From: Daira Hopwood Date: Mon, 18 May 2015 16:25:45 +0000 (+0100) Subject: New version of design notes for remote-to-local sync. refs #2418 X-Git-Url: https://git.rkrishnan.org/?a=commitdiff_plain;h=29f337c746c66cb7833016c6fa22b1568f60c354;p=tahoe-lafs%2Ftahoe-lafs.git New version of design notes for remote-to-local sync. refs #2418 Signed-off-by: Daira Hopwood --- diff --git a/docs/proposed/magic-folder/remote-to-local-sync.rst b/docs/proposed/magic-folder/remote-to-local-sync.rst index a42051e9..096eb13a 100644 --- a/docs/proposed/magic-folder/remote-to-local-sync.rst +++ b/docs/proposed/magic-folder/remote-to-local-sync.rst @@ -25,6 +25,16 @@ design. .. _otf-magic-folder-objective4: https://tahoe-lafs.org/trac/tahoe-lafs/query?status=!closed&keywords=~otf-magic-folder-objective4 +*Glossary* + +Folder +DMD: distributed mutable directory +Descendant: a direct or indirect child in a directory or folder tree +Subpath +Subfolder +Object: a file or directory + + *Detecting remote changes* Unlike the local case where we use inotify or ReadDirectoryChangesW to @@ -51,99 +61,311 @@ write_coordination.rst So, in order to achieve the goal of allowing multiple users to write to a Magic Folder, we cannot implement the Magic Folder as a single Tahoe-LAFS -mutable directory. -Instead, we create one mutable Tahoe-LAFS directory per client. The +DMD. [FIXME reword to allow for design 6.] +[In designs 1 to 5 inclusive] Instead, we create one DMD per client. The contents of the Magic Folder will be represented by the union of these -directories. Each client polls the other directories in order to detect +client DMDs. Each client polls the other client DMDs in order to detect remote changes. +Six possible designs were considered for the representation of subfolders +of the Magic Folder: + +1. All subfolders written by a given Magic Folder client are collapsed +into a single client DMD, containing immutable files. The child name of +each file encodes the full subpath of that file relative to the Magic +Folder. + +2. The DMD tree under a client DMD is a direct copy of the folder tree +written by that client to the Magic Folder. Not all subfolders have +corresponding DMDs; only those to which that client has written files or +child subfolders. + +3. The directory tree under a client DMD is a ``tahoe backup`` structure +containing immutable snapshots of the folder tree written by that client +to the Magic Folder. As in design 2, only objects written by that client +are present. + +4. *Each* client DMD contains an eventually consistent mirror of all +files and folders written by *any* Magic Folder client. Thus each client +must also copy changes made by other Magic Folder clients to its own +client DMD. + +5. *Each* client DMD contains a ``tahoe backup`` structure containing +immutable snapshots of all files and folders written by *any* Magic +Folder client. Thus each client must also create another snapshot in its +own client DMD when changes are made by other . (It can potentially batch +changes, subject to latency requirements.) + +6. The write coordination problem is solved by implementing `two-phase +commit`_. Then, the representation consists of a single DMD tree which is +written by all clients. + +.. _`two-phase commit`: https://tahoe-lafs.org/trac/tahoe-lafs/ticket/1755 + +Here is a summary of advantages and disadvantages of each design: [TODO: +express this as a table with the properties as rows and the designs as +columns. It may be useful to simplify/merge some of the properties and +use footnotes for more detailed explanation.] + +Key: +++ major advantage ++ minor advantage +- minor disadvantage +-- major disadvantage +--- showstopper + +123456+: All designs have the property that a recursive add-lease +operation starting from the parent Tahoe-LAFS DMD will find all of the +files and directories used in the Magic Folder representation. Therefore +the representation is compatible with `garbage collection`_, even when a +pre-Magic-Folder client does the lease marking. + +.. _`garbage collection`: https://tahoe-lafs.org/trac/tahoe-lafs/browser/trunk/docs/garbage-collection.rst + +123456+: All designs avoid "breaking" pre-Magic-Folder clients that read +a directory or file that is part of the representation. + +456++: Only these designs allow a readcap to one of the client +directories --or one of their subdirectories-- to be shared with other +Tahoe-LAFS clients (not necessarily Magic Folder clients), so that such a +client sees all of the contents of the Magic Folder. Note that this was +not a requirement of the OTF proposal, although it is useful. + +135+: A Magic Folder client has only one mutable Tahoe-LAFS object to +monitor per other client. This minimises communication bandwidth for +polling, or alternatively the latency possible for a given polling +bandwidth. + +1-: If the Magic Folder has many subfolders, their files will all be +collapsed into the same DMD, which could get quite large. In practice a +single DMD can easily handle the number of files expected to be written +by a client, so this is unlikely to be a significant issue. + +35--: When a Magic Folder client detects a remote change, it must +traverse an immutable directory structure to see what has changed. +Completely unchanged subtrees will have the same URI, allowing some of +this traversal to be shortcutted. + +24---: When a Magic Folder client detects a remote change, it must +traverse a mutable directory structure to see what has changed. This is +more complex and less efficient than traversing an immutable structure, +because shortcutting is not possible (each DMD retains the same URI even +if a descendant object has changed), and because the structure may change +while it is being traversed. Also the traversal needs to be robust +against cycles, which can only occur in mutable structures. + +45--: When a change occurs in one Magic Folder client, it will propagate +to all the other clients. Each client will therefore see multiple +representation changes for a single logical change to the Magic Folder +contents, and must suppress the duplicates. This is particularly +problematic for design 4 where it interacts with the preceding issue. + +1236+: A client does not need to make changes to its own DMD that repeat +changesthat another Magic Folder client had previously made. This reduces +write bandwidth and complexity. + +4---, 5--: There is the potential for client DMDs to get "out of sync" +with each other, potentially for long periods if errors occur. Thus each +client must be able to "repair" its client directory (and its +subdirectory structure) concurrently with performing its own writes. This +is a significant complexity burden and may introduce failure modes that +could not otherwise happen. + +6---: While two-phase commit is a well-established protocol, its +application to Tahoe-LAFS requires significant design work, and may still +leave some corner cases of the write coordination problem unsolved. + +[Daira: +* designs 2 and 3 have no significant advantages over design 1, while +requiring higher polling bandwidth and greater complexity due to the need +to create subdirectories. They should be rejected. +* design 4 should be rejected due to the out-of-sync problem, which is +severe and possibly unsolvable for mutable structures. +* for design 5, the out-of-sync problem is still present but possibly +solvable. However, design 5 is substantially more complex, less efficient +in bandwidth/latency, and less scalable in number of clients and +subfolders than design 1. It only gains over design 1 on the ability to +share directory readcaps to the Magic Folder (or subfolders) which was +not a requirement, and IMHO could be better satisfied by design 6 in +future. +* design 6 is an unsolved research problem and should be considered out +of scope for the time being. We can benefit from experience with design 1 +when switching to design 6 later.] + *Conflict detection* -there are several kinds of dragon +there are several kinds of dragon [*] -earth dragons: alice changes 'foo' locally while alice's gateway is writing 'foo'. +earth dragons: write/download and read/download collisions + +alice changes 'foo' locally while alice's gateway is writing 'foo' +locally (in response to a remote change) alice's gateway * writes a temporary file foo.tmp -* if 'foo' is clean, i.e. there are no pending notifications, it renames foo.tmp over foo - -there is a race condition where the local write notification occurs concurrently with the rename, in which case we may clobber the local write. -it is impossible to detect this (even after the fact) because we can't distinguish whether the notification was for the rename or for the local write. -(assertion: the type of event doesn't help, because the local write may also be a rename --in fact should be for a maximally well-behaved app-- -and a rename event doesn't include the from filename. also Windows which doesn't support atomic rename-onto.) +* if 'foo' is clean, i.e. there are no pending notifications, it moves +foo.tmp over foo [FIXME: if we want to preserve old versions then it +should rename the old version first; see below] +there is a race condition where the local write notification occurs +concurrently with the move, in which case we may clobber the local write. +it is impossible to detect this (even after the fact) because we can't +distinguish whether the notification was for the move or for the local +write. +(assertion: the type of event doesn't help, because the local write may +also be a move --in fact should be for a maximally well-behaved app-- +and a move event doesn't include the from filename. also Windows which +doesn't support atomic move-onto.) this race has a small window (milliseconds or less) OR: alice's gateway * writes a temporary file foo.new -* if 'foo' is clean, i.e. there are no pending notifications, it renames foo to foo.old and then foo.new to foo - -(this would work on Windows; note that the rename to foo.old will fail if the file is locked for writing) +* if 'foo' is clean, i.e. there are no pending notifications, it moves +foo to foo.old and then foo.new to foo +(this would work on Windows; note that the rename to foo.old will fail if +the file is locked for writing. We should probably handle that case as a +conflict.) +TODO: on Unix, what happens wrt inotify events if we rename a file while +it is open? Does the filename for the CLOSE_WRITE event reflect the new +name? did the notification event for the local change precede the write? -air dragons: alice sees a change by bob to 'foo' and needs to know whether that change is an overwrite or a conflict +air dragons: write/upload collisions + +we can't read a file atomically. therefore, when we read a file in order +to upload it, we may read an inconsistent version if it was also bring +written locally. + +the magic folder is still eventually consistent, but inconsistent +versions may be visible to other users' clients, +and may interact with conflict/overwrite detection for those users +the queuing of notification events helps because it means that if files +are written more quickly than the +pending delay and less frequently than the pending delay, we shouldn't +encounter this dragon at all. + +also, a well-behaved app will give us enough information to detect this +case in principle, because if we get a notification +of a rename-to while we're reading the file but before we commit the +write to the Tahoe directory, then we can abort that write and requeue +the file to read/upload +(there is another potential race condition here due to the latency in +responding to the notification. We can make it very unlikely by pausing +after reading the file and before uploading it, to allow time to detect +any notification that occurred as a result of a write-during-read) + +we have implemented the pending delay but we will not implement the +abort/re-upload for the OTF grant + + +fire dragons: distinguishing conflicts from overwrites + +alice sees a change by bob to 'foo' and needs to know whether that change +is an overwrite or a conflict i.e. is it "based on" the version that alice already had +for the definition of "based on", we build on the solution to the earth +dragon -for the definition of "based on", we build on the solution to the earth dragon -when any client uploads a file, it includes Tahoe-side metadata giving the URI of the last remote version that it saved +when any client uploads a file, it includes Tahoe-side metadata giving +the URI of the last remote version that it saved before the notification of the local write that caused the upload -the metadata also includes the length of time between the last save and the notification; if this is very short, -then we are uncertain about whether the writing app took into account the last save (and we can use that information +the metadata also includes the length of time between the last save and +the notification; if this is very short, +then we are uncertain about whether the writing app took into account the +last save (and we can use that information to be conservative about treating changes as conflicts). - -so, when alice sees bob's change, it can compare the URI in the metadata for the downloaded file, with the URI that +so, when alice sees bob's change, it can compare the URI in the metadata +for the downloaded file, with the URI that is alice's magic folder db. -(if alice had that version but had not recorded the URI, we count that as a conflict. -this is justified because bob could not have learnt an URI matching alice's version unless [alice created that version -and had uploaded it] or [someone else created that version and alice had downloaded it]) - -alice does this comparison only when it is about to write bob's change. if it is a conflict, then it just creates a -new file for the conflicted copy (and doesn't update its own copy at the bare filename, nor does it change its +(if alice had that version but had not recorded the URI, we count that as +a conflict. + +this is justified because bob could not have learnt an URI matching +alice's version unless [alice created that version +and had uploaded it] or [someone else created that version and alice had +downloaded it]) + +alice does this comparison only when it is about to write bob's change. +if it is a conflict, then it just creates a +new file for the conflicted copy (and doesn't update its own copy at the +bare filename, nor does it change its magic folder db) +filesystem notifications for filenames that match the conflicted pattern +are ignored -filesystem notifications for filenames that match the conflicted pattern are ignored +water dragons: resolving conflict loops -fire dragons: resolving conflict loops +suppose that we've detected a remote write to file 'foo' that conflicts +with a local write +(alice is the local user that has detected the conflict, and bob is the +user who did the remote write) -suppose that we've detected a remote write to file 'foo' that conflicts with a local write -(alice is the local user that has detected the conflict, and bob is the user who did the remote write) alice's gateway creates a 'foo.conflict_by_bob_at_timestamp' file -alice-the-human at some point notices the conflict and updates hir copy of 'foo' to take into account bob's writes -but, there is no way to know whether that update actually took into account 'foo.conflict_by_bob_at_timestamp' or not -alice could have failed to notice 'foo.conflict_by_bob_at_timestamp' at all, and just saved hir copy of 'foo' again -so, when there is another remote write, how do we know whether it should be treated as a conflict or not? -well, alice could delete or rename 'foo.conflict_by_bob_at_timestamp' in order to indicate that ze'd taken it into account. but I'm not sure about the usability properties of that -the issue is whether, after 'foo.conflict_by_bob_at_timestamp' has been written, alice's magic folder db should be updated to indicate (for the purpose of conflict detection) that ze has seen bob's version of 'foo' -so, I think that alice's magic folder db should *not* be updated to indicate ze has seen bob's version of 'foo'. in that case, when ze updates hir local copy of 'foo' (with no suffix), the metadata of the copy of 'foo' that hir client uploads will indicate only that it was based on the previous version of 'foo'. then when bob gets that copy, it will be treated as a conflict and called 'foo.conflict_by_alice_at_timestamp2' +alice-the-human at some point notices the conflict and updates hir copy +of 'foo' to take into account bob's writes + +but, there is no way to know whether that update actually took into +account 'foo.conflict_by_bob_at_timestamp' or not +alice could have failed to notice 'foo.conflict_by_bob_at_timestamp' at +all, and just saved hir copy of 'foo' again +so, when there is another remote write, how do we know whether it should +be treated as a conflict or not? +well, alice could delete or rename 'foo.conflict_by_bob_at_timestamp' in +order to indicate that ze'd taken it into account. but I'm not sure about +the usability properties of that +the issue is whether, after 'foo.conflict_by_bob_at_timestamp' has been +written, alice's magic folder db should be updated to indicate (for the +purpose of conflict detection) that ze has seen bob's version of 'foo' +so, I think that alice's magic folder db should *not* be updated to +indicate ze has seen bob's version of 'foo'. in that case, when ze +updates hir local copy of 'foo' (with no suffix), the metadata of the +copy of 'foo' that hir client uploads will indicate only that it was +based on the previous version of 'foo'. then when bob gets that copy, it +will be treated as a conflict and called +'foo.conflict_by_alice_at_timestamp2' which I think is the desired behaviour -oh, but then how do alice and bob exit the conflict loop? that's the usability issue I was worried about [...] -but if alice's client does update hir magic folder db, then bob will see hir update as an overwrite +oh, but then how do alice and bob exit the conflict loop? that's the +usability issue I was worried about [...] +if alice's client does update hir magic folder db, then bob will see hir +update as an overwrite even though ze didn't necessarily take into account bob's changes which seems wrong :-( -(bob's changes haven't been lost completely; they are still on alice's filesystem. but they have been overwritten in bob's filesystem!) -so maybe we need alice to delete 'foo.conflict_by_bob_at_timestamp', and use that as the signal that ze has seen bob's changes and to break the conflict loop -(or rename it; actually any change to that file is sufficient to indicate that alice has seen it) - - -water dragons: - -we can't read a file atomically. therefore, when we read a file in order to upload it, we may read an inconsistent version. -the magic folder is still eventually consistent, but inconsistent versions may be visible to other users' clients, -and may interact with conflict/overwrite detection for those users -the queuing of notification events helps because it means that if files are written more quickly than the -pending delay and less frequently than the pending delay, we shouldn't encounter this dragon at all. -also, a well-behaved app will give us enough information to detect this case (in principle), because if we get a notification -of a rename-to while we're reading the file but before we commit the write to the Tahoe directory, then we can abort that -write and re-upload - -we have implemented the pending delay but we will not implement the abort/re-upload for the OTF grant - - - +(bob's changes haven't been lost completely; they are still on alice's +filesystem. but they have been overwritten in bob's filesystem!) +so maybe we need alice to delete 'foo.conflict_by_bob_at_timestamp', and +use that as the signal that ze has seen bob's changes and to break the +conflict loop +(or rename it; actually any change to that file is sufficient to indicate +that alice has seen it) + + +aether dragons: handling renames + +suppose that a subfolder of the Magic Folder is renamed on one of the +Magic Folder clients. it is not clear how to handle this at all: + +* if the folder is renamed automatically on other clients, then apps that +were using files in that folder may break. The behavior differs between +Windows and Unix: on Windows, it might not be possible to rename the +folder at all if it contains open files, while on Unix, open file handles +will stay open but operations involving the old path will fail. either +way the behaviour is likely to be confusing. +* for conflict detection, it is unclear whether existing entries in the +magic folder db under the old path should be updated to their new path. +* another possibility is treat the rename like a copy, i.e. all clients +end up with a copy of the directory under both names. effectively we +treat the move event as a directory creation, and also pretend that there +has been a modification of the directory at the old name by all other +Magic Folder clients. this is the easiest option to implement. other design issues: -* choice of conflicted filenames (e.g. foo.by_bob_at_YYYYMMDD_HHMMSS[v].type) -* Tahoe-side representation of per-user folders \ No newline at end of file +* choice of conflicted filenames (e.g. +foo.by_bob_at_YYYYMMDD_HHMMSS[v].type) + +[*] the association of dragons with the classical Greek elements +admittedly owes more to modern fantasy gaming than historically or +culturally accurate dragon mythology. consider them just as codenames for +now