Search that kept working as the repository count grew

After fifteen years the institute had roughly forty thousand repositories, and the practical consequence was that prior work could not be found.

So it was rewritten. Researchers reimplemented solvers that already existed two directories away — the most expensive failure mode in research software, and invisible on any dashboard.

Full-text search was indexed across every repository, file, commit message, comment and snippet, on its own storage volume so rebuilds never contend with repository I/O.

Results are filtered by the permissions of whoever is searching, which let the institute index restricted projects alongside open ones instead of excluding them.

People read what already exists before they start now. That was not true before, at any scale.

Mercurial repositories stayed on Mercurial for the legacy pipelines that depend on them, under the same permission model as the Git projects beside them.