UPGRADE

upgrading monotone to 0.42
==========================

How to read this file:
  - Figure out what version of monotone you are upgrading from, let's
    say 0.22, for example.
  - Read down the list, and at each "version X or earlier" line, those
    instructions apply if X is larger than 0.22.  So, for instance,
    the "0.25 or earlier" instructions _and_ the "0.23 or earlier"
    instructions both apply.
  - If there are no entries that apply, then there is nothing to do.
    For instance, when upgrading from 0.23 to 0.24, no action was
    necessary.
  - No matter what, reading the release notes in NEWS is a good idea.
    Changes to things like command line switches, new commands, etc.,
    will be described there, not here.

If you are upgrading from:
  - 0.39 or earlier: The database format has changed.  You must run
    (after remembering to take a backup copy):
       $ mtn -d mydb.mtn db migrate
    on each of your databases.
  - 0.31 or earlier: The database format has changed.  You must run
    (after remembering to take a backup copy):
       $ mtn -d mydb.mtn db migrate
       $ mtn -d mydb.mtn db regenerate_caches
    on each of your databases.
  - 0.29 or earlier: The workspace format has changed.  You must run:
       $ mtn migrate_workspace
    in each of your workspaces.
  - 0.25 or earlier: Please read this section carefully.  We have
    modified the textual format used by revisions and manifests; old
    monotones cannot parse the new format, and new monotones cannot
    parse the old format, except to convert it to the new format.  See
    NEWS before upgrading.  Your project must rebuild its history;
    this is irreversible and creates a flag day -- everyone in your
    project must switch over at the same time.

    You will not be able to sync revisions between pre- and post-
    migration databases, and each project should only run the
    migration ONCE, on ONE database.  Pick a person to do the
    migration, they do it, once they're done, everyone else does a
    fresh pull from them.

    You should probably do some test migrations locally first to make
    sure everything seems to be working, before running your real
    migration.  To run the real migration:
      1) get everyone to commit and push their changes to the database
         chosen for migration (such as a central server if you use
         one); local dbs and working copies will be broken by this
         migration.
      2) shut down the server
      3) run 'cp server.db server.db.backup'
      4) run 'mtn db migrate --db server.db'
      5) run 'mtn db rosterify --db server.db'
      6) do whatever tests you like to make sure things look reasonable
      7) make sure that your server's firewall is set up to allow
         connections on the new netsync port: 4691
      8) start up the server
      9) tell everyone to do a new pull into a fresh database, using
         the new version of monotone.
    This migration is "best effort"; bugs in rename handling in the
    old code mean that some histories contain bogosities that cannot
    be migrated.  However, it makes an effort to preserve everything
    it can (including renames), and does guarantee that at each
    entry in the history graph, the tree layout and file contents are
    preserved exactly.

    The largest changes are that after upgrading, all revision ids
    will change, and, as a consequence, all certs must be reissued.
    This means that after upgrading, all certs will be signed by the
    person who performed the upgrade, rather than their original
    issuer.  This is a basic property of how signatures work, so
    there's little we can do about it.

    If you have any private branches, and are not your project's
    "designated migrator", then it is still possible to preserve your
    branches.  However, this requires some more understanding of what
    exactly is going on, and is outside the scope of this document.
    See http://monotone.ca/wiki/RosterifyingPrivateBranches/
    for one approach, and if you don't understand the logic behind
    those commands, then we're happy to explain things in more detail
    on the mailing list (monotone-devel@nongnu.org) or IRC (#monotone
    on irc.OFTC.net).

    We tentatively hope that this is the last time we will have to
    change the revision/manifest formats and have a flag day of this
    magnitude; but only experience will let us know for certain.

    If you have any questions or concerns about this process, please
    let us know via email (monotone-devel@nongnu.org) or discuss it on
    IRC (#monotone on irc.OFTC.net).

  - 0.23 or earlier: keys are now stored in ~/.monotone/keys (Unix,
    OS X), or %APPDATA%\Monotone\keys (Windows).  You must run 'db
    migrate' against each of your databases; this will automatically
    migrate the keys.  See NEWS for details.
    Command line syntax for 'serve' has changed; please adjust startup
    scripts accordingly.
    On Windows only, monotone now looks in a different place for the
    monotonerc file; see NEWS for 0.24 for details.

  - 0.21 or earlier: hooks governing netsync read permission have
    changed again; see NEWS for 0.22.

  - 0.19 or earlier: there are incompatible command line and server
    configuration changes; see NEWS for 0.20.

  - 0.18 or earlier: if you have created a ~/.monotonerc, rename it to
    ~/.monotone/monotonerc, so monotone will still find it.

  - 0.17: simply make a backup of your databases, just in case, and
    run "db migrate" on each.

  - 0.15 or 0.16: see below
  - 0.14 or earlier: see file README.changesets

upgrading from 0.15 or 0.16
---------------------------

there was still some residual badness in the revision graph code in
0.16.  we think we've caught it all now (we hope!), and there is now a
great deal more coordinated effort put into stopping any such thing
from sneaking in again, but... we have to do something about the
badness that's already there.

the solution is, another rebuild, like we did for the 0.15 -> 0.16
transition.  this is still obnoxious, still loses a little bit of
history information (every revision remains exactly reconstructable,
but rename information and information on who signed which certs are
both lost), and still requires coordination among your whole team to
pull of smoothly.  we're sorry.  we'll try not to do this again.

just in case we _do_ have to do it again, though, and to help make
this time smoother, we've added support for "epochs".  there's a whole
new section in the manual about all this, but basically, epochs are a
way to let monotone keep track of who's done a rebuild, so you can't
accidentally mix together pre-rebuild and post-rebuild databases.
(epochs will also come in handy when it's time to migrate away from
SHA1, for instance.)  this means you have more of a safety net than
last time, though you still have to coordinate within your team...

so, the basic procedure is: one designated person gets to perform the
rebuild, and deal with any missing files (see below).  everyone else
gets to 1) make sure the designated person has sync'ed with everyone,
because anything that's not in the designated person's database will
be lost, or at least, hard to deal with, 2) take a break, so the
designated person can rebuild and test and such without having to deal
with new commits.

if you're the designated person, then you should:
  1) make sure you have everyone's changes, and that they know they
     shouldn't make any more changes until you give the all-clear
  2) make a backup copy of your database.  seriously, do this.
        $ cp my.db my.db-backup
     if something goes wrong, and you don't have a backup, there may
     not be much we can do to help...
  3) dump your database to SQL text using your old version of
     monotone, and then reload it using your 0.17 version of
     monotone.  (this will migrate you from SQLite 2 to SQLite 3,
     which have different on-disk formats.)
        $ old-monotone --db=my.db db dump > my-db.dumped
        $ monotone --db=new.db db load < my-db.dumped
     (if you've been tracking the 0.17 development mainline, you can
     skip this step.  you still need to do all the others.)
  4) do the usual 'db migrate' command, for migrating a database to a
     later version:
        $ monotone --db=new.db db migrate
     (this will create the new tables needed for epoch support)
  5) rebuild your ancestry.  as mentioned above, this will lose
     renames and cert signing information (all certs that you trust
     will be re-issued with your signature; all other certs will be
     lost), but will also generate new changesets that actually make
     sense.
        $ monotone --db=new.db db rebuild
  6) check your database.  this is the rough equivalent of running a
     'fsck' or a 'scandisk' on your hard drive -- it just goes through
     and makes sure that everything looks good.  it's especially
     important for detecting missing files; see below.
        $ monotone --db=new.db db check 2>&1 | tee db-check.log
  7) look at db-check.log, and if any files were missing, note down
     their SHA1's, find the files, and load them into the database
     with "fload".
        $ monotone --db=new.db fload <missing-file
     after doing this, run 'db check' again, until you get a clean
     bill of health.  (if you see problems other than missing files
     and incomplete manifests/incomplete revisions, then there's
     something more gone wrong, and you might want to ask the list at
     monotone-devel@nongnu.org.)
  8) put the new database somewhere where people can netsync to it.
     (either by putting it there directly, or by creating a new
     database on your server and pushing to it.)
  9) tell everyone to create a new, empty database, and pull from your
     server.  (make sure they remember to copy their private keys over
     from their old database, using 'monotone privkey' and 'monotone
     read'.)

missing files
-------------

unfortunately, monotone 0.16 had a bug that may have led to loss of
data under certain, rare, circumstances.  the bug is that, 0.16's
"rebuild" (and probably "changesetify") commands incorrectly
constructed root revisions.  as a result, if you had any files that
were in the initial import of your project, and then later deleted,
without having had any changes made in between, then those files were
invisible to netsync, and not transferred by "push", "pull", or
"sync".  therefore, if you rebuilt your database at the 0.16 release,
and then pulled everything into a new database and deleted the
original database, monotone has lost its record of those files's
contents; if you attempted to check out a revision containing them you
would get an assertion error.

hopefully this should be a rare enough problem that most people don't
run into it.  the new 'db check' command detects this problem (indeed,
it's how it was discovered in the first place), so it should be easy
to find out whether you have it or not.  if you _do_ have it, you need
to find those missing files, and tell monotone about them.  this is
straightforward, though it may be tricky -- basically, 'db check' will
tell you the SHA1's of the missing data.  you need to find files that
have those SHA1's, and load them into your master database.  (that
would be the database that you just ran 'db rebuild', as above.)

some places to look:
  -- old releases of your software
  -- old copies of your monotone database (e.g., if you still have
     pre-changesetify backups of it)
     'cat file <SHA1>' is the useful command here
  -- the database that you originally ran 'rebuild' or 'changesetify'
     in, back at the 0.16 release.  since the problem only affects
     netsync, your initial database should be fine

getting the manifest of your initial revision may also be helpful,
since you can grep it to find the filenames of the missing files.  you
can do this by examining the output of 'cat revision <revision id>',
and then running 'cat manifest <manifest id>'.

once you've found the files, load them into your database with
'fload', e.g.:
  $ monotone --db=new.db fload <a-found-file

if you cannot find the relevant files anywhere, then there's a more
serious problem; your history is not fully reconstructible, and you
won't be able to use it with 0.17, because the bug is now fixed.  it
should still be possible to reconstruct all revisions after the
offending files were deleted, but this requires some code that doesn't
currently exist; if you actually are in this state, please email
<monotone-devel@nongnu.org> and we'll work something out.

in the future, the more rigorous checking monotone now does should
prevent problems like this from arising; in case you're nervous,
though, you can always run 'db check' to check for problems.