github Graphify-Labs/graphify v0.9.114
graphify 0.9.114

58 minutes ago

Performance work on the extract/build/update pipeline, plus a batch of extractor and build fixes. Every entry marked Perf preserves graph output exactly (byte-identical); the extractor fixes change output and share one AST cache schema bump (see the note at the end).

  • Perf: ids normalize_id gets an ASCII fast-path, pre-compiled regexes and an LRU cache, so the common (ASCII, repeated) identifier case skips the Unicode normalization loop. ~1.26x on the compute path and a large win on repeated ids, which recurs heavily across a corpus.
  • Perf: cache save_cached builds the portable on-disk payload copy-on-write instead of copy.deepcopy-ing every per-file result; ~20% faster on the cache-write transform, scaling with per-file graph size. The input is never mutated and the on-disk bytes are identical.
  • Perf: extract each source file is read from disk once (the key hash and the parse share one buffer) instead of twice on a cold build, and the tree-sitter Parser is reused per language within a worker instead of rebuilt per file (file reads roughly halved on the worker path).
  • Perf: extract the cross-file resolution finalize builds its shared node/edge indexes in a few fused passes instead of ~6 independent full scans of the graph; ~15% faster on the finalize stage.
  • Perf: watch the no-change topology comparison uses an order-independent content hash computed in one pass instead of sorting and re-serializing the whole graph; ~1.4x on that compare.
  • Perf: cluster the community partition is cached by a topology hash and reused when the graph is unchanged (~9x on a cluster cache hit), and the deterministic edge-sort key drops a per-edge json.dumps in favour of a cheap tuple key (identical ordering).
  • Perf: analyze suggest_questions' betweenness centrality now runs a flat-array Brandes implementation instead of networkx's dict-based one, with bit-for-bit identical scores; ~2.6x faster, growing with graph size (#5303, thanks @Max17190).
  • Perf: watch each preserved source path is rebased once per incremental update instead of recomputed per node and edge; the update output is byte-identical (#5322, thanks @Max17190).
  • Fix: java a field or class annotated with Lombok @Getter gets the synthesized getX()/isX() getter node javac sees, so obj.getX() resolves to it; boolean fields use isX, @Getter(AccessLevel.NONE) is respected, a hand-written getter is not duplicated, and a non-Lombok @Getter is ignored (#5148, thanks @Deepjyoti-Sarmah).
  • Fix: kotlin a member call on a receiver with a declared type (parameter, property, val x: T, or val x = T()) binds to that type's method across files instead of the last same-named method in the file, and interface calls dispatch to their implementers; an ambiguous or unknown receiver fails closed (#5355, thanks @xiongjianxu).
  • Fix: sql a PostgreSQL PL/pgSQL routine body that the grammar only parses through ERROR recovery (RETURN QUERY SELECT ... FROM t) has its table reads recovered as reads_from edges instead of dropped (#5349, thanks @Deepjyoti-Sarmah).
  • Fix: build when an AST-extracted edge and a semantic (LLM) copy of the same edge tie, the deterministic EXTRACTED edge is kept, so a later docs --update no longer deletes a code edge that a semantic copy had overwritten (#5334, thanks @Ildar-gn).
  • Fix: sql a T-SQL procedure whose header the grammar only partly parses (tree-sitter-sql 0.3.12+ misreads the name, absorbing unparenthesised params or a ""/]] escape) yields a single routine node instead of a duplicate under the misread id; a real routine sharing the misread's line is kept (#5236, thanks @Deepjyoti-Sarmah).
  • Fix: vbnet a method declared after a multi-line string literal is no longer dropped: the file is reparsed with multi-line strings flattened to a line-preserving placeholder first, so an unterminated string on one line no longer swallows the rest of the class; single-line files are parsed unchanged (#5320, thanks @Deepjyoti-Sarmah).
  • Fix: cpp a std::-qualified type (std::chrono::seconds, std::vector<T>) is no longer rewired onto a same-named project symbol; the std reference gets its own sourceless stub the corpus rewire skips, while the std type's template arguments still resolve to the project's own types (#5219, thanks @deba2k5).
  • Fix: python a plain import mod where mod sits on a sys.path.insert(...) directory (a non-package script dir) now binds mod.func() to the func function node, matched to the unique file on that path and failing closed on ambiguity or a stdlib name; a calls edge is no longer aliased onto a file node (#5253, thanks @madhusudanpatnaik).
  • Fix: serve a query that filters on the call relation now keeps cross-repo calls edges (written as relation=calls, context=cross_repo), which were silently dropped by the context-only filter (#5257, thanks @LHopitalOne).
  • Fix: build a member's structural method/contains edge to its owner is kept over an opposite-direction member-usage edge (a method reading its class's own static property or constant), so the member no longer vanishes from the type's walk; non-usage relations (inherits/implements) keep the existing rule (#2391, thanks @livingforelvis).

The AST cache schema is bumped to 50; warm caches re-run the affected Kotlin, Java, SQL, VB.NET, C++ and Python (sys.path import) extractions.

Don't miss a new graphify release

NewReleases is sending notifications on new releases.