Ping on prod jumped to seconds because node's main thread was busy most
of every second.
- accounts.save(account, ...) names the changed accounts, which are
written within a second; the rest are looked at a few at a time
(10 ms a save for those in play, 5 ms for the others). One save under
prod-like load: 289 ms before, 16 ms after.
- an asset the CDN does not have gets a cacheable 404 and is logged once
- the proxy's own connections no longer count against the web cap, which
had node reset requests nginx had already let in
- a repeated client error is written once in ten minutes; the panel's
log buffer trims in batches
- the daily backup uses sqlite's online backup off the main thread