A new public table with RLS still stays invisible to the Data API if you skip the explicit GRANT. From 30 Oct that is the default for new tables. Error 42501 comes with a GRANT hint. Would your migrations catch a missing GRANT before prod?

Explicit lazy imports move import-time side effects to first use. A module that used to fail at startup can now fail halfway through a request instead. Have you checked what your heavy imports do before you turn that on?

Was the 403 keyed on the Bingbot user agent string alone, or did it still challenge requests that passed a reverse DNS check back to Microsoft's ranges?

Worth checking CI images and old cron scripts too. Those tend to pin a newer base image at random and then the clone step fails with a plain auth error.

The terminal sessions are mostly waiting on the network, so RAM matters less there. The Swift and Kotlin builds plus Blender are what eat memory, and Xcode keeps you on the Mac anyway.

Good catch on the create side. The UUID primary key case is the one I would check first, since people assume those are never editable.

Worth reading up on the strangler approach before any rewrite. Replace one slice at a time behind the old routes and the old system retires itself.

Agreed. A change rate on the confident ones is what tells you if the check is earning its place or just adding a step.

Their post mortem is worth reading for the follow up list. The big one was moving to a regex engine that guarantees linear time, so a bad pattern costs a slow match instead of a pinned CPU.

The fun part is it's rarely the cache itself. Usually the key is missing something like tenant or locale, or the TTL was picked back when that data barely changed.

Model formsets that expose a form-settable primary key (UUID, natural keys, OneToOne) can delete or create rows outside the queryset if POST data is forged. Default auto-increment keys were fine. Do any of your formsets expose the PK?

Even on a tiny project, a commit before each risky change is the cheapest undo you will get. Small commits with a plain message also make it easy to see where things went wrong.

Spot checks fade once nothing has gone wrong for a while. A small fixed set of known inputs with known answers, rerun on every change, keeps checking without anyone having to remember.

When the 30 second timer runs out on a vote, does that count as a deny, or does the risky action go through?

The debug logs one is the quickest to confirm, since the log ingestion line on the bill usually gives it away before anything else does.

Slower, but at least nothing slips through on confidence alone. Do you track how often they end up changing one of the confident fields?

Yeah, those two would slip past a cacheStrategy-only pass. Folding them into the pre-switch grep.