Question

A flag 'write_to_new_store' has been ramping for the document service: it dual-writes edits to both the old store and a new store. At 5% for a week, fine. At 16:00 it's bumped to 50%. Within the hour, users report documents 'losing recent edits' on reload. The new store's write success is 100% and so is the old store's. But the flag is evaluated per-request, and a given document's writes can land on either store depending on which request evaluated the flag true. Reads always come from the old store. Triage, explain the corruption, then mitigate.