The design in f666699 was written before the code, and building it
changed a number of details, several of them after runs against the
real providers on the emulators. The doc now says what the code does:
- a sync run's steps in the order they run, the stale list, the
reconcile cursor, the owner check, the time margins, and accounts that
are left out when they refuse to list their collections;
- uploads in phases across accounts, the `ifInState` catch-up, account
refusals, series splits and copied events, creates adopting only in
their target collection, and poison markers;
- the contacts and calendar mappings as built (membership rows, text
limits, postal and relation rules, read-only rows and calendars,
exceptions, reminders under the old owner), and the columns the
providers refuse to return;
- the native module's scoping, the app's store, lifecycle and triggers,
and the failure matrix;
- decisions and limitations, with the questions still open.
Checked: every statement against the code of this branch.
The engine plans its provider reads and writes in JS, but only native
code can reach ContactsProvider and CalendarProvider. This adds that
bridge to the BulwarkDeviceSync module, with the rest of what the engine
and the app need from it:
- `query`, `applyBatch`, `readSyncState` and `readPhoto` take the
`ProviderQuery`/`ProviderOp[]` JSON of `src/device-sync/types.ts`,
check it (`sync/ProviderOps.kt`) and run it for one Android account of
our type (`sync/ProviderScope.kt`, `sync/ProviderIo.kt`). Every
selection gets the account ANDed in, every row an op names by id or
points to as a parent must be the account's, and every URI carries
`caller_is_syncadapter`, so nothing the sync writes marks rows dirty or
schedules an upload. The calls run one at a time on the module's own
thread through an unstable provider client, so a long batch never
stalls other module calls and a provider that dies fails the call
instead of the app.
- `showSyncProblem` and `clearSyncProblem` post and clear the one
notification per account that needs the user, e.g. "Sign in again".
- `setPushRoutes` stores which Android accounts each JMAP account feeds,
for the push router in `sync/PushRouting.kt`; the push services start
using it in a later commit.
- A run now ends exactly once and refuses a report that comes after
that. A sync requested while a run is open makes the run ask for
another sync when it ends, because SyncManager drops such requests as
duplicates. `onPerformSync` lets nothing escape any more: an exception
there killed the whole process, UI included.
- `ensureAccount` no longer hands an existing Android account to another
app account (`conflict`): its rows and SyncState belong to the first.
Payloads cross as JSON strings, which the bridge moves much faster than
nested maps. `src/device-sync/native.ts` is the typed JS side. The fake
provider follows the rules the emulator runs turned up: contacts
`data1`..`data14` read back as text, events and calendars the sync
adapter inserts have DIRTY NULL, attendee rows are addressed by id, an
op by id expects one row, and a vanished parent fails with `assert`
while another account's fails with `scope`.
Checked: 54 JVM tests (`gradlew :app:testDebugUnitTest`), debug and
release (R8) builds, `npm run typecheck` and `npm test` on this commit,
and on the API 32 emulator through the debug receiver: inserts with
back-references and group memberships, SyncState, photo writes and
scaled reads, asserts and expected counts, both scope refusals, the
499-op yield boundary, `tooLarge` for a 1.3 MB batch, and the
notification.
Issue #34 asks for contacts and calendars that show up in Android's own
apps and whose edits there reach the server. docs/device-sync.md is the
design the implementation follows: Android's sync framework drives the
runs, a thin Kotlin shell starts a headless JS task, and the engine is
TypeScript on a detached JMAP client, so token refresh, mTLS and session
discovery keep a single owner. A pure-Kotlin engine was rejected: it
would have had to redeem rotating refresh tokens outside that owner and
re-implement the client.
It pins down what every provider column holds, the JSContact and
JSCalendar mapping tables, how device edits are told apart from provider
normalisation (a baseline per row), how rows are matched when an editor
re-inserts them, the merge and conflict rules, uploads as patches with
`ifInState`, the native API, the failure matrix and the limitations.
The platform and server facts it relies on come from AOSP's providers
and calendar/contacts apps, Fossify, Etar and Stalwart v0.16.23's
source. These were checked on the local test server:
- map keys written through JMAP survive patches; `uid` duplicates are
refused; unknown and plural recurrence properties are rejected
- a start or `until` in a DST gap or overlap is silently dropped (a
daily rule came back unbounded); `Etc/GMT-2` keeps both
- `ifInState` yields `stateMismatch`; `relatedTo/<key>: null` does not
delete; result references to `/set` are refused
An adversarial review of the first draft found 20 paths that could lose
data, duplicate items or loop (own creates duplicated after a lost
response, unsynced edits dropped on deselection or eviction, Fossify
rewrites uploaded as edits, Etar moves turned into destroy + create,
and more); the design was changed for each before any engine code.
Linus Rath