[trungdong]

Makes resolving a full IRI linear in the number of registered namespaces, so reading RDF whose IRIs sit in many directories and that declares no prefixes stops being quadratic in prov's share. Closes #691.

Two scans grew with the namespace count on every new IRI. NamespaceManager tested every registered namespace to find the one that compacts a full IRI, and the PROV-O reader walked every binding of the source graph, including the one rdflib mints for each new directory, for every IRI the document could not name. A read with one record per directory minted a namespace per record, so both scans ran over a list that grew with the input.

The manager now keeps an index of the namespaces its dict holds, by URI length, and tests iri[:n] for each distinct length. One match is the scan's answer by construction, and two or more (nested namespaces, or two prefixes at one URI) fall back to the unchanged scan. The default namespace keeps its non-empty-local-part rule and the #599 parent-scope rule still runs first. A manager pickled by 3.2.3 has no index, so __setstate__ builds one; a current pickle and a deep copy carry theirs.

decode_document reads the source graph's bindings once, when it starts, and the graph lookup scans that snapshot. A binding the reader mints later is for a namespace it has already registered on the document, which resolves the IRI first, so the snapshot gives the same answers. An IRI the document cannot compact (one with no colon, or a blank-node label) reads the live bindings, as does any direct call to decode_container or decode_rdf_representation.

Every answer is the one main gives: a 20,000-sequence differential of the namespace manager (registrations, defaults, aliases, child and attached bundles, pickles and copies) and an 862-input differential of PROV-O reads (the RDF fixtures, the issue's N-Triples, multi-bundle TriG, rebound defaults and random graphs, at six hash seeds) found no difference in names, documents, namespaces, warnings or exceptions. NamespaceManager's docstring now says that namespaces are registered through add_namespace and set_default_namespace and that the dict's own write methods bypass its indexes; a namespace written into the dict directly no longer takes part in compaction.

The issue also proposed clearing the resolve cache per prefix instead of whole. That was built and measured, gained nothing once the graph snapshot was in place, and served a stale name in three cases, so add_namespace keeps clearing the whole cache and spec 5's F15 and section 4.4 are amended.

The issue's probe (N-Triples, one record per directory, five runs, medians, CPython 3.12.8):

Directories	main	this branch
500	0.142 s	0.043 s
2,000	2.037 s	0.547 s

The rest of the 2,000-directory read is rdflib's own minting (compute_qname).

ab.py main --python 3.12 on CPython 3.12.8, garbage collection off, load average 2.2:

benchmark                                                      base min   head min   change
benchmarks/test_bench.py::test_construct                         0.0445     0.0454   +2.0%
benchmarks/test_bench.py::test_deserialize[json]                 0.0776     0.0761   -2.0%
benchmarks/test_bench.py::test_deserialize[jsonld]               0.0811     0.0813   +0.3%
benchmarks/test_bench.py::test_deserialize[provn]                0.2184     0.2205   +1.0%
benchmarks/test_bench.py::test_deserialize[rdf]                  0.6585     0.6544   -0.6%
benchmarks/test_bench.py::test_deserialize[xml]                  0.1341     0.1339   -0.1%
benchmarks/test_bench.py::test_equality                          0.0387     0.0376   -2.8%
benchmarks/test_bench.py::test_prov_to_graph                     0.1740     0.1687   -3.1%
benchmarks/test_bench.py::test_serialize[json]                   0.0519     0.0520   +0.1%
benchmarks/test_bench.py::test_serialize[jsonld]                 0.0556     0.0557   +0.2%
benchmarks/test_bench.py::test_serialize[provn]                  0.0284     0.0279   -1.8%
benchmarks/test_bench.py::test_serialize[rdf]                    0.8105     0.8130   +0.3%
benchmarks/test_bench.py::test_serialize[xml]                    0.1047     0.1030   -1.6%
benchmarks/test_bench.py::test_unified                           0.0582     0.0584   +0.3%
OK

The benchmark document has few namespaces, so its rows do not move; the gain shows on the probe's shape.

[codacy-production]

Up to standards ✅

🟢 Issues 0 issues

Results:
0 new issues

View in Codacy

🟢 Metrics 31 complexity

Metric	Results
Complexity	31

View in Codacy

_{NEW Get contextual insights on your PRs based on Codacy's metrics, along with PR and Jira context, without leaving GitHub. Enable AI reviewer
TIP This summary will be updated as you push new changes.}

[coveralls]

coverage: 98.416% (+0.01%) from 98.405% — perf/3.3.0-iri-resolution into main