I run about thirty of these little gateways on substations, same pattern.
Overlay on for the root filesystem, yes. Do your RaspAP changes with the overlay switched off, reboot, switch it back on. Annoying but rare.
The buffer needs a place that survives reboots and power loss: a second partition on the card (or a USB stick) mounted read-write with ext4 and its journal, outside the overlay. Point SQLite there, with PRAGMA journal_mode=WAL and synchronous=NORMAL. A cut then costs you at most the last few writes, not the database. Move mosquitto’s persistence there too, or turn persistence off since your forwarder already buffers.
On Rachel’s question: yes, if you stamp each reading with ts in UTC milliseconds when the ESP sends it (only if the Pi’s clock is synced, which it is once 4G is up, and keep a fake-hwclock so it isn’t 1970 after a power cut). Send the backlog in batches per child:
{"esp-bed3": [{"ts": 1767254400000, "values": {"soil_pct": 41}},
{"ts": 1767254460000, "values": {"soil_pct": 41}}]}Records land at their own time in history. A message whose every record is older than a minute counts as a replay and has to fit the profile’s replay rate, so a long backlog drains over a while. Don’t panic when it doesn’t all appear in one second. And wait for the PUBACK before deleting rows. See Devices behind a gateway.