[rasswanth-s]

Summary

Second part of step 1 of the V1 transport split, built on #9561. ConnectionRouter stops calling GDriveConnection directly and goes through the connections from #9561, implemented on Drive. What is read from and written to Drive does not change, apart from one fix to the order the event log is read in.

Changes

• syft/sync/connections/drive/drive_connections.py: DriveP2PConnection, DriveOwnerStore, DriveCollections, DrivePeering and DriveWatcherStore, and drive_transport() to build them. Each calls the GDriveConnection method used today. Where a contract builds a method from simpler ones, the Drive class overrides it to keep the one Drive call it makes today.
• DriveP2PConnection only sends for now: a proposed change goes to the owner's inbox, file events to the scientist's outbox, with the file name as msg_id, and the result is DELIVERED. Reading and archiving messages stay on the GDriveConnection methods. Doing them through receive and ack changes how they behave (one rule for both folders, delete on ack, no read cursor), which lands with send-and-forget. receive and ack raise NotImplementedError.
• DriveWatcherStore raises NotImplementedError until watcher storage on Drive is built.
• connection_router.py:
  • Sends go through p2p.send.
  • The event log, checkpoints and rolling state go through owner_store.
  • The peer list, peer requests, key bundles and version files go through peering.
  • Datasets go through collections.
  • The router holds a Transport built from connections[0], and copy_connection calls clone_for_thread() instead of checking isinstance(GDriveConnection). It imports nothing from Drive.
• base_connection.py: two hooks, as_transport() and clone_for_thread().
• gdrive_transport.py:
  • implements the two hooks;
  • takes PEERS_META_KEY, SYFT_PEERS_VERSION and the message timestamp parser from naming.py, so there is one copy (they still import from here);
• peer_manager.py: load_peers takes the peer's platform from the transport instead of creating GdriveFilesPlatform().
• Fix: the event log is read oldest first. Drive lists files by name, newest first, and the owner restores its state by replaying event-log messages in that order. A client that restored from the event log alone ended with the older version of a file that changed twice. owner_get_all_accepted_event_file_ids, owner_download_all_raw_events_from_syftbox and download_raw_events_since_timestamp now sort by message time.

Testing

• tests/unit/test_drive_connections.py (20 tests): each Drive connection on the mock Drive; both router sends go through p2p with the right kind; the router keeps its transport until its connection changes; the router source holds no Drive code; restoring from the event log keeps the newest version of a file changed twice. The restore test and the ordering test fail without the fix.
• tests/unit/test_peers_json_version.py: the fake router gives a transport instead of connection_for_send_message; the assertions are unchanged.
• The unit, syft-rds, syft-enclave, syft-job and syft-datasets suites pass, and so do the integration tests in mock mode. Two migration tests in test_dataset_multicopy_delivery.py fail on macOS, on dev as well (/tmp against /private/tmp).

Asana task

https://app.asana.com/1/1185126988600652/project/1210542925864934/task/1219496828587148

[pjwerneck]

Only minor issues. Looks good.