Skip to content

rocket_launchRoadmap

The 0.6.x line ships as attested micro-releases. Each cut is CI-built and signed with SLSA Build Provenance v1. v0.7.0 is the graduation at the end of the carving work — not a parallel branch.

check_circleDelivered — the v0.6.x bulk-ingest line

The v0.6.x line landed the staged bulk loader benchmark and the post-benchmark threading work that made the 8.2-billion-triple load fast. Tasks T1–T6 shipped in v0.6.14:

TaskShippedWhat it delivered
T1–T6v0.6.14Post-benchmark threading — parallel STAGE COPY, concurrent index build, adaptive self-tune, and the resolve-strategy selector. 37 % faster than the v0.6.13 hash baseline (4 h 53 m / 466 K tps vs. 6 h 41 m / 340.7 K). See Scale & benchmarks.

rocket_launchToward v0.7 — graph carving (C1–C6)

v0.7 is gated on graph carving and per-graph index reorganization. This work ships as further v0.6.n micro-releases, each a clean increment, culminating in the v0.7.0 graduation:

StepReleaseWhat it delivers
C1 — Carve by queryv0.6.15Cut a query-defined slice out of the full source graph — the Carve(query) verb.
C2 — Re-encodev0.6.16Re-encode the carved slice into a fresh, small dictionary so the working set shrinks to the slice.
C3 — Full hexastore + string indexv0.6.17Full 6-way bidirectional hexastore plus a pg_trgm string index on the carved graph.
C4 — Park-graph lifecyclev0.6.18seal / unseal / unload — the park-graph lifecycle behind the Unload verb.
C5 — Right-sizingv0.6.19Right-size to the operator's hardware automatically — no postgresql.conf tuning required.
C6 — At-scale validationv0.7.0At-scale validation of the carved-reasoning pipeline → the v0.7.0 graduation.

These steps build the scale-meets-hardware chain end to end: ingest the full graph in parallel (shipped), carve a right-sized slice (C1/C2), index it (C3), park the rest (C4), right-size to the box (C5), and validate the whole flow at scale (C6).

schedulePer-pillar forward edges (the v0.6-FUTURE backlog)

Each pillar carries its own forward edge from the v0.6-FUTURE spec — performance and depth work, not missing pillar surface:

PillarForward-edge items
StoragePer-graph index reorganization for carved graphs — the full 6-way bidirectional hexastore + pg_trgm string index built per carved slice (C3, v0.6.17). The staged bulk loader, resolve-strategy selector, and big-RAM tuning already shipped in v0.6.14.
SPARQL queryexecutor.rs core-BGP carve · federated SERVICE · incremental materialisation · RDF 1.2 triple terms (gate E-011). The native SHACL-SPARQL engine is already shipped (mode => 'pgrdf') — see Validation.
InferenceReasoning over a carved, right-sized slice — single-threaded by design; the carving line (C1–C6) is what makes at-scale reasoning practical.
ValidationSHACL-SPARQL constraint execution is shipped as the authoritative mode => 'pgrdf' gate. The open upstream item is E-014 — the rudof 'sparql' mode returns the wrong verdict — tracked below.

scheduleDocumented upstream gates (not pgRDF defects)

Two backlog items are blocked on a third-party crate. They are documented upstream dependencies — the pgRDF side is built and shipped; the upstream side is the gate:

ErratumWhatUpstream gate
E-011RDF 1.2 triple terms and the crates.io publishBoth wait on gtfierro/reasonable#50. The crates.io publish is deliberately held until that lands — the tarball / OCI bundle are the consumption path meanwhile.
E-014rudof 'sparql' mode returns the wrong verdict on common SHACL-SPARQL topologiesThe shacl 0.3.2 SparqlEngine (E-012 — engine reachability — is resolved; 'sparql' now reaches the engine) returns conforms=true / 0 violations where the W3C answer is non-conforming. pgRDF beats rudof here: the shipped, authoritative mode => 'pgrdf' gate returns the correct verdict, so SHACL-SPARQL constraint execution is delivered. The open item is purely the rudof verdict bug, tracked upstream on rudof. See SHACL-SPARQL.

The full upstream-gate detail lives on the SPARQL forward edge.

See also

MIT licensed. Documentation for pgRDF — built with VitePress, served via GitHub Pages.