Open Stack

On Thu, Dec 25, 2014 at 01:52:43PM +0400, Sergey Kraynev wrote:
>    Hi all.
>    In the last time we got on review several patches, which removes old
>    deprecation properties [1],A 
>    and one mine [2].
>    The aim is to delete deprecated code and redundant tests. It looks simple,
>    but the main problem, which we met, is backward compatibility.
>    F.e. user has created resource (FIP) with old property schema, i.e. using
>    SUBNET_ID instead of SUBNET. On first look nothing bad will not happen,
>    because:

FWIW I think it's too soon to remove the Neutron subnet_id/network_id
properties, they were only deprecated in Juno [1], and it's evident that
users are still using them on icehouse [2]

I thought the normal deprecation cycle was at least two releases, but I
can't recall where I read that.  Considering the overhead of maintaining
these is small, I'd favour leaning towards giving more time for users to
update their templates, rather then breaking them via very aggresive
removal of deprecated interfaces.

I'd suggest some or all of the following:

- Add a "planned for removal in $release" to the SupportStatus string
  associated with the deprecation, so we can document the planned removal.
- Wait for at least two releases between deprecation and removal, and
  announce the interfaces which will be removed in the release notes for
  the release before removal e.g:
    - Deprecated in Juno
    - Announce planned removal in Kilo release notes
    - Remove in "J"

[1] https://review.openstack.org/#/c/82853/
[2] http://lists.openstack.org/pipermail/openstack/2015-January/011156.html

>    1. handle_delete use resource_id and any changes in property schema does
>    not affect other actions.
>    2. If user want to use old template, he will get adequate error message,
>    that this property is not presented in schema. After that he just should
>    switch to new property and update stack using this new property.
>    In the same time we have one big issues for shadow dependencies, which is
>    actual for neutron resources. The simplest approach will not be worked
>    [3], due to old properties was deleted from property_schema.
>    Why is it bad ?
>    - we will get again all bugs related with such dependencies.
>    - how to make sure:A 
>    A  A  - create stack with old property (my template [4])
>    A  A  - open horizon, and look on topology
>    A  A  - download patch [2] and restart engine
>    A  A  - reload horizon page with topology
>    A  A  - as result it will be different
>    A 
>    I have some ideas about how to solve this, but none of them is not enough
>    good for me:
>    A - getting such information from self.properties.data is bad, because we
>    will skip all validations mentioned in properties.__getitem__
>    A - renaming old key in data to new or creating copy with new key is not
>    correct for me, because in this case we actually change properties
>    (resource representation) invisibly from user.
>    A - as possible we may leave old deprecated property and mark it something
>    like (removed), which will have similar behavior such as for
>    implemented=False. I do not like it, because it means, that we never
>    remove this "support code", because wants to be compatible with old
>    resources. (User may be not very lazy to do simple update or something
>    else ...)
>    - last way, which I have not tried yet, is usingA _stored_properties_data
>    forA extraction necessary information.
>    So now I have the questions:
>    Should we support such case with backward compatibility?A 
>    If yes, how will be better to do it for us and user?
>    May be we should create some strategy forA removingA A deprecated
>    properties ?

Yeah, other than the process issues I mentioned above, Angus has pointed
out some technical challenges which may mean property removal breaks
existing stacks.  IMHO this is something we *cannot* do - folks must be
able to upgrade heat over multiple versions without breaking their stacks.

As you say, delete may work, but it's likely several scenarios around
update will break if the stored stack definition doesn't match the schema
of the resource, and maintaining the internal references to removed or
obsolete properties doesn't seem like a good plan long term.

Could we provide some sort of migration tool, which re-writes the
definition of existing stacks (via a special patch stack update maybe?)
before upgrading heat?

Steve