Synnefo » snf-ganeti

=============================================================================

Management of storage types and disk templates, incl. storage space reporting

=============================================================================

.. contents:: :depth: 4

Background

==========

Currently, there is no consistent management of different variants of storage

in Ganeti. One direct consequence is that storage space reporting is currently

broken for all storage that is not based on lvm technolgy. This design looks at

the root causes and proposes a way to fix it.

Proposed changes

================

We propose to streamline handling of different storage types and disk templates.

Currently, there is no consistent implementation for dis/enabling of disk

templates and/or storage types.

Our idea is to introduce a list of enabled disk templates, which can be

used by instances in the cluster. Based on this list, we want to provide

storage reporting mechanisms for the available disk templates. Since some

disk templates share the same underlying storage technology (for example

``drbd`` and ``plain`` are based on ``lvm``), we map disk templates to storage

types and implement storage space reporting for each storage type.

Configuration changes

---------------------

Add a new attribute "enabled_disk_templates" (type: list of strings) to the

cluster config which holds disk templates, for example, "drbd", "file",

or "ext". This attribute represents the list of disk templates that are enabled

cluster-wide for usage by the instances. It will not be possible to create

instances with a disk template that is not enabled, as well as it will not be

possible to remove a disk template from the list if there are still instances

using it.

The list of enabled disk templates can contain any non-empty subset of

the currently implemented disk templates: ``blockdev``, ``diskless``, ``drbd``,

``ext``, ``file``, ``plain``, ``rbd``, and ``sharedfile``. See

``DISK_TEMPLATES`` in ``constants.py``.

Note that the abovementioned list of enabled disk types is just a "mechanism"

parameter that defines which disk templates the cluster can use. Further

filtering about what's allowed can go in the ipolicy, which is not covered in

this design doc. Note that it is possible to force an instance to use a disk

template that is not allowed by the ipolicy. This is not possible if the

template is not enabled by the cluster.

The ipolicy also contains a list of enabled disk templates. Since the cluster-

wide enabled disk templates should be a stronger constraint, the list of

enabled disk templates in the ipolicy should be a subset of those. In case the

user tries to create an inconsistent situation here, gnt-cluster should emit

a warning.

We consider the first disk template in the list to be the default template for

instance creation and storage reporting. This will remove the need to specify

the disk template with ``-t`` on instance creation. Note: It would be

better to take the default disk template from the node-group-specific

ipolicy. However, when using the iallocator, the nodegroup can only be

determined from the node which is determined by the iallocator, which in

turn needs the disk-template first. To solve this

chicken-and-egg-problem we first need to extend 'gnt-instance add' to

accept a nodegroup in the first place.

Currently, cluster-wide dis/enabling of disk templates is not implemented

consistently. ``lvm`` based disk templates are enabled by specifying a volume

group name on cluster initialization and can only be disabled by explicitly

using the option ``--no-lvm-storage``. This will be replaced by adding/removing

``drbd`` and ``plain`` from the set of enabled disk templates.

Up till now, file storage and shared file storage could be dis/enabled at

``./configure`` time. This will also be replaced by adding/removing the

respective disk templates from the set of enabled disk templates.

There is currently no possibility to dis/enable the disk templates

``diskless``, ``blockdev``, ``ext``, and ``rdb``. By introducing the set of

enabled disk templates, we will require these disk templates to be explicitly

enabled in order to be used. The idea is that the administrator of the cluster

can tailor the cluster configuration to what is actually needed in the cluster.

There is hope that this will lead to cleaner code, better performance and fewer

bugs.

When upgrading the configuration from a version that did not have the list

of enabled disk templates, we have to decide which disk templates are enabled

based on the current configuration of the cluster. We propose the following

update logic to be implemented in the online update of the config in

the ``Cluster`` class in ``objects.py``:

- If a ``volume_group_name`` is existing, then enable ``drbd`` and ``plain``.

(TODO: can we narrow that down further?)

- If ``file`` or ``sharedfile`` was enabled at configure time, add the

respective disk template to the list of enabled disk templates.

- For disk templates ``diskless``, ``blockdev``, ``ext``, and ``rbd``, we

inspect the current cluster configuration regarding whether or not there

are instances that use one of those disk templates. We will add only those

that are currently in use.

The order in which the list of enabled disk templates is built up will be

determined by a preference order based on when in the history of Ganeti the

disk templates were introduced (thus being a heuristic for which are used

more than others).

The list of enabled disk templates can be specified on cluster initialization

with ``gnt-cluster init`` using the optional parameter

``--enabled-disk-templates``. If it is not set, it will be set to a default

set of enabled disk templates, which includes the following disk templates:

``drbd`` and ``plain``. The list can be shrunk or extended by

``gnt-cluster modify`` using the same parameter.

Storage reporting

-----------------

The storage reporting in ``gnt-node list`` will be the first user of the

newly introduced list of enabled disk templates. Currently, storage reporting

works only for lvm-based storage. We want to extend that and report storage

for the enabled disk templates. The default of ``gnt-node list`` will only

report on storage of the default disk template (the first in the list of enabled

disk templates). One can explicitly ask for storage reporting on the other

enabled disk templates with the ``-o`` option.

Some of the currently implemented disk templates share the same base storage

technology. Since the storage reporting is based on the underlying technology

rather than on the user-facing disk templates, we introduce storage types to

represent the underlying technology. There will be a mapping from disk templates

to storage types, which will be used by the storage reporting backend to pick

the right method for estimating the storage for the different disk templates.

The proposed storage types are ``blockdev``, ``diskless``, ``ext``, ``file``,

``lvm-pv``, ``lvm-vg``, ``rados``.

The mapping from disk templates to storage types will be: ``drbd`` and ``plain``

to ``lvm-vg``, ``file`` and ``sharedfile`` to ``file``, and all others to their

obvious counterparts.

Note that there is no disk template mapping to ``lvm-pv``, because this storage

type is currently only used to enable the user to mark it as (un)allocatable.

(See ``man gnt-node``.) It is not possible to create an instance on a storage

unit that is of type ``lvm-pv`` directly, therefore it is not included in the

mapping.

The storage reporting for file storage will report space on the file storage

dir, which is currently limited to one directory. In the future, if we'll have

support for more directories, or for per-nodegroup directories this can be

changed.

For now, we will implement only the storage reporting for non-shared storage,

that is disk templates ``file``, ``lvm``, and ``drbd``. For disk template

``diskless``, there is obviously nothing to report about. When implementing

storage reporting for file, we can also use it for ``sharedfile``, since it

uses the same file system mechanisms to determine the free space. In the

future, we can optimize storage reporting for shared storage by not querying

all nodes that use a common shared file for the same space information.

In the future, we extend storage reporting for shared storage types like

``rados`` and ``ext``. Note that it will not make sense to query each node for

storage reporting on a storage unit that is used by several nodes.

We will not implement storage reporting for the ``blockdev`` disk template,

because block devices are always adopted after being provided by the system

administrator, thus coming from outside Ganeti. There is no point in storage

reporting for block devices, because Ganeti will never try to allocate storage

inside a block device.

RPC changes

-----------

The noded RPC call that reports node storage space will be changed to

accept a list of <storage_type>,<key> string tuples. For each of them, it will

report the free amount of storage space found on storage <key> as known

by the requested storage_type. Depending on the storage_type, the key would

be a volume group name in case of lvm, a directory name for the file-based

storage, and a rados pool name for rados storage.

Masterd will know through the mapping of storage types to storage calculation

functions which storage type uses which mechanism for storage calculation

and invoke only the needed ones.

Note that for file and sharedfile the node knows which directories are allowed

and won't allow any other directory to be queried for security reasons. The

actual path still needs to be passed to distinguish the two, as the type will

be the same for both.

These calculations will be implemented in the node storage system

(currently lib/storage.py) but querying will still happen through the

``node info`` call, to avoid requiring an extra RPC each time.

Ganeti reporting

----------------

`gnt-node list`` can be queried for the different disk templates, if they

are enabled. By default, it will just report information about the default

disk template. Examples::

  > gnt-node list

  Node                       DTotal DFree MTotal MNode MFree Pinst Sinst

  mynode1                      3.6T  3.6T  64.0G 1023M 62.2G     1     0

  mynode2                      3.6T  3.6T  64.0G 1023M 62.0G     2     1

  mynode3                      3.6T  3.6T  64.0G 1023M 62.3G     0     2

  > gnt-node list -o dtotal/drbd,dfree/file

  Node      DTotal (drbd, myvg) DFree (file, mydir)

  mynode1                 3.6T                    -

  mynode2                 3.6T                    -

Note that for drbd, we only report the space of the vg and only if it was not

renamed to something different than the default volume group name. With this

design, there is also no possibility to ask about the meta volume group. We

restrict the design here to make the transition to storage pools easier (as it

is an interim state only). It is the administrator's responsibility to ensure

that there is enough space for the meta volume group.

When storage pools are implemented, we switch from referencing the disk template

to referencing the storage pool name. For that, of course, the pool names need

to be unique over all storage types. For drbd, we will use the default 'drbd'

storage pool and possibly a second lvm-based storage pool for the metavg. It

will be possible to rename storage pools (thus also the default lvm storage

pool). There will be new functionality to ask about what storage pools are

available and of what type. Storage pools will have a storage pool type which is

one of the disk templates. There can be more than one storage pool based on the

same disk template, therefore we will then start referencing the storage pool

name instead of the disk template.

``gnt-cluster info`` will report which disk templates are enabled, i.e.

which ones are supported according to the cluster configuration. Example

output::

  > gnt-cluster info

  [...]

  Cluster parameters:

    - [...]

    - enabled disk templates: plain, drbd, sharedfile, rados

    - [...]

``gnt-node list-storage`` will not be affected by any changes, since this design

is restricted only to free storage reporting for non-shared storage types.

Allocator changes

-----------------

The iallocator protocol doesn't need to change: since we know which

disk template an instance has, we'll pass only the "free" value for that

disk template to the iallocator, when asking for an allocation to be

made. Note that for DRBD nowadays we ignore the case when vg and metavg

are different, and we only consider the main volume group. Fixing this is

outside the scope of this design.

With this design, we ensure forward-compatibility with respect to storage

pools. For now, we'll report space for all available disk templates that

are based on non-shared storage types, in the future, for all available

storage pools.

Rebalancing changes

-------------------

Hbal will not need changes, as it handles it already. We don't forecast

any changes needed to it.

Space reporting changes

-----------------------

Hspace will by default report by assuming the allocation will happen on

the default disk template for the cluster/nodegroup. An option will be added

to manually specify a different storage.

Interactions with Partitioned Ganeti

------------------------------------

Also the design for :doc:`Partitioned Ganeti <design-partitioned>` deals

with reporting free space. Partitioned Ganeti has a different way to

report free space for LVM on nodes where the ``exclusive_storage`` flag

is set. That doesn't interact directly with this design, as the specifics

of how the free space is computed is not in the scope of this design.

But the ``node info`` call contains the value of the

``exclusive_storage`` flag, which is currently only meaningful for the

LVM storage type. Additional flags like the ``exclusive_storage`` flag

for lvm might be useful for other disk templates / storage types as well.

We therefore extend the RPC call with <storage_type>,<key> to

<storage_type>,<key>,[<param>] to include any disk-template-specific

(or storage-type specific) parameters in the RPC call.

The reporting of free spindles, also part of Partitioned Ganeti, is not

concerned with this design doc, as those are seen as a separate resource.

.. vim: set textwidth=72 :

.. Local Variables:

.. mode: rst

.. fill-column: 72

.. End: