Benji

This release pins two package dependencies to older versions as newer releases of these dependencies broke Benji.\
Fixes 49.\
Fixes 51.

Notable changes:

* Added a new I/O module `rbdaio` which uses the asynchronous API of `librbd`. Performance results in relation to `rbd`
have been mixed but performance should be at least 10-20% higher on restore. In one case performance has been
increased tenfold.

* Almost all Bash helper scripts have been rewritten in Python. The new scripts are calling Benji via the command line
just like before. This is intentional to minimize the interdependence between Benji and these helpers. The
scripts are examples only and not part of the API. There still is one example Bash script at `scripts/ceph.sh` to show
how to interact with Benji via Bash. The helpers have additional dependencies which can be installed with
`pip install benji[helpers]`.

* The Prometheus metrics exported by `benji-k8s` have changed:

* Backup metrics now longer include the `auxiliary_data` label.
* Command metrics now longer include the `arguments` label. The arguments have been folded into the `command` label.

* `benji-k8s`: The included scripts have been replaced by Python scripts and are using the new helper modules. They
should be calling compatible.

* `benji-k8s`: All calls to `kubectl` have been replaced with direct API requests. The official Python client for
Kubernetes is used. `kubectl` is still included in the image.

* Helm chart: Volumes and volume mounts are now configurable via `values.yaml`. This is mostly for getting the Ceph
credentials into the container but could also be used to mount file-based storage.

* Helm chart: The PostgreSQL chart dependency was updated from 2.7.6 to 4.2.2. This is the last chart which uses
PostgreSQL 10 and requires no upgrade of the database data structures.

* `benji-k8s` and Helm chart: The image was simplified to only include the Kubernetes specific scripts and `kubectl`.
Instead of running backups or other jobs via `crond` inside the container, the Helm chart now generates separate
`CronJob`s inside of Kubernetes. This is in preparation of the move to custom resources and an operator.

* An experimental and as of yet unfinished REST API has been added. The environment variable `BENJI_EXPERIMENTAL`
has to be set to `1` to enable the new `rest-api` subcommand. The API currently only services one request at a time,
which limits its usefulness. The REST API has additional dependencies, they can be installed with `pip install
benji[rest-api]`.

Notable changes:

* URL parsing of I/O resources is now conforming to standards. Especially for the RBD I/O module the two slashes
directly after the colon are no longer valid and have to be removed (`rbd://pool/image` -> `rbd:pool/image`).

* Added I/O module for iSCSI. It is based on `libiscsi` and requires no elevated permissions. Please see the
documentation as Benji requires a special version of the `libiscsi` Python bindings. The module is single-threaded
and synchronous, so performance will be limited. Contributions are welcome!

* The algorithm used by `benji enforce` has seen an overhaul and should be more comprehensible as the time categories
are based on natural time boundaries (start of the hour, day, week, month, and year) now.

* Added a restore helper script (`images/benji-k8s/scripts/benji-restore-pvc`) for Kubernetes. This script is intended
to be run on a management system with access to the Kubernetes cluster and can restore a version into a new or
an existing PVC/PV pair.

* The container images are now based on the Python 3.6 included in EPEL. The RBD support has been updated to Ceph
Nautilus. Nautilus also added RADOS and RBD Python bindings for Python 3.6 which are now used instead of building
them themselves.

Notable changes:

* Added `fsfreeze` support to the `benji-k8s` Docker image. Just add the `benji-backup.me/fsfreeze: yes` annotation to
the PVC. Kubernetes hosts are accessed via pods which are deployed by a DaemonSet, see the Helm chart for details.

* Use bulk inserts to speed up backups of images based on a previous version. This also decreases memory usage.

* Switched from in-memory block lists to an iterator based approach. This will increase performance and decrease
memory usage when backing up large images.

* Fixed a wrong index on the `blocks` table. This should also increase performance. The database will need to be
migrated with `benji database-migrate`.

* Laid the foundation for structured logging.

* Removed database table `stats` and assorted code and commands. Statistics are now kept together with the other
version metadata in the `versions` table. This means they are also removed when the version is removed. If
you want to keep historic statistics you need to export them beforehand with `benji -m ls` or
`benji metadata-export`. This is a breaking change and you might need to adjust your scripts. As statistics
are now included in a version's metadata the metadata version has changed to `1.1.0`. Old metadata backups
and exports with a metadata version of `1.0.0` can be imported by the current version. The statistics will
be empty in that case. The database will need to be migrated with `benji database-migrate`.

* Fixed a bug in the time calculation of `benji enforce` which could lead to a late expiration of versions,
the timing was a few hours off.

I'd like to thank olifre and adambmedent for their testing efforts!

Notable changes:

* Documentation updates
* Added new CLI command `benji storage-stats` to get storage usage information
* Fixed backup progress reporting on console
* Added `fdatasync()` calls to the `file` module to ensure backup integrity in case of a system crash or a
power failure (will impact performance when using the `file` I/O module)
* Speed up command line completion
* Added `benji completion` CLI command

Fixes a naming problem with the `--override-lock` CLI option of `benji rm` and `benji cleanup`