Module Trimming

Module and class trimming

cvista ships ~84 modules (PyVista's measured closure) out of VTK's ~160, and within those modules drops classes PyVista never touches. The result is a ~37 MB wheel vs stock's ~120 MB. The full rationale and measurements are in docs/build-internals.md (levers 1–11). This skill is the operational guide.

The three levers

All three live in cvista-config/. The class lists are append-only and closure-closed by design — every entry is there because nothing kept refers to it.

1. Module deny-list — cvista-config/_modules_minimal.cmake VTK_BUILD_ALL_MODULES OFF; only PyVista's closure is enabled via WANT/YES. This is ~53 direct C++ imports plus the IO format readers reached through the object factory plus the rendering impl modules (RenderingContextOpenGL2, RenderingGL2PSOpenGL2).

2. Lever A — NOWRAP — cvista-config/_nowrap_classes.cmake (~1,173 classes) The C++ compiles, but the Python wrapper is skipped. Closed under header references from kept classes plus the vtkmodules bundled imports. Removing an entry is zero-risk (you only add a wrapper back). Adding an entry needs a check that no kept Python path imports it.

3. Lever B — NOCOMPILE — cvista-config/_nocompile_classes.cmake (~742 classes) Dropped entirely: no compile, no wrapper, no hierarchy. Closed under C++ source references, transitive ::New() bases, and generated ObjectFactory registrations. Adding an entry needs closure analysis (nothing kept may reference it, directly or via factory). Removing an entry is safe.

The promotion hooks for both levers live in CMake/vtkModule.cmake.

Restoring a class (the common case)

A PyVista path needs vtkFooBar, but it was trimmed and import fails or a filter is missing.

1. Find where it was dropped:

   grep -rn vtkFooBar cvista-config/_nowrap_classes.cmake cvista-config/_nocompile_classes.cmake
   

2. Two ways to restore:
   • Permanent (the right fix when PyVista genuinely needs it): delete the line from the list. From _nowrap_classes.cmake is zero-risk. From _nocompile_classes.cmake, also restore any bases / referenced classes it needs so the closure holds, then prove a fresh configure + build + import.
   • Ad-hoc / to test (no canonical-list edit): pass it at configure time. This is the supported override:

     cmake ... -DCVISTA_KEEP_CLASSES="vtkFooBar;vtkBazQux"
     

     CVISTA_KEEP_CLASSES subtracts each class from both deny-lists, so it compiles and wraps, and leaves the canonical lists untouched. Set via env for build-cvista.sh. (Added in #106.)
3. Validate: configure in a fresh build dir, build, python smoke-cvista.py, then the bitexact gate if the class affects a filter path.

Dropping a new class to shrink further

1. Confirm nothing kept references it: search C++ sources for the type, check for ::New() callers and ObjectFactory registrations, and confirm no kept Python path imports it.
2. NOWRAP first (cheap, reversible). Only move to NOCOMPILE once you have proven the full closure, since NOCOMPILE removes the hierarchy too.
3. Add the entry to the appropriate list (append-only).
4. Validate in a fresh build dir (a dirty cache hides generate-time breakage): configure, build, smoke-cvista.py, then the PyVista regression suite — the bar is zero new failures vs stock 9.6.2.

Enabling a whole module

Edit cvista-config/_modules_minimal.cmake to WANT/YES the module. Watch for the WANT silent-drop cascade: forcing a module to NO silently removes its dependents while configure still succeeds, so a missing module can surface far from where you turned it off. Re-run a fresh configure and read the module report.

Traps (read before extending)

• NOCOMPILE filters classes, not files. A class listed in a module's SOURCES or TEMPLATES still compiles; do not delete its .cxx/.h to "help".
• Not every Testing/ directory is deletable; some carry a vtk.module or are referenced at the top level.
• Structurally-required disabled modules (MPI, Catalyst, WebGPU, Java, SerializationManager) must stay declared even though they are off.
• Source pruning does not change which classes compile (NOCOMPILE already excluded them), so a configure + compile + import-smoke is sufficient to prove a prune; the full proof is still the PyVista suite + bitexact gate.