXT/QUERY, FOR row IN <query>, cursors, composite/%ROWTYPE vars, exception handlers, dynamic EXECUTE, and DML-in-body (read path). | | Raster tile pyramids | ✅ | Shipped: the raster complement to EG-KG.domains.map-tiles's vector-tile (MVT) server (EG-KG.domains.raster-build/EG-KG.domains.raster-fetch, crates/eg-geo/src/raster.rs). A Raster (georeferenced Web-Mercator coverage grid: bbox + width×height × bands, serde-persistable like a Geometry) resamples per XYZ tile — Raster::tile(z,x,y) fetches a 256² RasterTile (nearest-neighbour, nodata outside), Raster::build_pyramid(z_min,z_max) emits the full pyramid with per-zoom tile counts. Includes a hand-rolled, dependency-free PNG codec (to_png/decode_png: stored-DEFLATE + hand-computed CRC-32/Adler-32, NO image/png/flate2), so eg-geo adds zero deps and the no-new-C-deps invariant holds (cargo tree links no image/png/flate2). Raw band tiles via RasterTile::data. Extends EG-KG.domains.map-tiles/306. | | PL/pgSQL procedural bodies | 🗺 | Stored-procedure/function procedural execution (loops, variables, control flow) beyond the shipped SQL views + DDL. | | Memory → weights distillation | ✅ | Distilling consolidated agent-memory (EG-KG.compute.hierarchical-summary-tier-eg/221) into model weights (a fine-tune/LoRA export), beyond the retrieval-time context assembly (EG-195), now SHIPS — on the agent-utilities side (CONCEPT:AU-KG.memory.memory-weights-distillation-export): MemoryWeightsDistiller reads the consolidated/procedural memory tiers and renders SFT/DPO corpora, plus a GRPO renderer for EG-099 :Trajectory/:Step reward sequences, then hands off via graph_analyze action=distill_memory to the train-model workflow, which runs the LoRA/SFT/GRPO fine-tune in agents/data-science-mcp (GPU-gated) and deploys the checkpoint through the model-registry role-bind seam, optionally hot-loading it onto a live vLLM/SGLang server (POST /v1/load_lora_adapter). See Agent Memory. | | Calvin distributed read-lock (OLLP) | ✅ | The Calvin global-sequencer total order + deterministic vote-free execution + crash-replay recovery ship (EG-KG.txn.calvin-deterministic-ordering); the distributed OLLP reconnaissance/read-lock phase now SHIPS (EG-KG.txn.concept-2): an OLLP reconnaissance read predicts a data-dependent txn's read/write set and a deterministic OrderedLockManager (keyed by the GlobalSeq total order) acquires read+write locks in sequence order, giving full serializable isolation of conflicting sequenced txns (the deterministic-execution phase then runs lock-free); and the multi-node sequencer epoch fan-in ships too (EG-KG.txn.concept-3: epoch_fan_in deterministically merges each node's per-epoch sequenced inputs into one identical global order per epoch). src/raft/cross_shard_txn.rs, gated calvin/harness (out of full); proven live in xshard_harness. The OLLP recon-staleness RESTART now ships too (EG-KG.txn.concept-4): when recon_still_valid is false under the held ordered locks, acquire_ollp_with_restart re-reconnoiters and re-submits the txn at a fresh GlobalSeq bounded by max_restarts, deterministic within the single-active-sequencer ordering domain (proven by calvin_ollp_stale_recon_is_restarted_and_commits_serializably). Remaining open (EG-KG.txn.concept-3 scope, not a correctness gap here): routing a restarted txn into a specific epoch of the multi-node fan-in. | | SQLite .db file I/O | ✅ | The SQLite-dialect NDJSON-over-TCP wire ships (EG-KG.query.concept-3); reading/writing an on-disk sqlite3 .db file now SHIPS too (EG-KG.query.eg-feature import / EG-KG.query.full-protocol export — Method::ImportSqliteFile/ExportSqliteFile over the user-table store). Behind the sqlite-file feature: it pulls rusqlite with the bundled C sqlite3 (no pure-Rust SQLite-file writer exists), it is part of the one main build (rusqlite bundles its C sqlite3, so no external toolchain is needed). |

Shipped in Program B — remaining engine tail (EG-3.x)

The last engine features of Program B, closing the distribution / robotics / GPU tail. Each lands with tests + a docs/concepts.md entry and is feature-gated OUT of the Pi tier (the heavy deps stay optional).

Distribution

  • Cross-region async read-replica tier — a bounded monotone-LSN replication log the primary serves over /replicate?since=<lsn> + an async follower pull-loop that applies the tail via the canonical wal::apply path, so a distant region gets a local eventually-consistent read copy with bounded staleness (EG-322). Beyond the synchronous multi-Raft groups + EG-KG.ontology.federation-client federated read.
  • Capacity guardrails — a per-target circuit breaker (Closed→Open→HalfOpen), a per-tenant hard concurrency quota, and global-high-water backpressure fronting the replica/transport paths (EG-323). Complements the EG-320 QoS scheduler with absolute ceilings + fail-fast.
  • Full Calvin deterministic-ordering commit — a global sequencer total order + Raft-replicated input log + vote-free deterministic execution + crash-replay recovery, a third cross-shard commit branch opt-in alongside 2PC + Paxos-Commit-lite (EG-KG.txn.calvin-deterministic-ordering). (Distributed OLLP read-lock phase + multi-node sequencer fan-in remain — see forward roadmap.)

Robotics

  • ROS2 bridge over rosbridge-WebSocket — engine CDC events ↔ ROS2 topics via the standard rosbridge_suite JSON-over-WebSocket protocol to a rosbridge_server (advertise/subscribe/publish/ unsubscribe), with NO CycloneDDS/rmw/DDS C stack — a pure-Rust tokio-tungstenite client (EG-KG.domains.robotics-gpu-distribution).
  • Native DDS/RTPS transport seam + leg — the DdsTransport trait (src/server/dds.rs, EG-KG.ingest.dds-transport) puts the WS bridge and a native DDS/RTPS transport behind ONE interface, so the CDC↔ROS2 path targets EITHER (both put the identical std_msgs/String payload on the wire via the shared EG-KG.domains.robotics-gpu-distribution shaping). The native leg (ros2-dds feature) speaks real RTPS via the pure-Rust rustdds crate — NO CycloneDDS/rmw/C toolchain, so it actually builds in CI and is covered by a real RTPS loopback pub/sub test. Kept OUT of the main build (only the opt-in full-extras layer). The CycloneDDS-C rmw leg stays a toolchain-gated future option.

GPU

  • GPU-accelerated distance/tensor — a DistanceBackend/TensorBackend dispatch seam (EG-KG.compute.gpu-distance-seam) with the pure-Rust CPU backend as the compiled-in fallback, plus a real cudarc-backed CUDA backend (EG-KG.backend.real-cuda-tensor-backend) that NVRTC-compiles the batch-distance/elementwise kernel and launches it on a device, else falls back to CPU. cudarc is dynamic-loading, so the leg builds with no CUDA toolkit and a pi build links none of it. EG-KG.compute.tensor-gpu-distance adds a GPU-gated parity test (compiled under gpu-cuda) that asserts the real CUDA kernel == the CPU ground truth whenever a device is present and SKIPs cleanly otherwise — so the kernels auto-validate on any GPU host (e.g. GB10) without breaking GPU-less CI.

Shipped in Program B

Program B (waves B-1..B-6, CONCEPT:EG-KG.query.gds-call-procedures..320) closed the deferred tail. Every item below is now ✅ — see the capability matrix for the per-operation evidence and concepts for the authoritative CONCEPT:EG-* definitions.

SQL / tables

  • ALTER TABLE beyond ADD COLUMN — DROP COLUMN, RENAME COLUMN, RENAME TO, ALTER COLUMN TYPE (with data migration), DROP CONSTRAINT on the durable user-table catalog (EG-KG.query.rename-table-moves-catalog).
  • Real ParadeDB BM25 — real relevance scoring + highlighted snippets behind @@@ / paradedb.score() / snippet() (EG-311, replacing the placeholder 1.0).
  • Real pgvector ANN pushdown — ORDER BY col <-> $1 LIMIT k pushes to the eg-ann HNSW/IVF index with an exact re-rank (EG-KG.query.real-pgvector-ann-top).

Lakehouse interop (LTAP)

  • eg-lake LTAP tier — Parquet-on-object-store materialization + Delta transaction log + Iceberg-REST catalog + real Iceberg v2 Avro manifest/manifest-list writer (EG-KG.storage.eg-iceberg-avro-manifest/EG-KG.storage.iceberg-manifest-list) + LSN-style as-of snapshots, so external lakehouse engines (Databricks/Spark/Trino/DuckDB) read the engine's tables with zero ETL (EG-KG.storage.lsn-as-snapshot-returns).

Graph / Cypher

  • GDS over Cypher — CALL gds.<algo>(…) YIELD … (PageRank/Louvain/WCC/SCC/betweenness/Dijkstra/similarity) over eg-compute (EG-298).

Vector / ANN

  • HNSW index — hierarchical-navigable-small-world graph index for higher recall-per-probe than IVF-PQ, with insert/search/serde-persist (EG-KG.retrieval.hnsw-vector-index).
  • Cross-shard kNN scatter-gather — a kNN query scatters across per-shard eg-ann indexes and merges to a deterministic global top-k via the merge_topk leaf (EG-319, completing EG-KG.retrieval.scatter-gather).

Broker / streams

  • Exactly-once broker delivery — idempotent-producer dedup (producer id + sequence) + the stream/confirm/ack ops exposed over the AMQP confirm.select + MQTT 5 wire frames (EG-314).
  • Live CEP subscription — register a CEP pattern, subscribe, receive pushed matches fed by the CDC bus (EG-KG.query.protocol-types).
  • Redis pub/sub + S3 multipart — RESP SUBSCRIBE/PSUBSCRIBE/PUBLISH + MULTI/EXEC (EG-KG.ontology.resp2-resp3-codec-round) and S3 multipart upload + range GET (EG-KG.ontology.object-put-get-head), both EG-307.

RDF / OBDA

  • ICV write-path enforcement — a commit guard rejects a transaction that would introduce a SHACL-as-constraint violation, configurable enforce/warn (EG-KG.ontology.rdf-update-guard).
  • Full R2RML Turtle parse — standard R2RML mapping documents drive an OBDA virtual graph (EG-305).

Observability

  • PromQL extended function set — _over_time family, delta/idelta/deriv, topk/bottomk/quantile, label_replace/label_join, clamp* (EG-302).
  • OTel export + Prometheus remote-write + OTLP — the engine emits its own metrics/traces to an external OTel collector + accepts a Prometheus remote-write receiver (EG-316).

GIS

  • Shapefile / KML / GeoParquet I/O — ESRI Shapefile, KML/KMZ, and GeoParquet reader/writer round-tripping eg-geo geometries + attributes (EG-KG.domains.geo-formats).
  • Routing turn-restrictions + time-windows — turn-restriction penalties + time-dependent edge weights on the EG-KG.domains.geo-routing router (EG-KG.domains.geo-partitioning).

KV-cache

  • Real KV warm-tier compression — zstd (optional lz4) replacing the RLE fallback (EG-315).

Durability / wire surface / scheduling

  • Durable RBAC persistence — roles/grants + agent identities persist to redb + reload at boot (EG-KG.compute.durable-rbac-identity-persistence).
  • Tensor-op CAS write-back — derived tensors persist into the content-addressed tensor store on the exec path (EG-304).
  • Durable JSONPath index — the inverted JSONPath index persists to redb + rehydrates at boot + feeds planner cost Stats (EG-308).
  • Federated typed result fusion — schema-aware SQL + SPARQL column-union + typed dedup/merge (EG-KG.query.schema-typed-fusion-sql).
  • Memory/scene/trajectory wire-Ops — the agent-memory + scene-graph + trajectory library APIs exposed over the wire as additive Methods + dispatch + WAL replay (EG-KG.memory.eg-batch-decay-caller).
  • Real-time QoS/SLO scheduler — per-tenant/priority admission + deadline scheduling + backpressure (EG-320).

Analytics program — numeric kernel (one kernel, two surfaces)

  • P1 (this release) — numeric kernel + Python shim — crates/eg-numeric (faer + ndarray, BLAS/LAPACK-free): the curated numpy op-surface (reductions/stats, element-wise, the linalg-6 + LinAlgError, random) as a pure Rust rlib, exposed on Surface A (Python extension epistemic_graph.numeric, feature python) and consumed by agent_utilities.numeric.xp (AU-KG.backend.lmcache-native-connector); gated behind a numeric cargo feature (in the one main build); parity-proven np.allclose vs numpy incl. nan/inf/singular/empty edge cases (AU-KG.compute.numeric-kernel). Done.
  • P2–P3 — migrate the 598 numpy sites — swap the 32 light-op files then the 6 linalg files in agent-utilities from numpy to the xp shim. ✅ Shipped (agent-utilities 1.2.0/1.3.0, KG-2.313): 34 files migrated, zero numeric regressions; numpy now survives only in the [test] extra as the np.allclose parity reference.
  • P4 — Surface B engine operators — DataFusion SQL UDFs/UDAFs over the kernel + kernel-backed batch vector ops. In progress: cosine_sim/l2_normalize/zscore scalar UDFs + covariance UDAF registered on the SQL surface (SELECT zscore(price) FROM …, SELECT cosine_sim(a.emb, b.emb) … run in-engine), plus the BatchL2Normalize engine Method (EG-KG.query.surface-b-numeric-operators/EG-KG.compute.l2-normalize-batch-vectors), all behind the numeric feature (in the one main build). Also shipped: the svd(vec_col)/pca(vec_col,k) column→matrix UDAFs (EG-KG.query.svd-eg-pca-column/EG-KG.query.concept-6 — a row-buffering accumulator marshals a column of vectors into a dense ndarray::Array2, then faer svdvals/eigh), plus kmeans(vec_col,k) (EG-KG.query.kmeans-clustering-half-one — the same column→matrix accumulator driving a new pure-Rust eg-numeric::cluster Lloyd + k-means++ kernel, no linfa/BLAS; one List<Int64> cluster label per row). The differentiator — shipped (EG-KG.query.eg-3): cross-modal join → analytics in-engine — one SQL statement joins graph ⋈ vector ⋈ timeseries and runs pca/kmeans/covariance over the joined result set (impossible in numpy — no data layer), E2E-proven in crates/eg-query/tests/cross_modal_analytics.rs. Deferred (next P4 increment): graph-algo/timeseries unification under the kernel via native Method surfaces. ▶ ✅
  • P5 — drop numpy/scipy from agent-utilities; the kernel is the dep, shipped as epistemic-graph[numeric] (ONE package). 🗺 The publish blocker is cleared (EG-KG.compute.tensor-gpu-distance): release-build.yml builds the pyo3 Surface-A kernel and folds its .so into the epistemic-graph node wheel as epistemic_graph.numeric (scripts/inject_numeric_kernel.py) — no separate eg-numeric package on PyPI. So downstreams hard-depend on epistemic-graph[numeric] and the numpy drop becomes purely an agent-utilities dependency swap.

See architecture/numeric-kernel.md.

Client drivers

  • Multi-language client drivers (B1.7) — thin client bindings for the Program-B engine Methods that had no client surface: the native broker + append-log streams (EG-275..284/314), RBAC admin (EG-KG.compute.feature), online backup/restore (EG-090), and NL→query (EG-080). A FULL Python surface (client.broker/.rbac/.admin + query.nl_query) plus THIN generated-from-the-Method-list JS (clients/js) and Go (clients/go) bindings over the same framed-MessagePack transport; Pi-contract preserved (no heavy deps). See interfaces/clients.md (EG-328).

Shipped in 2.2.0

Everything below was previously listed here as roadmap and is now ✅ — see the capability matrix for the per-operation evidence and concepts for the authoritative CONCEPT:EG-* definitions.

SQL — full Postgres + multi-wire parity

  • Compound/complex DML WHERE, INSERT … SELECT, UPDATE…FROM/DELETE…USING (EG-045/046/047).
  • ON CONFLICT upsert + RETURNING (EG-KG.query.delete-returning-sees-row); wire transactions BEGIN/COMMIT mixing graph + user tables with RYOW + TransactionStatus (EG-KG.compute.kg-transaction-is-pinned); CREATE VIEW / DROP VIEW durable catalog (EG-072).
  • SQL DDL + arbitrary user tables + COPY.

SPARQL — full Stardog/GraphDB parity

  • Content negotiation (XML/CSV/TSV/Turtle; EG-KG.ontology.content-negotiation-serializers), rich FILTER (regex/arithmetic/IN/builtins; EG-KG.ontology.rich-filter), FROM/FROM NAMED (EG-KG.ontology.from-from-named), sub-SELECT · SERVICE federation · MINUS · negated property sets (EG-KG.ontology.sub-select/052/055/056), SPO/POS triple index + selectivity join-ordering (EG-KG.ontology.capability-catalog).
  • ASK/CONSTRUCT/DESCRIBE + UPDATE + the W3C /sparql endpoint; true named-graph quad dataset.

OWL / RDF validation

  • rdfs:range in EL completion (EG-KG.ontology.owl-reasoning), OWL-DL tableau (EG-KG.ontology.concept-2), SWRL/RuleML rules (EG-KG.ontology.concept-3).
  • SHACL Core validation (eg-shacl, EG-KG.ontology.concept-6), ShEx validation (eg-shex, EG-KG.compute.concept-2), and Integrity Constraint Validation with guard mode (EG-KG.ontology.wired-into-commit-write).

Graph — full Neo4j parity + Bolt wire

  • Cypher REMOVE (EG-KG.query.cypher-execution), ORDER BY/SKIP/WITH/OPTIONAL MATCH/OR/aggregation/DISTINCT (EG-KG.query.eg-extend-read-side), var-length + fixed hops + path binding (EG-KG.query.concept-2); writes (CREATE/MERGE/SET/DELETE).
  • Neo4j Bolt v4.4 wire adapter so native drivers connect (EG-KG.query.bolt-wire-protocol).

GraphQL

  • Subscriptions over CDC (EG-064), fragments/variables/directives (EG-KG.query.fragments-variables-directives), Relay pagination (EG-KG.query.graphql-cursors).
  • Apollo Federation v2 subgraph (EG-295) + enterprise hardening (APQ, depth/cost limits, introspection toggle; EG-KG.domains.graphql-enterprise-hardening); GraphQL mutations.

Time-series · Vector · Blob

  • Op::Window execution (EG-KG.query.streaming-execution), per-point retention trim (EG-KG.temporal.bucket-cutoff-trim), hybrid metadata pre-filtering (EG-070), content-defined chunking (EG-KG.storage.backward-manifest-read).
  • Exact/flat vector index + recall@k harness (EG-KG.query.concept-5).

Multi-wire adapters (one WireProtocol trait)

  • The WireProtocol/WireSession keystone with pgwire refactored behind it (EG-KG.compute.subsystems-reference), plus SQLite (EG-KG.query.concept-3), MySQL/MariaDB (EG-KG.query.kg-2), MSSQL TDS (EG-KG.query.hand-rolled-tds-server), Redis RESP (EG-KG.ontology.resp2-resp3-codec-round), and the S3 REST surface (EG-KG.ontology.object-put-get-head).

Message broker (Phase Y)

  • The native engine task queue ClaimNext (EG-KG.compute.atomically-claim-oldest-pending) grown into a RabbitMQ-class broker: exchanges/routing + AMQP wire (EG-275), MQTT (EG-281) and STOMP (EG-KG.ontology.stomp-frame-codec-unit) wires, DLQ (EG-KG.compute.dead-letter-queues), TTL/expiry (EG-KG.compute.message-ttl-expiry), priority queues (EG-KG.compute.priority-queues), delayed/scheduled delivery (EG-KG.compute.delayed-scheduled-delivery), consumer groups + prefetch/QoS (EG-KG.compute.groups-qos-prefetch-honoring), publisher confirms + consumer acks (EG-KG.compute.publisher-confirms-consumer-qos), and replayable append-log streams (EG-283).

Observability (Phase T)

  • Log ingestion (OTLP/_bulk/syslog; AU-KG.ingest.self-ingest), Parquet-on-object-store segments (EG-KG.retrieval.observability-search), log search + _search API (EG-KG.query.concept-4), distributed traces (EG-163), VRL ingest pipelines (EG-165), PromQL query API (EG-172), and super-cluster federated search (EG-KG.ontology.federation-client).

GIS (eg-geo)

  • Spatial modality (EG-KG.ontology.singles-concept) built to real-GIS depth: full geometry model (EG-KG.domains.geometry-collections), DE-9IM + RCC8/Egenhofer (EG-KG.ontology.de-9im-relations/155), constructive algebra (EG-KG.ontology.concept-9), geodesic ops (EG-KG.ontology.concept-8), CRS registry + reprojection (EG-KG.domains.coordinate-reference-system/262), durable R-tree (EG-KG.domains.spatial-strtree-index), GeoJSON/WKB/GPX I/O (EG-KG.domains.geojson-gpx-formats), map tiling + MVT (EG-KG.domains.map-tiles), raster tile pyramids + hand-rolled PNG (EG-KG.domains.raster-build/339), Shapefile/KML/GeoParquet I/O (EG-KG.domains.geo-formats), routing/isochrones/TSP (EG-KG.domains.geo-routing), map-anchored tasks (EG-KG.domains.geo-task), and GeoSPARQL (EG-KG.ontology.concept-10).

New modality engines + subsystems

  • eg-tensor N-D array store (EG-085), eg-stream windowed CEP (EG-KG.query.pipelined-execution), eg-kvcache tiered + shared KV-cache with the vLLM/LMCache server endpoint (EG-185/186/187), and the agent-memory primitives — summary tier (EG-KG.compute.hierarchical-summary-tier-eg), episodic→semantic consolidation (EG-KG.compute.consolidate-cluster), decay/reinforcement (EG-222), trajectory memory (EG-099), LeanRAG retrieval (EG-195).

Unified planner / UQL · Distribution

  • Op::Foreign name → ForeignSourceSpec resolution (EG-073), NL → query dual-mode (EG-078/079/080).
  • Multi-Raft groups + GroupRouter + online resharding, N-participant cross-shard 2PC with crash recovery, parallel-commit + read-only-participant fast path (EG-KG.txn.cross-shard), non-blocking commit (EG-KG.txn.harness-crash).

Reading this as a contributor? The forward-roadmap table above is the whole of what's left. Each item lands with tests + a docs/concepts.md entry, and flips to ✅ in the capability matrix as it merges.