During the sync we collect a list of directories which should be empty
and attempt to rmdir them at the end of the sync. If the directories
are not empty then the rmdir will fail, logging a message but not
erroring the sync.
Multiple directories (up to --checkers worth) are scanned at once.
This uses much less memory than the previous scheme - only the amount
of memory needed to hold an entire directory listing of objects.
For directory based remotes the speed is unchanged.
For bucket based remotes, instead of doing one API call to list the
whole bucket, it does multiple calls, one for each pseudo directory.
However these are done in parallel so in practice this seems to speed
up directory listings.
This replaces the existing sync method as it performs faster and uses
less memory.
The old sync method is available with the temporary --old-sync-method
flag.
Fixes#517Fixes#439Fixes#236Fixes#1067
Optional interfaces are becoming more important in rclone,
--track-renames and --backup-dir both rely on them.
Up to this point rclone has used interface upgrades to define optional
behaviour on Fs objects. However when one Fs object wraps another it
is very difficult for this scheme to work accurately. rclone has
relied on specific error messages being returned when the interface
isn't supported - this is unsatisfactory because it means you have to
call the interface to see whether it is supported.
This change enables accurate detection of optional interfaces by use
of a Features struct as returned by an obligatory Fs.Features()
method. The Features struct contains flags and function pointers
which can be tested against nil to see whether they can be used.
As a result crypt and hubic can accurately reflect the capabilities of
the underlying Fs they are wrapping.
This commits adds support for tracking of file renames if `track-renames` flag is set,
and it then performs server-side renames for remotes that support it, i.e.
remotes that implement either the `Mover` or the `Copier` interface.
Fixes#888
Due to a logic error files stored on remotes which support modtime but
not hashes weren't being transferred when updating with a file of the
same size but different modtime. Instead the modtime of the remote
file was being set to that of the local file.
In practice this affected crypt with all remotes except Amazon Drive
and Dropbox.
* Make move command check for overlapping remotes and refuse to run
* Do copy/delete rather than all the copies then all the deletes
* Doesn't purge the source - this was unexpected behaviour see #512 and #416
* Add -list-retries flag to test suite to control retries
This changes the semantics of `move` slightly. However it now errs on
the side of not deleting stuff.
* Factor sync/copy/move into its own file
* Make fatal errors abort the sync
* Make Copy return errors
* Make Sync/Copy/Move return the last Copy error if there was one
* Prioritise returning Fatal errors
* NoRetry errors are returned if no other types of errors
