simplesshd/rsync/rsync3.txt

-*- indented-text -*-

Notes towards a new version of rsync
Martin Pool <mbp@samba.org>, September 2001.


Good things about the current implementation:

  - Widely known and adopted.

  - Fast/efficient, especially for moderately small sets of files over
    slow links (transoceanic or modem.)

  - Fairly reliable.

  - The choice of running over a plain TCP socket or tunneling over
    ssh.

  - rsync operations are idempotent: you can always run the same
    command twice to make sure it worked properly without any fear.
    (Are there any exceptions?)

  - Small changes to files cause small deltas.

  - There is a way to evolve the protocol to some extent.

  - rdiff and rsync --write-batch allow generation of standalone patch
    sets.  rsync+ is pretty cheesy, though.  xdelta seems cleaner.

  - Process triangle is creative, but seems to provoke OS bugs.

  - "Morning-after property": you don't need to know anything on the
    local machine about the state of the remote machine, or about
    transfers that have been done in the past.

  - You can easily push or pull simply by switching the order of
    files.

  - The "modules" system has some neat features compared to
    e.g. Apache's per-directory configuration.  In particular, because
    you can set a userid and chroot directory, there is strong
    protection between different modules.  I haven't seen any calls
    for a more flexible system.


Bad things about the current implementation:

  - Persistent and hard-to-diagnose hang bugs remain

  - Protocol is sketchily documented, tied to this implementation, and
    hard to modify/extend

  - Both the program and the protocol assume a single non-interactive
    one-way transfer

  - A list of all files are held in memory for the entire transfer,
    which cripples scalability to large file trees

  - Opening a new socket for every operation causes problems,
    especially when running over SSH with password authentication.

  - Renamed files are not handled: the old file is removed, and the
    new file created from scratch.

  - The versioning approach assumes that future versions of the
    program know about all previous versions, and will do the right
    thing.

  - People always get confused about ':' vs '::'

  - Error messages can be cryptic.

  - Default behaviour is not intuitive: in too many cases rsync will
    happily do nothing.  Perhaps -a should be the default?

  - People get confused by trailing slashes, though it's hard to think
    of another reasonable way to make this necessary distinction
    between a directory and its contents.


Protocol philosophy:

   *The* big difference between protocols like HTTP, FTP, and NFS is
    that their fundamental operations are "read this file", "delete
    this file", and "make this directory", whereas rsync is "make this
    directory like this one".


Questionable features:

  These are neat, but not necessarily clean or worth preserving.

  - The remote rsync can be wrapped by some other program, such as in
    tridge's rsync-mail scripts.  The general feature of sending and
    retrieving mail over rsync is good, but this is perhaps not the
    right way to implement it.


Desirable features:

  These don't really require architectural changes; they're just
  something to keep in mind.

  - Synchronize ACLs and extended attributes

  - Anonymous servers should be efficient

  - Code should be portable to non-UNIX systems

  - Should be possible to document the protocol in RFC form

  - --dry-run option

  - IPv6 support.  Pretty straightforward.

  - Allow the basis and destination files to be different.  For
    example, you could use this when you have a CD-ROM and want to
    download an updated image onto a hard drive.

  - Efficiently interrupt and restart a transfer.  We can write a
    checkpoint file that says where we're up to in the filesystem.
    Alternatively, as long as transfers are idempotent, we can just
    restart the whole thing.  [NFSv4]

  - Scripting support.

  - Propagate atimes and do not modify them.  This is very ugly on
    Unix.  It might be better to try to add O_NOATIME to kernels, and
    call that.

  - Unicode.  Probably just use UTF-8 for everything.

  - Open authentication system.  Can we use PAM?  Is SASL an adequate
    mapping of PAM to the network, or useful in some other way?

  - Resume interrupted transfers without the --partial flag.  We need
    to leave the temporary file behind, and then know to use it.  This
    leaves a risk of large temporary files accumulating, which is not
    good.  Perhaps it should be off by default.

  - tcpwrappers support.  Should be trivial; can already be done
    through tcpd or inetd.

  - Socks support built in.  It's not clear this is any better than
    just linking against the socks library, though.

  - When run over SSH, invoke with predictable command-line arguments,
    so that people can restrict what commands sshd will run.  (Is this
    really required?)

  - Comparison mode: give a list of which files are new, gone, or
    different.  Set return code depending on whether anything has
    changed.

  - Internationalized messages (gettext?)

  - Optionally use real regexps rather than globs?

  - Show overall progress.  Pretty hard to do, especially if we insist
    on not scanning the directory tree up front.


Regression testing:

  - Support automatic testing.

  - Have hard internal timeouts against hangs.

  - Be deterministic.

  - Measure performance.


Hard links:

  At the moment, we can recreate hard links, but it's a bit
  inefficient: it depends on holding a list of all files in the tree.
  Every time we see a file with a linkcount >1, we need to search for
  another known name that has the same (fsid,inum) tuple.  We could do
  that more efficiently by keeping a list of only files with
  linkcount>1, and removing files from that list as all their names
  become known.


Command-line options:

  We have rather a lot at the moment.  We might get more if the tool
  becomes more flexible.  Do we need a .rc or configuration file?
  That wouldn't really fit with its pattern of use: cp and tar don't
  have them, though ssh does.


Scripting issues:

  - Perhaps support multiple scripting languages: candidates include
    Perl, Python, Tcl, Scheme (guile?), sh, ...

  - Simply running a subprocess and looking at its stdout/exit code
    might be sufficient, though it could also be pretty slow if it's
    called often.

  - There are security issues about running remote code, at least if
    it's not running in the users own account.  So we can either
    disallow it, or use some kind of sandbox system.

  - Python is a good language, but the syntax is not so good for
    giving small fragments on the command line.

  - Tcl is broken Lisp.

  - Lots of sysadmins know Perl, though Perl can give some bizarre or
    confusing errors.  The built in stat operators and regexps might
    be useful.

  - Sadly probably not enough people know Scheme.

  - sh is hard to embed.


Scripting hooks:

  - Whether to transfer a file

  - What basis file to use

  - Logging

  - Whether to allow transfers (for public servers)

  - Authentication

  - Locking

  - Cache

  - Generating backup path/name.

  - Post-processing of backups, e.g. to do compression.

  - After transfer, before replacement: so that we can spit out a diff
    of what was changed, or kick off some kind of reconciliation
    process.


VFS:

  Rather than talking straight to the filesystem, rsyncd talks through
  an internal API.  Samba has one.  Is it useful?

  - Could be a tidy way to implement cached signatures.

  - Keep files compressed on disk?


Interactive interface:

  - Something like ncFTP, or integration into GNOME-vfs.  Probably
    hold a single socket connection open.

  - Can either call us as a separate process, or as a library.

  - The standalone process needs to produce output in a form easily
    digestible by a calling program, like the --emacs feature some
    have.  Same goes for output: rpm outputs a series of hash symbols,
    which are easier for a GUI to handle than "\r30% complete"
    strings.

  - Yow!  emacs support.  (You could probably build that already, of
    course.)  I'd like to be able to write a simple script on a remote
    machine that rsyncs it to my workstation, edits it there, then
    pushes it back up.


Pie-in-the-sky features:

  These might have a severe impact on the protocol, and are not
  clearly in our core requirements.  It looks like in many of them
  having scripting hooks will allow us

  - Transport over UDP multicast.  The hard part is handling multiple
    destinations which have different basis files.  We can look at
    multicast-TFTP for inspiration.

  - Conflict resolution.  Possibly general scripting support will be
    sufficient.

  - Integrate with locking.  It's hard to see a good general solution,
    because Unix systems have several locking mechanisms, and grabbing
    the lock from programs that don't expect it could cause deadlocks,
    timeouts, or other problems.  Scripting support might help.

  - Replicate in place, rather than to a temporary file.  This is
    dangerous in the case of interruption, and it also means that the
    delta can't refer to blocks that have already been overwritten.
    On the other hand we could semi-trivially do this at first by
    simply generating a delta with no copy instructions.

  - Replicate block devices.  Most of the difficulties here are to do
    with replication in place, though on some systems we will also
    have to do I/O on block boundaries.

  - Peer to peer features.  Flavour of the year.  Can we think about
    ways for clients to smoothly and voluntarily become servers for
    content they receive?

  - Imagine a situation where the destination has a much faster link
    to the cloud than the source.  In this case, Mojo Nation downloads
    interleaved blocks from several slower servers.  The general
    situation might be a way for a master rsync process to farm out
    tasks to several subjobs.  In this particular case they'd need
    different sockets.  This might be related to multicast.


Unlikely features:

  - Allow remote source and destination.  If this can be cleanly
    designed into the protocol, perhaps with the remote machine acting
    as a kind of echo, then it's good.  It's uncommon enough that we
    don't want to shape the whole protocol around it, though.

    In fact, in a triangle of machines there are two possibilities:
    all traffic passes from remote1 to remote2 through local, or local
    just sets up the transfer and then remote1 talks to remote2.  FTP
    supports the second but it's not clearly good.  There are some
    security problems with being able to instruct one machine to open
    a connection to another.


In favour of evolving the protocol:

  - Keeping compatibility with existing rsync servers will help with
    adoption and testing.

  - We should at the very least be able to fall back to the new
    protocol.

  - Error handling is not so good.


In favour of using a new protocol:

  - Maintaining compatibility might soak up development time that
    would better go into improving a new protocol.

  - If we start from scratch, it can be documented as we go, and we
    can avoid design decisions that make the protocol complex or
    implementation-bound.


Error handling:

  - Errors should come back reliably, and be clearly associated with
    the particular file that caused the problem.

  - Some errors ought to cause the whole transfer to abort; some are
    just warnings.  If any errors have occurred, then rsync ought to
    return an error.


Concurrency:

  - We want to keep the CPU, filesystem, and network as full as
    possible as much of the time as possible.

  - We can do nonblocking network IO, but not so for disk.

  - It makes sense to on the destination be generating signatures and
    applying patches at the same time.

  - Can structure this with nonblocking, threads, separate processes,
    etc.


Uses:

  - Mirroring software distributions:

  - Synchronizing laptop and desktop

  - NFS filesystem migration/replication.  See
    http://www.ietf.org/proceedings/00jul/00july-133.htm#P24510_1276764

  - Sync with PDA

  - Network backup systems

  - CVS filemover


Conflict resolution:

  - Requires application-specific knowledge.  We want to provide
    policy, rather than mechanism.

  - Possibly allowing two-way migration across a single connection
    would be useful.


Moved files: <http://rsync.samba.org/cgi-bin/rsync.fom?file=44>

  - There's no trivial way to detect renamed files, especially if they
    move between directories.

  - If we had a picture of the remote directory from last time on
    either machine, then the inode numbers might give us a hint about
    files which may have been renamed.

  - Files that are renamed and not modified can be detected by
    examining the directory listing, looking for files with the same
    size/date as the origin.


Filesystem migration:

  NFSv4 probably wants to migrate file locks, but that's not really
  our problem.


Atomic updates:

  The NFSv4 working group wants atomic migration.  Most of the
  responsibility for this lies on the NFS server or OS.

  If migrating a whole tree, then we could do a nearly-atomic rename
  at the end.  This ties in to having separate basis and destination
  files.

  There's no way in Unix to replace a whole set of files atomically.
  However, if we get them all onto the destination machine and then do
  the updates quickly it would greatly reduce the window.


Scalability:

  We should aim to work well on machines in use in a year or two.
  That probably means transfers of many millions of files in one
  batch, and gigabytes or terabytes of data.

  For argument's sake: at the low end, we want to sync ten files for a
  total of 10kb across a 1kB/s link.  At the high end, we want to sync
  1e9 files for 1TB of data across a 1GB/s link.

  On the whole CPU usage is not normally a limiting factor, if only
  because running over SSH burns a lot of cycles on encryption.

  Perhaps have resource throttling without relying on rlimit.


Streaming:

  A big attraction of rsync is that there are few round-trip delays:
  basically only one to get started, and then everything is
  pipelined.  This is a problem with FTP, and NFS (at least up to
  v3).  NFSv4 can pipeline operations, but building on that is
  probably a bit complicated.


Related work:

  - mirror.pl http://freshmeat.net/project/mirror/

  - ProFTPd

  - Apache

  - http://freshmeat.net/search/?site=Freshmeat&q=mirror&section=projects

  - BitTorrent -- p2p mirroring
    http://bitconjurer.org/BitTorrent/