📣 New Release: codeanalyzer-python 1.6.0

[github-actions[bot]]

codeanalyzer-python 1.6.0 makes level-4 analysis of large projects use about 7x less memory, with the same output. It also raises the minimum Python version to 3.11.

⚠️ Breaking changes

• Python 3.11 or newer is now required. Python 3.9 and 3.10 are no longer supported (requires-python = ">=3.11"). The new memory-saving records need pydantic v2, and below 3.11 the dependency pins installed pydantic v1.
  Migration: install canpy under Python 3.11+. The tomli backport and the pydantic-v1 code paths are gone. codeanalyzer.schema.PYDANTIC_V2 still imports, and is always True.
• The schema's leaf record types are now dataclasses, not pydantic models. This affects in-process users only: Span, BodyNode, the cfg/cdg/ddg/summary/param_* edges, PyCallEdge, PyCallsite, PyCallArgument, PySymbol, PyComment, PyVariableDeclaration, PyCallableParameter, PyDecorator and PyImport no longer have .model_dump() or .model_validate().
  Migration: use dataclasses.asdict(...) or the constructor. The container models (PyModule, PyClass, PyCallable, PyApplication, Analysis) and the codeanalyzer.schema helpers handle them transparently. The analysis.json format and the schema version are unchanged.

Highlights

• About 7x less memory at level 4 (Cut level-4 memory ~7x: slotted leaf records, streamed JSON/cache/cypher, early release of intermediates #224). On a 6,685-module project (Odoo 20.0), peak memory fell from 42.6 GB to 5.7 GB for JSON output, and from 65.1 GB to 5.5 GB for --emit neo4j. That project now fits on a 16 GB laptop. On the same input, analysis.json and the analysis cache have the same content as 1.5.4. The Neo4j snapshot has the same statements, though edge batches can come in a different order.
• Streamed output. analysis.json, the cache and graph.cypher are written piece by piece, not built as whole-tree copies in memory. analysis.json now uses compact JSON separators, so the whitespace differs but the content is the same.
• Ray memory controls. The symbol-table phase runs on a private Ray cluster with object spilling turned off, and the cluster shuts down when the phase finishes. Two new environment variables control it: CANPY_RAY_OBJECT_STORE caps the object store (in bytes, default 1 GB), and CANPY_RAY_CPUS caps the number of workers.

Thanks to @saurabhjha1 for the memory work, their first contribution!

Upgrade

pip install -U "codeanalyzer-python==1.6.0"

For the optional live Neo4j push: pip install -U "codeanalyzer-python[neo4j]==1.6.0"

Links: GitHub release · CHANGELOG