rdflib-turbo

Performance

Matched public Graph pipeline

Parse → OWL-RL → SHACL: about 11× in this measured synthetic run. Both arms use public Graph APIs, expand into the same source Graph and validate in place. The Turbo trio keeps one RDF dataset in memory throughout; a separate non-timed ownership regression checks the actual shared route without adding spies to timed children.

Measured 31 October 2026, macOS 27.0.1 (arm64), CPython 3.13.16, native uv 0.13.5: original registry wheels of rdflib-turbo 0.2.0a23 / owlrl-turbo 0.2.0a1 / pyshacl-turbo 0.7.1, versus RDFLib 7.6.0 / OWL-RL 7.6.2 / pySHACL 0.40.1. These are medians of five fresh children per arm after a separate warmup. The total includes Graph construction/parsing, OWL-RL closure construction/expansion and public validation/report construction; imports, fixture text and SHAPES preparation are outside. Complete scoped/generalized facts, literal metadata, report/text, namespace/effect/flag gates passed for rdflib.NORMALIZE_LITERALS=True. The separate NORMALIZE_LITERALS=False check differs in a global-setting side effect and has no performance ratio.

Peak RSS is absolute whole-process memory captured before correctness encoding; it includes imports, setup and retained results. This single-host result is not a universal speed or memory claim. See the portable numeric provenance and method/locked replay. From a pyshacl-turbo checkout, use a fresh output directory:

uv run --no-project --python 3.13.16 python -I -B benchmarks/matched_graph_driver.py \
    --cohort benchmarks/matched-graph-releases.json --roles baseline --widths 1000 \
    --repeats 5 --output benchmark-artifacts/matched-graph-readme

Historical measurements and provenance

N-Triples serialization: 4.4–4.9× faster than RDFLib on 100,000 synthetic triples. The candidate measured on 23 October 2026 at c196691 serializes directly from the selected native graph and returns UTF-8 bytes without a Python Unicode copy. It preserves RDFLib's URI rejection and caught-transaction behavior; the corrected implementation is measured below.

Measured on macOS ARM64 with CPython 3.13.16 on 23 October 2026: medians of six rotated cold processes per case, using the published baseline wheel and an independently audited release build of the candidate. Timings cover serialization only; whole-process peak RSS includes imports, input construction and retained output, measured before correctness parsing (decimal MB). Complete exact-term and blank-node-ID comparisons cover small/default/named/blank-graph controls; the 100,000-triple cases use a matching streaming count/hash witness, which is probabilistic rather than a complete row comparison. A separate 1,024-triple Unicode test reduced Python-traced UTF-8 peak allocations from 177,730 to 48,461 bytes, using the same cold child in both projects; that excludes native allocations. These are synthetic NT-only measurements, not end-to-end or universal speedup claims. The shared host's one-minute load ranged from 16.05 to 25.26 and large-case CPU variation from 1.3% to 12.6%; absolute timings reflect that load. See results and provenance and reproduction instructions.

Compare RDFLib + OWL-RL + pySHACL with rdflib-turbo, owlrl-turbo and pyshacl-turbo used together. The Turbo pipeline keeps one native dataset through parsing, in-place inference and SHACL validation.

End-to-end measurements for 1,000 synthetic nodes:

Turbo uses more memory in these runs: 8–46% higher whole-process peak RSS. Ordinary RDF inference closures and complete SHACL report graphs match. Generalized RDF statements and report text differ, so these timings are not a full-output speedup claim or a claim of complete drop-in parity.

Measured in CI on 21 October 2026, using source-built wheels and Python 3.13: rdflib-turbo 0.2.0a7, owlrl-turbo 0.1.0a3, pyshacl-turbo 0.6.0 versus rdflib 7.6.0, owlrl 7.6.2, pyshacl 0.40.1. Times are medians of seven samples after warmup and include data parsing, shape compilation, inference, validation and full report construction. Shapes are parsed before timing. Memory is the median peak RSS of three fresh processes, including the interpreter, imports and retained output. Shared CI hosts and small synthetic cases do not establish universal speedups or large-dataset capacity.

See the complete Linux measurements, macOS measurements and continuously recorded benchmark history for all samples, source identities, dependency locks and other dataset sizes.

Overview

Functional alpha of a native RDF dataset runtime for Python, built on Oxigraph and exact-term compatibility work extracted from pyshacl-turbo.

Companion turbo projects: pyshacl-turbo and owlrl-turbo.

uv add rdflib-turbo==0.2.0a23

from rdflib_turbo import Dataset

# The data stays native between parsing, querying and owlrl-turbo inference.
data = Dataset().parse(data='<urn:alice> <urn:name> "Alice" .')
for row in data.query('SELECT ?name WHERE { <urn:alice> <urn:name> ?name }'):
    print(row.name)
print(data.serialize(format="nquads"))

The alpha provides native parsing and serialization, SELECT/ASK/CONSTRUCT/DESCRIBE, exact literal spellings and datatype identity, named graphs, independent copies, and explicit RDFLib conversion. Query results are RDFLib Result objects; returned terms are Python RDFLib objects. Native operations release Python's GIL. A lock covers each complete native read or write. Parsing and inference commit only on success. Parsing separate documents gives blank nodes separate scopes.

Supported native format names: turtle, nt, nquads, trig, xml, n3, json-ld. Store accepts UTF-8 text, bytes or a local path; its source URL and file-object inputs are not implemented. Graph and Dataset use RDFLib input-source handling. Dataset queries use the default graph unless default_union=True is selected. Low-level Store serializers reject graph-only formats with named-graph data; Graph and Dataset retain their RDFLib format semantics.

This is not a complete drop-in RDFLib replacement. It uses the rdflib_turbo namespace. Dataset and Graph provide inference/SHACL integration over the native Store, which also supplies persistence, SPARQL Update and explicit transactions. Dataset and StoreDataset name the same RDFLib Dataset facade. Arbitrary RDFLib storage plugins, initial query bindings, custom functions and remote SERVICE execution are not provided. RDFLib conversions copy data and explicitly preserve caller blank-node identifiers. Query results are eagerly materialized. Performance results apply to their recorded versions and workloads; complete conformance is not claimed.

RDFLib Graph and dataset facades

Import Graph and Dataset explicitly from rdflib_turbo; importing rdflib continues to select the original RDFLib classes.

from rdflib import Literal, URIRef
from rdflib_turbo import Graph, StoreDataset

alice, name = URIRef("urn:alice"), URIRef("urn:name")
graph = Graph()
graph.add((alice, name, Literal("Alice")))
assert list(graph.objects(alice, name)) == [Literal("Alice")]
assert graph.query('ASK { <urn:alice> <urn:name> "Alice" }').askAnswer

dataset = StoreDataset()
dataset.default_graph.add((alice, name, Literal("default")))
named = dataset.graph(URIRef("urn:people"))
with dataset.transaction() as pending:
    pending.graph(URIRef("urn:people")).add((alice, name, Literal("Alice")))
    assert len(named) == 0              # Other views see committed state
assert list(named.objects(alice, name)) == [Literal("Alice")]

These facades reuse RDFLib terms, namespace management, serializers and graph utility algorithms. Statements, pattern matching, counts and SPARQL remain in native storage; no reference Memory graph or Python SPARQL evaluator runs. Committed Graph and Dataset iteration reads bounded native batches. Union statements and their lazy context iterators share one pinned snapshot. Explicit transaction iteration retains its pending-state reader; query results and serialization can still materialize larger results.

StoreDataset graph views share their native backend and namespace bindings. Queries use the default graph unless StoreDataset(default_union=True) is selected; a named Graph query uses its own graph as the query's default graph. Dataset-level queries can address named graphs with GRAPH. Graph(store=store, identifier=...) and StoreDataset(store=store) also accept an existing native Store, including a persistent Store. Namespace bindings are facade metadata and are not persisted.

Use Graph.transaction() or StoreDataset.transaction() as a context manager. Pending views see the same transaction's writes; normal exit commits statements and namespace changes together, while exceptions abort them. Native operation failures abort even when caught inside the scope. Nested transactions and manual commit()/rollback() on the RDFLib adapter are unsupported. Parsing is atomic and uses RDFLib input-source handling followed by the native parser. JSON-LD accepts version and base; ordinary string/file input in the default 1.1 mode uses the native parser. Input sources exposing JSON-LD HTTP context-link metadata execute the selected RDFLib plugin so that metadata is preserved. Options such as context, skolemize, encoding, extract_all_scripts, and version 1.0 execute the selected RDFLib JSON-LD parser directly into the same native transaction. Plugin errors abort its writes and namespace changes. Unknown options, unsupported versions, generalized RDF output, and keyword options for other formats raise explicit errors. Plugin parsing has its own Python execution cost. JSON-LD preserves explicitly supplied blank-node IDs; generated IDs remain fresh. Other document syntaxes, including Turtle, scope blank-node labels to each parse operation.

SPARQL uses the native processor. Queries accept strings or supported prepared RDFLib queries; updates require strings. Both accept RDFLib initBindings, including variables hidden by nested projections. Parameter terms remain native; they are never inserted into query text. Graph.update() uses that graph for unqualified reads and writes, including graphs identified by blank nodes, and retains the enclosing transaction. Use StoreDataset.update() for dataset updates, including explicit GRAPH clauses. Alternate processors, prepared updates, quoted graphs and arbitrary RDFLib store plugins are unsupported. See the facade decision, scoped UPDATE contract and bounded read contract.

Transactional storage

from rdflib_turbo import Store

with Store() as store:                 # In-memory default
    with store.transaction() as tx:
        tx.parse(data='<urn:alice> <urn:name> "Alice" .')
        tx.update('INSERT DATA { <urn:alice> <urn:active> true }')
        assert len(tx) == 2             # Sees this transaction's writes
        assert len(store) == 0          # Readers see committed state
    assert len(store) == 2               # Committed at normal scope exit

with Store("./database") as store:     # Persistent RocksDB directory
    store.update('INSERT DATA { <urn:alice> <urn:name> "Alice" }')

Exceptions abort the transaction. A native operation failure also aborts it if its exception is caught inside the scope. Transaction handles become unusable after exit. Parsing and updates outside a scope are individually atomic. Another writer or close on the same handle fails immediately while a transaction is active. Store queries use the default graph unless default_union=True is requested. Named graphs and exact literal spelling/datatype identity survive commit and reopen.

Python wheels include RocksDB support; Store() still defaults to memory. Rust-only builds enable persistence with --features persistence. Closing a store releases its database handle. Store and transaction inference use the existing rdfs, owlrl and both compatibility modes directly on the native transaction:

with Store() as store:
    with store.transaction() as tx:
        tx.parse(data="<urn:Child> <http://www.w3.org/2000/01/rdf-schema#subClassOf> <urn:Parent> . <urn:alice> a <urn:Child> .")
        tx.infer(inference="rdfs", destination="urn:inferred")
        assert tx.query("ASK { GRAPH <urn:inferred> { <urn:alice> a <urn:Parent> } }").askAnswer

Inference reads the union of graph contexts and writes only new facts to the requested destination (the default graph when omitted). It sees pending updates; errors abort the entire transaction. Standalone store.infer(...) is atomic. See the direct inference contract. Dataset, Store and Transaction also accept axiomatic_triples=True and datatype_axioms=True on infer; Dataset supports these flags on infer_delta as well. Both flags default to false and apply only to that invocation. They add the selected reference RDFS/OWL-RL axiom tables before reasoning. Programmatic results retain generalized RDF terms, including literal subjects; document formats keep their own syntax contracts. Global Python datatype-conversion hooks are not supported. See the axiom options contract. The low-level Store has no namespace-binding API. Graph and StoreDataset provide namespace management through their RDFLibStore adapter, without registering a named RDFLib storage plugin. Compiled pyshacl-turbo shapes validate Store and open Transaction directly; stored data is never copied into Dataset for inference or validation.

The shared native engine includes sh:singleLine from the SHACL 1.2 Core Working Draft of 14 October 2026. Borrowed Graph and StoreDataset validation retains literal identity and reports the draft constraint through the same native owner. This selected, at-risk draft coverage does not establish complete SHACL 1.2 conformance.

from pyshacl_turbo import compile_shapes
from rdflib_turbo import Store, ValidationError

shapes = compile_shapes('''@prefix sh: <http://www.w3.org/ns/shacl#> .
  <urn:S> sh:targetNode <urn:s>;
    sh:property [sh:path <urn:name>; sh:minCount 1].''')
with Store("./database") as store:
    with store.transaction(shapes=shapes, iterate_rules=True) as tx:
        tx.parse(data='<urn:s> <urn:name> "Alice".')
        tx.infer(inference="rdfs")

The gate runs after the body and commits only after a conforming report has been constructed. Nonconformance raises ValidationError carrying report and text, and aborts all pending writes, inference and rule additions. Exceptions abort the same owner; caught native execution or report-processing errors leave the transaction closed. Direct shapes.validate(tx) returns a report and lets the caller choose acceptance. Companion facades share the provider's one compiled engine and compiled program types.

Oxigraph buffers the transaction delta in memory. The base dataset is not copied for rollback; transaction changes, parsed input documents and materialized query or serialization results can still be large. See the storage decision and transaction ownership contract.

Inference and SHACL

from owlrl_turbo import expand
from pyshacl_turbo import validate

expand(data, inference="rdfs", inplace=True)
conforms, report, text = validate(data, shacl_graph="shapes.ttl")

This shared companion path accepts Dataset, Graph, Store and Transaction owners. Graph views retain their selected physical scope; Dataset validation reads its union. StoreDataset is the same class as Dataset. Converting native facades through Dataset.from_rdflib is explicitly rejected to prevent copying input.

The canonical-facade integration requires coordinated provider and companion artifacts with matching runtime and engine identities; installed-wheel acceptance of this source integration remains pending. Parsing, in-place inference and ordinary validation retain the same native data object. Only the validation report becomes an RDFLib graph. Rules stage changes privately and commit only after success when inplace=True is requested. See the SHACL ownership ADR, the runtime ADR and source provenance.

Requires Python 3.11+. macOS wheels require macOS 11 or later on Intel and ARM64. Source builds require Rust 1.99.1, a C++20 compiler and libclang. Both project code and extracted inference code are MIT OR Apache-2.0; vendored sources retain licenses. Copyright 2026 DataRoad Inc, Delaware, USA, trading as Legra.

Development

Development uses native uv project commands and the committed uv.lock. Run make dev to sync the project and test tools, then make check. Add or update Python dependencies with uv add; run Ruff, mypy and pytest through uv run. Build wheels and source distributions with uv build. The coverage and release harnesses create isolated uv projects and install the exact built wheels.

The native heap profiler and allocation budget explain phase-level memory measurements and the inference insertion journal.

The shared companion-project benchmark suite compares parsing, serialization, queries, inference, conversion costs and complete SHACL pipelines against the reference packages. With all three repositories as siblings, run make benchmark-companions. It uses pinned published wheels and separate timing and peak RSS workers; it does not use local editable builds.

Use make dev and make check. The initial alpha is published only by the verified GitHub Actions workflow; see release instructions.

make dev
uv run --locked --extra test ruff check python tests scripts
uv run --locked --extra test pytest
cargo fmt --check
cargo clippy -- -D warnings

Coverage

Coverage badges update from successful main-branch CI measurements, including the daily scheduled run. Python lines and branches are separate; Rust badges measure owned native source exercised by Rust and Python integration tests. Rust branch coverage and dependency/vendored-patch coverage are not measured by these badges.

CI requires 100% measured Python package line and branch coverage. Coverage is separate from full API compatibility and standards conformance.