Skip to content

CHANGELOG.md

Latest commit

 

History

History
176 lines (134 loc) · 8.03 KB

CHANGELOG.md

File metadata and controls

176 lines (134 loc) · 8.03 KB
Raw

Changelog — python-mip 2.0

Changes since 1.17.1

Performance

  • LinExpr.add_var and xsum hot paths optimised (single dict operation instead of two; bypass dispatch overhead for the common Var-only case), cutting 11–21% off constraint-heavy model build times.
  • HiGHS backend gained a CFFI C-array cache (matching CBC's bulk-add architecture) that accumulates pending columns/rows and flushes via Highs_addCols/Highs_addRows, plus name→index dicts to avoid unnecessary flushes on variable/constraint lookups.
  • Anonymous variables no longer pay the str.format/str.encode cost of auto-generated var(N) names.

A new reproducible benchmark suite (benchmarks/, documented in docs/bench.rst) was added to measure and track these improvements — comparing python-mip/CBC, python-mip/HiGHS, and python-mip/Gurobi model-build speed across CPython 3.14 and PyPy 3.11, against raw gurobipy and highspy (high-level and batch-numpy APIs), on four problems (n-Queens, TSP, RCPSP, and the new Capacitated Facility Location Problem). Highlights:

  • python-mip/Gurobi on PyPy is 2–4× faster than on CPython, depending on model structure, and remains the only way to drive Gurobi from PyPy at all (gurobipy has no PyPy wheel).
  • For sparse, variable-heavy models python-mip/Gurobi is competitive with or faster than raw gurobipy even on CPython; for dense-constraint models gurobipy's own loop is faster on CPython, but PyPy closes much of that gap.
  • Against highspy's own recommended vectorized numpy-array API, python-mip/HiGHS is competitive with or faster than it on sparser/irregular problems (n-Queens, TSP), while the vectorized API pulls ahead on dense, rectangular problems (e.g. CFLP).

New Features

Model.add_vars() batch variable creation

Model.add_vars(n, ...) / VarList.add_vars(n, ...) create n variables in a single call, avoiding per-call dispatch overhead compared to n calls to add_var().

New LP solve controls: TRUNCATED status, max_iter, lp_method, lp_preprocess

  • OptimizationStatus.TRUNCATED for LP solves stopped early by a time or iteration limit (populates primal/dual solution and objective bound whenever feasible, instead of reporting a misleading OPTIMAL/INFEASIBLE).
  • Model.max_iter caps simplex/barrier iterations for LP solves, wired through CBC, HiGHS and Gurobi.
  • Model.optimize(lp_method=...) selects dual simplex, primal simplex or IPM/barrier — added to HiGHS and Gurobi (CBC already had it).
  • New CBC-specific LP_Method.RACING and LP_Method.RECOMMEND values: RACING runs several LP configurations (dual simplex, primal/Idiot, primal/Sprint) in parallel threads and keeps the first to reach optimality (needs Model.threads >= 2; safely degrades to RECOMMEND with a log message if fewer threads are available); RECOMMEND uses CBC's ML-based per-instance LP method selection. Both are ignored (fall back to their own default) by HiGHS and Gurobi. LP_Method.AUTO (the default) already delegates to RECOMMEND when sequential and RACING when 2+ threads are set, so no behaviour changes for existing code — these are purely new, explicit opt-in values.
  • Model.optimize(relax=True, lp_preprocess=True) (CBC-specific) enables CBC's INT_PARAM_LP_FAST_PREPROCESS knapsack bound-tightening before an LP relaxation solve.

Infrastructure & Distribution

  • cbcbox bumped to >=2.935, improving binary reliability across platforms.
  • HiGHS backend migrated to highspy; added TRUNCATED status and max_iter/lp_method (dual/primal simplex, IPM/barrier) support, also ported to the Gurobi backend.
  • Python 3.14 added to the CI test matrix (alongside 3.10–3.13 and PyPy 3.11).

Bug Fixes

  • CBC string parameters migrated to the INT_PARAM API; fixed verbose reset on clear().
  • Removed an obsolete objective-sense save/restore workaround around Cbc_reset().
  • Stale highspy→highsbox error message corrected.

Previous release: python-mip 1.17 / 1.17.1

New Features

HiGHS Solver Support

python-mip now ships with full support for the HiGHS open-source solver as a first-class backend (alongside CBC and Gurobi). HiGHS is a high-performance solver for LP and MIP problems with a permissive MIT licence.

Key capabilities added:

  • Full LP and MIP solve via HiGHS C API (through highsbox)
  • Warm-start (basis handoff) for LP re-solves
  • relax=True support in optimize()
  • Variable and constraint inspection/modification
  • Correct handling of UNBOUNDED vs INFEASIBLE status
  • Reduced memory footprint and improved file read/write consistency

HiGHS is installed as an optional dependency: pip install mip[highs].

macOS Apple Silicon (M1/M2/M3) Native Support

CBC now runs natively on Apple Silicon via a pre-built ARM64 binary, replacing the previous Rosetta 2 x86_64 fallback.

Infrastructure & Distribution

CBC Binaries via cbcbox

The bundled CBC shared libraries (.so, .dylib, .dll) have been removed from the python-mip source tree. CBC binaries are now distributed through the cbcbox PyPI package, which provides pre-built wheels for:

  • Linux x86_64 and aarch64 (ARM64)
  • macOS x86_64 and arm64
  • Windows x64

cbcbox is a dedicated package whose sole job is to ship up-to-date CBC binaries for all major platforms. This decoupling means future CBC upgrades are released without touching python-mip itself. The minimum required version is cbcbox>=2.902.

Automated PyPI Publishing

A new GitHub Actions workflow (.github/workflows/publish.yml) automatically publishes to PyPI whenever a v* tag is pushed. It uses OIDC Trusted Publisher authentication — no API tokens to rotate.

Modernised CI Matrix

Platform OS
Linux x86_64 ubuntu-24.04
Linux aarch64 ubuntu-24.04-arm (new)
macOS ARM64 macos-15 (new)
Windows x64 windows-2025 (new)

Python versions tested: 3.10, 3.11, 3.12, 3.13, PyPy 3.11.

Bug Fixes

  • CBC re-solve correctness: A bug introduced by newer CBC versions caused stale solution data to be returned when optimize() was called multiple times on the same model. Fixed by calling Cbc_reset() before each Cbc_solve(), with objective sense saved and restored around the reset.
  • isfile import missing in SolverCbc.read(): os.path.isfile was used but not imported, causing a NameError when loading a model from a file.
  • Windows DLL loading: On Python 3.8+, Windows ignores PATH when resolving DLL dependencies. Fixed by calling os.add_dll_directory() on the cbcbox bin/ directory.
  • Empty LinExpr in constraints: Constraints containing an empty linear expression were not handled correctly. Fixed by Sebastian Heger (#237).

Breaking Changes / Compatibility

  • Minimum Python version raised to 3.10. Python 3.8 and 3.9 have reached end-of-life and are no longer tested or supported.
  • Bundled CBC libraries removed — cbcbox is now a required dependency (installed automatically via pip).
  • gurobipy version constraint relaxed to >=10 (no upper bound).
  • cffi version constraint relaxed to >=1.15 (no upper bound).
  • highsbox version constraint relaxed to >=1.10.0 (no upper bound).

Acknowledgements

This release was a team effort. Thank you to everyone who contributed:

  • Robert Schwarz — HiGHS interface: initial implementation (PR #332) and extensive improvements (PR #418), including objective setter fix, option types, test coverage and highsbox migration. Co-authored with Bernard Zweers and Miguel Hisojo.
  • Túlio Toffolo — macOS Apple Silicon support, HiGHS testing infrastructure, CI modernisation, and many quality-of-life fixes.
  • Sebastian Heger — Bug fix for constraints with empty linear expressions (#237).
  • Dominik Peters — Removed upper limit on supported Python versions (#408).
  • Adeel Khan — HiGHS _core library support.
  • Haroldo Santos — cbcbox integration, CBC bug fixes, CI/CD automation, and release management.