proton-drive-linux/docs/MILESTONE-3.0.0-STATUS.md

3.0.0: what is done and what is left

Tracks docs/MILESTONE-3.0.0.md. Update it with every commit that moves a phase.

Done

• Phase 0: fake Drive, simulation runs, CI job (2.9.0).
• Phase 1: queued-op preconditions and own sealed revisions in the database (2.9.0).
• Phase 2: local ids, inode = local id, one tree for every mount (2.10.0).
• Phase 3, first part: mkdir, create, rename, unlink and rmdir on the mount are recorded locally and sent from the queue; "local_first": false brings back the old path (5e0f30c).
• Fixes from the simulation and account runs since then: B125, B127, B130 and B141 to B180.
• Sync issues, first cut (§6.2): refusals sorted into issues, shown at once in the app, tray and CLI; Export button and pdfs sync export; pdfs sync issues.

Phase 3: left for 3.0.0

1. Sync issues (§6.2): first cut done.
   • The drain sorts a refusal from Drive into an issue kind (quota, access, missing, limit, rejected) and stores it on the queued op (schema 38).
   • An op with an issue counts as "needs attention" at once, not after six failures.
   • The queue listing carries the issue kind and whether its content can be exported.
   • Export a queued file's content (control request, pdfs sync export).
   • pdfs sync issues, and the issue shown in pdfs sync queue.
   • The app's Sync page shows each issue with what the user can do; Export and Retry buttons; translated in every catalog.
   • Discard an issue on purpose (after export, or for a change with no content), and undo the local change it stood for. Not in the first cut.
   • Show the issue on the node itself (file browser emblem or status). Not in the first cut.
2. Planner and executor (§5.4, §5.5): ops keyed by local id; remove mint_local_uid, finish_create and adopt_real_uid; a landing create sets remote_uid on its row.
   • Queued ops carry lid and parent_lid (schema 39). The claim reads the parent's uid through parent_lid, and a landing folder rewrites the ops inside it by local id.
   • Node rows link to their parent by lid (nodes.parent_lid, schema 40), and a restart places each node under its parent's row as it is now.
   • Queries over children, subtrees and ancestors by parent_lid, and no rewrite of the children's parent_uid when a folder lands (R3), nor of the ops made inside it (R2).
   • No more placeholders: a node made here goes by its local id until it lands. In four parts:
     • 6a: a node made here goes by local~<lid>, a function of its row, instead of a minted local~<ms>-<n>; schema 41 rewrites the old placeholders (nodes, ops, pins).
     • 6b: an op is found by its node's local id under either uid, the claim keeps one op per local id on the wire, and the drain sends the uid the row has now; remap_local_uid is gone.
     • 6c: children, subtrees, ancestors and ops' parents by lid, so a landing create rewrites no other row or op; what is read for a uid names the parent as its row has it now. The row keeps its stand-in in nodes.uid rather than NULL: it is as stable, and NULL would have changed some 40 node queries for no change in behaviour. Pins stay keyed by uid and move with the row in the landing transaction (B148).
     • 6d: a node written under its stand-in after it landed updates its landed row instead of bringing the placeholder row back, so a write that finds its create landed takes the uid from the row and no longer polls the tree for up to 2 s (landed_uid is gone). In memory, pending, the uploads and the tree stay keyed by the uid Drive knows: ops are matched by lid in the database (6b), hidden and creating only ever hold remote uids, and landed_row_uid is the one place at the mount's edge where a stand-in a handle still holds becomes the real uid.
   • An executor that runs ops in dependency order, many at once: 16 drain workers instead of 3, and a worker that lands an op wakes the idle ones, which slept through the ops it made claimable. The claim already keeps the order: a create waits for its parent's uid, a node's ops go one at a time. Ops stay blocking calls on worker threads: each one spends its time waiting on Drive, so tasks would buy nothing threads do not.
3. Speed (§7): run independent ops in parallel, batch trashes and moves; measure 1,000 files and 100 folders on Wi-Fi against LAN (target: within 10 %).
   • The measurement: a_thousand_files_drain_as_fast_on_wifi_as_on_lan in sim/daemon.rs, run with PDFS_SIM_MEASURE=1. It prints the times and asserts nothing yet.
   • Baseline on eb1c105, three drain threads: the syscalls take 3.7 s on LAN and 4.3 s on Wi-Fi; the drain takes 138 s on LAN and 938 s on Wi-Fi, 6.8 times as long.
   • Sixteen workers: the syscalls take 5.2 s on LAN and 5.0 s on Wi-Fi, which meets the target; the drain takes 47 s on LAN and 247 s on Wi-Fi. Most of a create on Wi-Fi is its read-back waiting for Drive to list the new node (250 ms, then 1 s), so batching (step 8) helps trashes and moves, not creates.
4. Done-when checks:
   • Simulation passes on the Wi-Fi profile: one_client_on_a_slow_link failed in CI on seeds 2 and 3. Two bugs (B127, B151) and a harness gap: a settle did not read the log, so it could not explain a known bug's damage. All three fixed; not yet confirmed in CI. Replays then found two races on the LAN profile too (B152, B153), both fixed. CI on 7269cb5 failed LAN seed 9, a regression from the B153 fix (B154), and Wi-Fi seed 2, a write whose answer was lost (B155). Both fixed. One slow-link run of seed 2 failed once with a rename answering ENOENT as a folder move landed. Not reproduced since; it may be B152. Replays of slow-link seed 4 failed two of six times on a file written again after its create lost its answer (B156, fixed). With sixteen drain workers (16fca10), seed 4 failed every replay on a write after an upload that was not read back (B157, fixed). Seed 3 then failed four of ten runs on two renames of one file landing out of order (B158, fixed). Flaky-link seed 4 failed one of ten runs on a lookup that answered EAGAIN while the root was listed again after a restart (B159, fixed). Slow-link seed 4 also had folders reach Drive 30 s late, after the drain's idle poll (B160, fixed). Slow-link seed 3 failed on a file moved while its create was on the wire, which the mount showed under its old name (B161, fixed). Seed 4 failed on a folder removed and made again, whose create adopted the removed folder (B162, fixed). CI on 903452a passed every profile. CI on f9b1638 failed slow-link seed 2: a step that unlinks an open file, appends and reads it back hung for 150 s. The B130 matcher did not fire, and CI kept no stacks; one replay passed. It may be a kernel notice sent from a runtime worker by another path than B130's, which a68dae9 (not yet pushed then) also takes off the runtime. CI now keeps a hung run's stacks.
   • Known-bug list in sim/run.rs is empty: none open; B125, B127, B129, B130 and B132 fixed but not yet confirmed by the runs.
   • A clean account run. The one on a6f36d5 (2026-10-04, LAN: 427 passed, 12 failed) failed B113 and B121 on every on-demand mount, two mirror-to-mirror move cases, and the journal check. B121 was a bug (B149, fixed); B113's case still expected creates to reach Drive at once and now accepts queued ones. The move cases were the harness: it retried every source after one was busy, and expected a refusal where a mirror to mirror move carries the local copy along. The journal showed a conflict copy retrying into a deleted folder (B150, fixed). Not confirmed yet. The run on 2.10.0 (2026-10-04: 260 passed, 3 failed) failed a new file that read back empty (B163, fixed), a 128-entry folder that listed 129 (B164, fixed), and its cleanup, which ran while the daemon was restarting: the stop waited on a busy thread until systemd killed it (B165, bounded). The next run on the working tree (2026-10-04: 259 passed, 4 failed) found a wide folder dropping files whose create landed while it was listed (B166, fixed), or listing one empty: the new file that read back empty again (B166 too). Its cleanup found no control socket because it ran 1.5 s after the last restart, before the new daemon had bound it: pdfs status reports that as not running, which the run's queue wait takes for an empty queue (the script's side). Looking into it found the start spending 1.7 s rebuilding listings (B167, fixed). Its journal check caught a file removed with its folder while uploading being sent again, to the root (B168, fixed), and a stop that let a mirror pass go on until the runtime ended under it (B169, fixed). CI on 9bbc994 failed one_client_on_a_good_link: a write opened as an upload landed took the revision it replaced as its base, and the next write failed with EIO (B170, fixed). The suite's seed 9 found a deleted file's trash waiting out its create's backoff (B171, fixed). Run three at a time on 2502af6, seed 8 failed two of nine runs: a file renamed while its upload was read back showed its old name again (B172, fixed). The loop on d850265 failed seed 12 once in nine runs: a file deleted just as its create landed stayed on Drive (B173, fixed). The suite on 378667f failed seed 6: a rename in the same gap was refused with EACCES (B174, fixed). The loop on 378667f failed seed 8: a file replaced by a rename in that gap stayed on Drive (B175, fixed), and seed 12: a write closed as its file's create landed was dropped (B176, fixed). The account run on 378667f: a folder listed as its create landed asked Drive by its stand-in and answered EIO (B177, fixed). A folder trashed while sync was paused kept a create below it as a trash: a paused drain worker had claimed it (B178, fixed). A file read as its create landed read empty (B179, fixed). The suite on 0f3c06f: a stop gave up on a drain worker that slept through its wake-up (B180, fixed). The account run on 903452a (2026-10-04, Wi-Fi: 262 passed, 1 failed) failed only its cleanup, which found no control socket: after the mirror target's last restart the queue wait took the missing daemon for an empty queue (the script's side, fixed in 6539c73).
5. Release: 2.10.0 is tagged; the 2.9.0 tag still waits for the user's go-ahead.

After 3.0.0

• Remove the old online path in filesystem.rs and the local_first switch (the release after 3.0.0).
• Phase 4: one applier for remote changes.
• Phase 5: mirror folders on the same planner and executor.
• Any time: Size and block geometry as types; a deadline on the remote calls outside the drain (the drain's have one: link::upload_deadline, Core::block_on_within).
• Open questions in §12 of the milestone.
• PLAN.md: the acceptance script notices network drops and reports interrupted instead of TIMEOUT.