What's happening now

The filesystem needs some path validation stuffs independent of the
SVN path utilities.  A filesystem path is a well-defined Thing that
should be held a safe distance away from future changes to SVN's
general path library.


Incorrectnesses

We must ensure that node numbers are never reused.  If we open a node,
svn_fs_delete it, and then create new nodes, what happens when the
original node structure suddenly comes to refer to an entirely
different node?  Files become directories?

We should convert filenames to some canonical Unicode form, for
comparison.

Does everyone call svn_fs__check_fs who should?

svn_fs_delete will actually delete non-empty directories, if they're
not cloned.  This is inconsistent; should it be fixed?

Does every operation on a deleted node or completed transaction fail
gracefully?

Produce helpful error messages when filename paths contain null
characters.


Uglinesses

Fix up comments in svn_fs.h for transactions.

Add `public name' member to filesystem structure, to use to identify
the filesystem in error messages.  When driven by DAV, this could be a
URL.

When a dag function signals an error, it has no idea what the path of
the relevant node was.  But node revision ID's are pretty useless to
the user.  tree.c should probably rewrap some errors.

svn_fs__getsize shouldn't rely on a maximum value for detecting
overflow.

The use of svn_fs__getsize in svn_fs__parse_id is ugly --- what if
svn_vernum_t and apr_size_t aren't the same size?

Consider some macros or accessory functions for referencing the pieces
of the NODE-REVISION skel (instead of seeing stuff like
node->children->next->next and such other unreadable rubbish)


Slownesses

We don't store older node revisions as deltas yet.

The delta algorithm walks the whole tree using a single pool, so the
memory used is proportional to the size of the target tree.  Instead,
it should use a separate subpool every time it recurses into a new
directory, and free that subpool as soon as it's done processing that
subdirectory, so the memory used is proportional to the depth of the
tree.

We should move as much real content out of the NODE-REVISION skel as
possible; the skels should be holding only small stuff (node kind,
flags).
- File contents and deltas should be moved out to a `contents' table.
  The NODE-REVISION skel should simply contain a key into that table.
- Directory contents should be moved out to a `directories' table,
  with a separate table entry for each directory entry.  Keys into the
  table should be of the form `NODE-ID ENTRY-NAME NODE-REVISION', and
  values should be node revision ID's, or the word `deleted'; to look
  up an entry named E in a directory whose node revision is N.R,
  search for the entry `N E x', where x is the largest number present
  <= R.
- Property lists should be moved out to a table `properties', indexed
  similarly to the above.  We could deltify property contents the
  same way we do file contents.


Amenities

Extend svn_fs_copy to handle mutable nodes.

Long term ideas:

- directory entry cache:
  Create a cache mapping a node revision id X plus a filename component
  N onto a new node revision id Y, meaning that X is a directory in
  which the name N is bound to ID Y.  If everything were in the cache,
  this function could run with no I/O except for the final node.

  Since node revisions never change, we wouldn't have to worry about
  invalidating the cache.  Mutable node objects will need special
  handling, of course.

- fulltext cache:
  If we've recently computed a node's fulltext, we might want to keep
  that around in case we need to compute one of its nearby ancestors'
  fulltext, too.  This could be a waste, though --- the access
  patterns are a mix of linear scan (backwards to reconstruct a given
  revision) and random (who knows what node we'll hit next), so it's
  not clear what cache policy would be effective.  Best to record some
  data on how many delta applications a given cache would avoid before
  implementing it.

- delta cache:
  As people update, we're going to be recomputing text deltas for the
  most recently changed files pretty often.  It might be worthwhile to
  cache the deltas for a little while.

- Handle Unicode canonicalization for directory and property names
  ourselves.  People should be able to hand us any valid UTF-8
  sequence, perhaps with precomposed characters or non-spacing marks
  in a non-canonical order, and find the appropriate matches, given
  the rules defined by the Unicode standard.

Keeping repositories alive in the long term: Berkeley DB is infamous
for changing its file format from one revision to the next.  If someone
saves a Subversion 1.0 repository on a CD somewhere, and then tries to
read it seven years later, their chance of being able to read it with
the latest revision of Subversion is nil.  The solution:

- Define a simply XML repository dump format for the complete
  repository data.  This should be the same format we use for CVS
  repository conversion.  We'll have an import function.

- Write a program that is simple and self-contained --- does not use
  Berkeley DB, no fancy XML tools, uses nothing but POSIX read and
  seek --- that can dump a Subversion repository in that format.

- For each revision of Subversion, make a sample repository, and
  archive a copy of it away as test data.

- Write a test suite that verifies that the repository dump program
  can handle all of the archived formats.