A weekly analysis is a better fit for pandas than for PHP, and the integration should be the narrowest thing that works.
the boundary that works:
PHP writes a CSV to object storage and emits an event
Python reads it, writes a result CSV, emits an event
PHP reads the result
what this avoids:
a shared database schema
a shared ORM model in two languages
a deployment order between them
what it costs: latency, and a file format nobody owns.A file with a documented schema is a weaker coupling than a shared table and is much easier to reason about during an incident, because the artefact is inspectable. The failure mode is the format drifting silently — a column added on one side and ignored on the other — so a header check on read is worth the six lines. The latency is real and is the reason this suits batch work and nothing interactive.