Code Review

Thanks for addressing the follow-up comments from PR #3031! I've reviewed the changes and overall they look good. Here are my findings:

✅ Strengths

1. Good Feature Addition: The new order parameter for build_matrix() adds useful flexibility for choosing between CSC and CSR sparse formats with proper validation.

2. Comprehensive Test Coverage: The new test_build_matrix_order() test is well-structured and covers:

   • Both CSC and CSR output formats
   • Consistency between formats (same values)
   • Invalid input validation
   • Both SCIPY and COO backends via parametrization

3. Improved Variable Naming: The rename from lhs/rhs in div() and rmul() improves clarity:

   • In div(): lhs → rhs (correct: operand is divisor)
   • In rmul(): lhs → rhs (correct: constant multiplier is on right in code flow)

4. Better Documentation: Enhanced docstrings with parameter descriptions and return type annotations.

5. Unnecessary Copy Removed: coo_backend.py:815 - removed redundant .copy() (already copying unconditionally on line 825).

6. Dead Code Cleanup: Removed unused pre-allocation arrays in coo_matmul() (lines 391-394).

🔍 Minor Issues & Suggestions

1. Inconsistent Type Hints (cvxpy/lin_ops/backends/base.py:199)

The change from:

self.variable_ids = variable_ids if variable_ids is not None else None

to:

self.variable_ids = variable_ids

This is cleaner, but could be even clearer. The conditional was redundant (no-op), so the removal is correct. However, consider if variable_ids should have a default empty set instead of None when not provided, to avoid potential None-checking throughout the codebase. If None is intentional, this is fine.

2. Return Type Annotation Consistency

In base.py, the return type uses union syntax:

sp.csc_array | sp.csr_array

This is Python 3.10+ syntax. Verify the minimum Python version requirement. If supporting older versions, use:

Union[sp.csc_array, sp.csr_array]

3. Test Coverage: Missing Rust Backend

The test_build_matrix_order() test only parametrizes over SCIPY_CANON_BACKEND and COO_CANON_BACKEND, but the Rust backend also implements the order parameter. Consider adding it to the parametrization:

@pytest.mark.parametrize("backend_name", [
    s.SCIPY_CANON_BACKEND, 
    s.COO_CANON_BACKEND,
    pytest.param(s.RUST_CANON_BACKEND, marks=pytest.mark.skipif(...))
])

🧪 Testing Recommendations

1. Edge Case: Empty Matrix: Test build_matrix() with empty constraint lists or zero-size matrices to ensure both CSC and CSR handle it correctly.

2. Integration Test: Consider adding a higher-level integration test that verifies CSR format works end-to-end in actual optimization problems (if not already covered elsewhere).

3. Performance Note: Document when CSR might be preferred over CSC (e.g., row-based operations), though this may be outside the scope of this PR.

🔒 Security & Performance

• ✅ No security concerns identified
• ✅ No performance regressions expected (the changes are additive with sensible defaults)
• ✅ Memory efficiency improved by removing unnecessary allocations

📝 Code Quality

• ✅ Follows existing code style
• ✅ Proper error handling with clear error messages
• ✅ Good documentation

Verdict

Approve with minor suggestions. The changes are well-implemented and improve code quality. The suggestions above are minor improvements for consideration but don't block merging.

Great work on the cleanup! 🎉

@SteveDiamond @Transurgeon can you review my commit? I had intended order to determine the flattening used for the first 2 tensor dimensions.

Benchmarks that have stayed the same:

   before           after         ratio
 [21bf98cf]       [9dc0dd38]
      293±0ms          315±0ms     1.08  slow_pruning_1668_benchmark.SlowPruningBenchmark.time_compile_problem
     38.9±0ms         40.8±0ms     1.05  matrix_stuffing.SmallMatrixStuffing.time_compile_problem
      1.10±0s          1.14±0s     1.04  gini_portfolio.Cajas.time_compile_problem
      11.4±0s          11.8±0s     1.04  simple_LP_benchmarks.SimpleLPBenchmark.time_compile_problem
      256±0ms          263±0ms     1.03  simple_QP_benchmarks.SimpleQPBenchmark.time_compile_problem
     14.2±0ms         14.4±0ms     1.01  simple_QP_benchmarks.ParametrizedQPBenchmark.time_compile_problem
      1.87±0s          1.88±0s     1.01  simple_QP_benchmarks.UnconstrainedQP.time_compile_problem
      934±0ms          942±0ms     1.01  simple_LP_benchmarks.SimpleScalarParametrizedLPBenchmark.time_compile_problem
      12.9±0s          13.0±0s     1.01  finance.CVaRBenchmark.time_compile_problem
      1.07±0s          1.07±0s     1.00  finance.FactorCovarianceModel.time_compile_problem
      4.97±0s          4.97±0s     1.00  optimal_advertising.OptimalAdvertising.time_compile_problem
      1.37±0s          1.37±0s     1.00  matrix_stuffing.ParamConeMatrixStuffing.time_compile_problem
      167±0ms          167±0ms     1.00  high_dim_convex_plasticity.ConvexPlasticity.time_compile_problem
      22.4±0s          22.3±0s     1.00  sdp_segfault_1132_benchmark.SDPSegfault1132Benchmark.time_compile_problem
      4.56±0s          4.55±0s     1.00  huber_regression.HuberRegression.time_compile_problem
      280±0ms          279±0ms     1.00  matrix_stuffing.ParamSmallMatrixStuffing.time_compile_problem
      345±0ms          343±0ms     0.99  gini_portfolio.Yitzhaki.time_compile_problem
      236±0ms          234±0ms     0.99  gini_portfolio.Murray.time_compile_problem
      676±0ms          668±0ms     0.99  matrix_stuffing.ConeMatrixStuffingBench.time_compile_problem
      5.17±0s          5.11±0s     0.99  svm_l1_regularization.SVMWithL1Regularization.time_compile_problem
     14.4±0ms         14.3±0ms     0.99  simple_LP_benchmarks.SimpleFullyParametrizedLPBenchmark.time_compile_problem
      2.85±0s          2.81±0s     0.98  quantum_hilbert_matrix.QuantumHilbertMatrix.time_compile_problem
      524±0ms          514±0ms     0.98  semidefinite_programming.SemidefiniteProgramming.time_compile_problem
      859±0ms          834±0ms     0.97  simple_QP_benchmarks.LeastSquares.time_compile_problem
      1.75±0s          1.64±0s     0.94  tv_inpainting.TvInpainting.time_compile_problem