scikit-tda

scikit-tda is a bundle. Installing it brings in ripser-py, persim, kmapper and tadasets together, on the premise that someone arriving at topological data analysis without a topology background should not have to work out which four packages they need before they can run anything. Started by Nathaniel Saul and Christopher Tralie in 2019, it names the organization that maintains all of them.

The design is worth stating plainly because it is the useful thing about it: each member stands alone. The bundle wraps independent libraries rather than vendoring them, so nothing here forces an all-or-nothing adoption, and a project that wants only diagram distances installs only persim.

What a metapackage does and does not promise

A bundle’s implicit claim is compatibility — that these versions of these packages work together. scikit-tda’s dependency list carries essentially no upper bounds, so the claim is weaker than the packaging suggests: it asserts these are the members, and leaves these versions cohere to the solver and to whatever happens to be current the day you install. That is normal for umbrella packages and it is not a defect, but it means the bundle is a convenience rather than a tested configuration, and a foundry that wants the latter has to pin the members itself.

The bundle also lags its members. It is a thin repository with far fewer commits than any package inside it, which is what you would expect and is worth knowing before reading its release date as a signal about the health of the stack.

The install that does not install

The one thing scikit-tda exists to do — bring the stack in with a single install — is the thing it cannot do from conda channels alone, because tadasets is on no channel. Every other member resolves; the smallest and least consequential one does not, and the umbrella inherits the gap.

That inversion is why the corpus put its packaging effort where it did. Submitting tadasets to conda-forge is a smaller piece of work than anything else in this family and it unblocks the whole bundle.

In this corpus

scikit-tda-environment is the fixture to reach for when the question is whether the ecosystem coheres rather than whether one engine is fast. It sits at L1 behind scikit-tda-recipe — which is the one recipe here that has never been built, so its status is plausible rather than verified — and behind tadasets-recipe, which has a pull request open. The techniques are persistent-homology and Mapper.