## Changes
This PR pays some tech debt by refactoring sync diff computation into
interfaces that are more robust.
Specifically:
1. Refactor the single diff computation function into a `SnapshotState`
class that computes the target state only based on the current local
files making it more robust and not carrying over state from previous
iterations.
2. Adds new validations for the sync state which make sure that the
invariants that downstream code expects are actually held true. This
prevents a class of issues where these invariants break and the
synchroniser behaves unexpectedly.
Note, this does not change the existing schema for the snapshot, only
the way the diff is computed, and thus is backwards compatible (ie does
not require a schema version bump).
## Tests
<!-- How is this tested? -->
## Changes
This change replaces usage of the `repofiles` package with the `filer`
package to consolidate WSFS code paths.
The `repofiles` package implemented the following behavior. If a file at
`foo/bar.txt` was created and removed, the directory `foo` was kept
around because we do not perform directory tracking. If subsequently, a
file at `foo` was created, it resulted in an `fs.ErrExist` because it is
impossible to overwrite a directory. It would then perform a recursive
delete of the path if this happened and retry the file write.
To make this use case work without resorting to a recursive delete on
conflict, we need to implement directory tracking as part of sync. The
approach in this commit is as follows:
1. Maintain set of directories needed for current set of files. Compare
to previous set of files. This results in mkdir of added directories and
rmdir of removed directories.
2. Creation of new directories should happen prior to writing files.
Otherwise, many file writes may race to create the same parent
directories, resulting in additional API calls. Removal of existing
directories should happen after removing files.
3. Making new directories can be deduped across common prefixes where
only the longest prefix is created recursively.
4. Removing existing directories must happen sequentially, starting with
the longest prefix.
5. Removal of directories is a best effort. It fails only if the
directory is not empty, and if this happens we know something placed a
file or directory manually, outside of sync.
## Tests
* Existing integration tests pass (modified where it used to assert
directories weren't cleaned up)
* New integration test to confirm the inability to remove a directory
doesn't fail the sync run
shreyas-goenka
Pieter Noordhuis