Page-first pagination for the submodule listing (#2404 follow-up)¶
Named as the structural follow-up in 2404-factor-resolution-indexes.md: the indexes there harden the factor-resolution branch, but the EXPLAINs on the issue showed the dominant cost is the emission aggregate, which touches every entry in the module to serve one page.
What changed¶
DataEntryRepository.get_submodule_data: when the requested sort/filter can
be decided from data_entries alone (_page_first_eligible), the page's
entry ids are resolved first (_page_first_entry_ids) and both the emission
aggregate and the per-row factor-resolution join are restricted to just
those ids instead of the whole module.
Ineligible whenever sort or filter reads something outside data_entries —
kg_co2eq (backed by the emission aggregate), a factor-backed sort/filter
column (detected the same way the existing count-query join guard already
does, by string-matching the rendered SQLAlchemy expression for factors/
building_rooms), or a handler default_where referencing either. Travel,
buildings and headcount entries keep their existing shape unconditionally —
their sorts read joined entities the ids-only query has no way to reproduce.
Why results cannot diverge¶
Both paths call the same _resolved_factor_id subquery and the same
emission-aggregate query shape — page-first only adds an id IN (...)
restriction to each. There is no second resolution implementation to drift.
test_page_first_equals_full_shape_across_pages and
..._with_filter assert byte-identical SubmoduleResponse output (items,
order, kg totals, resolved factor) between the page-first path and the
original shape forced on, across an offset boundary — the case where a
double-applied offset would silently return an empty or wrong page while
every same-page test stayed green.
test_page_first_actually_engages guards against the eligibility gate
silently going dead (returning None for a call that should qualify), which
would make every equivalence test above pass vacuously.
Deliberately not changed¶
- Travel/buildings/headcount branches, and any handler
default_whereor filter readingFactor/BuildingRoom— not touched, not optimized. - The count query's own conditional factor-join logic (already page-scoped correctly) — untouched.