A daily puzzle has a strange data profile: one row per day, read by a lot of people in the same hour, written once. That is about the friendliest workload a database can have, which is a good reason not to use one.
The stack I ended up with is a static site, a precomputed puzzle file per day, and a tiny amount of client-side logic. Here is the shape of it.
Puzzles are generated offline. A build script picks the target, finds a route of the intended length, and writes a JSON file: the start word, the target, the accepted word set for that day, and the hint tiers. The file is small enough to inline for the current day and cache forever for past days.
Because the puzzle is a static asset, the daily round costs one request and no compute. The unlimited mode uses the same rule engine with a random seed instead of the current day's file, so practice never touches the daily data. Both modes share the validator, which is the part I want to be identical: a practice mode that accepts different moves than the daily is a bug factory.
The interesting constraint is midnight. Something has to decide which daily file is current, and doing that on the client means trusting the device clock. The compromise is to derive the day from the server's date header on first load, cache it, and revalidate on visibility change. Players who change their clock get a slightly different experience, which is not worth defending against.
The archive is where the static approach really pays off. One JSON file per day, immutable, cacheable, and searchable client-side. No migrations, no cold starts, and nothing to page in.
You can see the result on the daily puzzle and the unrestricted version at the unlimited practice mode. The part I would keep if I rebuilt it is the separation: generation is a build step, the rule engine is shared, and the only moving part at request time is static file serving.



