fix(#516): review round 10 - the change watch is one JobScheduler job that re-arms itself, startup reconcile can't crash the app

This commit is contained in:
makiolaj committed 2026-11-14 12:53:58 +01:00
1 parent a902f5f862
commit ad5f0f4662
6 files changed
+203 -186

No files matched your search

+23 -23
View File
@@ -485,29 +485,29 @@ of the backup keeps its last one. The pre-#411 fixed `calendula-backup.ics` is
the combined series' stable copy for the install that wrote it, and an
install that already had a folder stays on the combined layout.
**Backing up after a change (#516)** is opt-in. `BackupScheduler.watchChanges`
arms a one-off `BackupWorker` with a WorkManager content-URI trigger on the
provider's root URI (the provider notifies every change there, whoever made it:
this app, a sync adapter or another app), a 2-minute settle delay capped at 15
minutes, and battery-not-low. A trigger fires once and observes only while its
job waits, so there are two slots: a change run arms the other one as it starts
(catching changes made during the run) and checks again as it ends;
`watchStep` keeps exactly one watch waiting, queueing a new one behind a run
that still shows as running and cancelling a second one. Runs share the
scheduled backup's lock. A change run only rewrites the stable copies and never
retries: the dated files and their retention stay on the interval, so a day of
edits can't rotate the history out. Every run that writes something is recorded
on the status line; the failure notification comes once per failure streak, as
it reaches the threshold, and the next success of any run clears it. Since the
trigger fires for any provider change (reminders, other calendars, sync
bookkeeping), each series' file (events in a fixed order) is hashed without its
`DTSTAMP`s (`auto_backup_fingerprints`, reset with the folder and kept only
while the option is on) and a change run
skips the series whose hash (bound to the copy's name) and stable copy are
unchanged, without even listing the folder when no hash moved; the hash is
dropped before a copy is rewritten. A copy that went missing is restored by the
next run that lists the folder; one a still-loading listing doesn't show, or
one edited in place outside the app, is left to the next scheduled run.
**Backing up after a change (#516)** is opt-in. `BackupChangeJob` is a
JobScheduler job with a content trigger on the provider's root URI (the provider
notifies every change there, whoever made it: this app, a sync adapter or
another app), a 2-minute settle delay capped at 15 minutes. A trigger fires
once, so each firing queues a change run (a unique `BackupWorker`, battery not
low, queued behind one already running) and schedules the next watch under the
same id before it finishes, which carries over the changes seen meanwhile. Such
a job can't be persisted; `BackupScheduler.reconcile` re-arms it on every
process start (the boot receivers start one), without replacing a waiting watch.
Runs share the scheduled backup's lock. A change run only rewrites the stable
copies and never retries: the dated files and their retention stay on the
interval, so a day of edits can't rotate the history out. Every run that writes
something is recorded on the status line; the failure notification comes once
per failure streak, as it reaches the threshold, and the next success of any run
clears it. Since the trigger fires for any provider change (reminders, other
calendars, sync bookkeeping), each series' file (events in a fixed order) is
hashed without its `DTSTAMP`s (`auto_backup_fingerprints`, reset with the folder
and kept only while the option is on) and a change run skips the series whose
hash (bound to the copy's name) and stable copy are unchanged, without even
listing the folder when no hash moved; the hash is dropped before a copy is
rewritten. A copy that went missing is restored by the next run that lists the
folder; one a still-loading listing doesn't show, or one edited in place outside
the app, is left to the next scheduled run.
**A backup carries the series' deleted and moved occurrences.** `EXDATE` lives on
the master row and is part of `EventExportProjection`; `IcsWriter` marks an