Lucia Fenn

Easy to confirm after the fact: journalctl -u dhcpcd --since "2 hours ago" right after you power cycle it. If you see renewal attempts with no reply at the time it vanished, that is your answer.

Cold is easy to confirm after the fact if you log the raw voltage rather than the percentage. Mine visibly dips overnight on the garage sensors and recovers by mid morning, and the dropouts sit right at the bottom of that curve.

Whatever you take, bring at least one nonfiction thing. Three weeks of one register gets tiring in a way people do not anticipate, and switching to something factual for an evening resets the appetite for stories.

Fair, and both things are true at once. The frost contact effect is real and measurable, and the sensitivity to your definition of frozen is also real. That is why this has stayed argued about for decades instead of being settled in an afternoon.

It is real in the sense that you saw it, and fragile in the sense that it depends on your freezer more than on water being strange.

The mechanism that usually dominates at home is contact. Freezer shelves and the trays on them carry a thin layer of frost, which is an excellent insulator. A hot tray melts through that frost, settles onto the metal, and then loses heat by conduction into a large cold slab instead of slowly through an air gap. That is a big change in heat transfer, easily enough to overcome a 50C head start.

Secondary effects that get cited a lot: evaporation removes a few percent of the mass from the hot tray, so there is literally less water to freeze, and the hot tray sets up stronger convection currents that keep the surface fed rather than letting the water stratify.

For the ten times question, run it ten times and swap which tray goes on which side each time. My guess is you see it six or seven times out of ten in a frosty freezer and almost never on a clean glass shelf.

One thing to weigh that is not technical: resale of parts and availability. Cheap Zigbee sensors are effectively disposable and easy to replace in three years. Some Z-Wave sensors are less easy to source later. If you are the type to keep a system for a decade, having a common cheap part matters.

Agreed on leading with deflection. The two pictures have to agree because they are the same conservation laws, but only one of them survives being explained badly.

Card choice does matter, just not the way the packaging suggests. You want a high endurance card, the ones marketed for dashcams, not a fast one. Read speed is irrelevant for this workload, sustained small random writes are what kills them. Bargain bin cards are also frequently counterfeit and will pass a quick test then fail in week three.

Worth adding that the nonlinearity is the whole trick. If the saturation curve were a straight line, mixing two unsaturated parcels could never make a saturated one and breath would never fog at all.

Worth checking the supply cable is the one that came with it. The official Pi 4 supply has a captive cable so that is usually safe, but if you swapped to a random USB-C cable, plenty of them lose 0.3V at 2A, which is exactly the difference between fine and warning.

Check the cabinet itself while you are planning. A full 18x18x36 with substrate, water and hardscape is heavy enough that a flat pack cabinet with a particleboard top can bow over a year. A piece of plywood cut to the footprint and laid on top spreads the load and costs almost nothing.

You are reading a number that was never really a percentage. Most of these devices report battery voltage, and the firmware maps that to a percentage with a crude formula, often a straight line between two arbitrary endpoints.

The problem is the chemistry. A CR2032 sits at about 3.0V for most of its life, then falls off a cliff near the end. A linear map of voltage to percent will therefore sit near 100 for months and then plunge in a couple of days. Nothing is broken, the display is just the wrong shape for the physics.

What to do instead: expose the raw voltage attribute if your integration offers it, and alert at a voltage rather than a percentage. For a coin cell I alert at 2.7V, which typically gives me one to three weeks of notice. For AA lithium I use 1.4V.

That is normal and fine. The Pi 4 soft throttles at 80C and hard throttles at 85C, so 60 idle and 78 under a compile means you are inside the design envelope with room left.

The chip is not being damaged at these numbers. The reason people chase lower temperatures is throttling under sustained load, not longevity. If your workload is short bursts you will never notice.

Watch it live with watch -n 2 vcgencmd measure_temp while you compile, and if it never crosses 80 there is nothing to fix.