On 07/02/2012 05:35 PM, Mike Kolesnik wrote:
Hi All,
I would like to hear opinions about a behaviour that I think is problematic in
REST API handling of logical networks.
-- Intro --
Today in the REST API we are exposing two collections for "logical
network" related entities.
First is a top level collection which is out of any context at the address
http://engine/api/networks.
Second is a sub-collection in the context of a cluster:
http://engine/api/cluster/xxx/networks
The network itself is defined per DC level, so for each DC you would have
at least one logical network for management, which has some properties such
as STP, MTU, etc..
The top level collection is used to create/delete such network entities.
The sub-collection in the context of a Cluster is used to attach/detach a
network from the DC of that cluster.
The network in the context of a cluster has some additional information, let's
say for example 'status' of the network:
If a network is defined on all hosts in the cluster then it's status is
'Operational'.
If a network is not defined on some of the hosts in the cluster then it's
status is 'Not Operational'[1].
-- Problem --
The problem is that details which are only relevant in context of a
cluster, are still displayed in the root context as well (e.g: 'status').
this is 2.2 bug, behaviour has to be the same as in /api/storagedomains
and /api/datacenters/xxx/storagedomains - 'status' should be addressed indeed.
(only use-case for status in /api/networks is when network yet not attached
to any cluster)
This can, in certain cases, cause unexpected behaviour.
For example, let's consider this topology:
Data Center A
|
|\____ Network 'red'
|\____ Cluster A1
| \______ Network 'red' attached
\____ Cluster A2
\______ Network 'red' attached
If the 'status' is the same on all the clusters that the network is attached to
(A1, A2) then there will be one element in the top level collection, with the
network details and the 'status' field representing the state (which is same
for all networks in the cluster contexts of the cluster).
If, however, the status is not the same (ie. on A1 the network is
'Operational' and on A2 it is 'Non Operational') then the top-level
collection will show two elements for the network, where all network
details are the same and only the 'status' field is different.
This is problematic IMHO for several reasons:
1. Showing one network in certain states, and multiple copies of this
network in other states is not optimal, to say the least.
2. In the top-level collection there is no indicator of the cluster for which
the network is displayed, so there is no way to differentiate between the
two 'red' network elements (they will have same id, name, etc.).
3. There is a certain asymmetry between the remove action[2] and the
result in that you would expect: you either remove a network but in the
result you would see several elements removed.
this is not correct, remove from /api/networks is not considered as *detach* from
clusters as well, but only as remove from /api/networks which is deletes network
from the DC, so it can't be 1:N,
if network still attached to cluster/s, backend should block this operation by CDA,
this is not api prerogative anyway.
-- Proposed Solutions --
Personally I can think of several solutions to this problem:
1. Declare the top-level collection as a collection of all networks that are
either attached to cluster or not, and if they are indeed attached then
show the details for each cluster, including a link to the cluster.
2. Declare the top-level collection as a collection of all networks that are
defined in data-centers, but they will not contain any cluster specific
data, and thus each entry is unique.
Solution #2 is breaking the API backwards-compatibility,
since it includes
removing certain fields that have appeared today (namely 'status' and
'display') but IMO would give a better experience since the top-level
collection is actually used for managing networks, and not their attachment
to clusters which should be done in the context of each cluster.
i vote for #2, that's the right way to go forward (but only if leaving properties
in /api/networks creates logical collisions as in 'status' case)
as for now it's only: 'status', 'display', right?
I would like to hear what suggestions you have to solve this problem or if
you prefer either of the above solutions.
-- Footnotes --
[1] In 3.1 this is slightly different, but for the sake of simplicity I didn't
specify the new behaviour.
[2] Currently you can't update the network if it's attached to any cluster,
but perhaps in the future this would be possible.
Regards,
Mike
--
Michael Pasternak
RedHat, ENG-Virtualization R&D