Thank you for the response. I totally understand the effort and your problems with getting it through at the time. Your design is completely inline with what's currently present in Nova for EC2, no doubt about that. I did whatever I could to review your patches and consider if it's worth to go forward with them in current circumstances. I believe it'll add more complications than value if we go on with them.

The main design problems, introduced before you, won't go away:
1. User isolation for shared objects is close to impossible to implement in this model. 2. Marked resources listing when describing tags, will require going to all of the possible different APIs and their databases eventually and then compiling the result.

The stackforge/ec2-api implementation fortunately had no constraints or previously implemented code with some conceptual problems. So it could and did store the tags and their associations with resources separately. It allowed efficient searching for both describing tags and resources.

Strategically if, as I understand, eventual aim is to switch to separate ec2-api solution, it makes little sense (to me) to add more functionality, especially partial functionality (no describe_volumes, describe_snapshots and even if added, no tagging for other resources), to current nova code. If the decision was made to enhance nova with new features like this, I'd still be for a separate table in DB for all of the tags and their associations - it would've made universal, complete and efficient solution with one effort.

And again, I more than agree with this:
...I can only wish that the patches got more attention when it was possible to get them merged :)

But that's a different story.

Best regards,
  Alex Levine

On 2/4/15 4:32 PM, Rushi Agrawal wrote:
Thanks Alex for your detailed inspection of my work. Comments inline..

On 3 February 2015 at 21:32, Alexandre Levine < <>> wrote:

    I'm writing this in regard to several reviews concering tagging
    functionality for EC2 API in nova.
    The list of the reviews concerned is here:,n,z

    I don't think it's a good idea to merge these reviews. The
    analysis is below:

    *Tagging in AWS*

    Main goal for the tagging functionality in AWS is to be able to
    efficiently distinguish various resources based on user-defined

    "Tags enable you to categorize your AWS resources in different
    ways, for example, by purpose, owner, or environment.
    You can search and filter the resources based on the tags you add."

    (quoted from here:

    It means that one of the two main use-cases is to be able to use
    Tags as filter when you describe something. Another one is to be
    able to get information about particular tag with all of the
    resources tagged by it.
    Also there is a constraint:

    "You can tag public or shared resources, but the tags you assign
    are available only to your AWS account and not to the other
    accounts sharing the resource."

    The important part here is "shared resources" which are visible to
    different users but tags are not shared - each user sees his own.
    **Existing implementation in nova

    *Existing implementation of tags in nova's EC2 API covers only
    instances. But it does so in both areas:
    1. Tags management (create, delete, describe,...)
    2. Instances filtering (describe_instances with filtering by tags).
    The implementation is based on storing tags in each instance's
    metadata. And nova DB sqlalchemy level uses "tag:" in queries to
    allow instances describing with tag filters.

    I see the following design flaws in existing implementation:

    1. It uses instance's own metadata for storing information about
    assigned tags.
    - it doesn't scale when you want to start using tags for other
    resources. Following this design decision you'll have to store
    tags in other resources metadata, which mean different services
    APIs and other databases. So performance for searching for tags or
    tagged resources in main use cases should suffer. You'll have to
    search through several remote APIs, querying different metadatas
    to collect all info and then to compile the result.
    - instances are not shared resources, but images are. It means
    that, when developed, metadata for images will have to store
    different tags for different users somehow.

    2. EC2-specific code ("tag:" searching in novaDB sqlalchemy)
    leaked into lower layers of nova.
    - layering is violated. There should be no EC2-specifics below EC2
    API library in nova, ideally.

All of the Nova-EC2 mapping happens in Nova's DB currently. See InstanceIdMapping model in nova/db/sqlalchemy/ EC2 API which resides in Nova will keep using the Nova database as long as it is functional.

    - each other service will have to implement the same solution in
    its own DB level to support tagging for EC2 API.

    *Proposed review changes**
    The review in question introduces tagging for volumes and
    snapshots. It follows design decisions of existing instance
    tagging implementation, but realizes only one of the two use
    cases. It provides "create", "delete", "describe" for tags. But it
    doesn't provide describe_volumes or describe_snapshots for filtering.

I honestly forgot about those two methods. I can implement them.

    It suffers from the design flaws I listed above. It has to query
    remote API (cinder) for metadata. It didn't implement filtering by
    "tag:" in cinder DB level so we don't see implementation of
    describe_volumes with tags filtering.

Cinder do support filtering based on tags, and I marked the work as TODO in . This was not the reason why I didn't implement describe_volumes and describe_snapshots. Those two methods just missed my attention :)

Nova's EC2 API's tag filtering is also done in-memory presently if I'm correct, as Nova's API doesn't support filtering only on the basis of tag names or tag values alone..

    *Current stackforge/ec2-api tagging implementation**

    *In comparison, the implementation of tagging in
    stackforge/ec2-api, stores all of the tags and their links to
    resources and users in a separate place. So we can efficiently
    list tags and its resources or filter by tags during describing of
    some of the resources. Also user-specific tagging is supported.


    *Keeping in mind all of the above, and seeing your discussion
    about deprecation of EC2 API in nova, I don't feel it's a good
    time to add such a half-baked code with some potential problems
    into nova.*
    *I think it's better to concentrate on cleaning up, fixing,
    reviving and making bullet-proof whatever functionality is
    currently present in nova for EC2 and used by clients.

EC2 API already shares database with Nova's, so the tight coupling between EC2 API and Nova's database is not going to go away till the time EC2 API server/controller is present in Nova. Nova instance metadata is being used as EC2 instance tags, and what the above-referenced spec is doing is is very similar: Cinder volume metadata is being used as EC2 volume tags, and similarly for volume snapshots. I don't see a difference between instances and volumes and volume snapshots in the sense that they all are non-share-able (yet).

I completely understand that these patches look like feature additions. I started working on them first in January 2014 ( ), and at that time it was just a sincere effort to improve EC2 API using the first possible way I could find out. Since we have not deprecated the in-Nova EC2 support yet, and we are yet to reach a concrete plan to move forward, I am tempted to ask for allowing this patch to be considered for review..

I am fine if people think these patches shouldn't be allowed to go in. I can only wish that the patches got more attention when it was possible to get them merged :)

Rushi Agrawal

    Best regards,
      Alex Levine

    OpenStack Development Mailing List (not for usage questions)

OpenStack Development Mailing List (not for usage questions)

OpenStack Development Mailing List (not for usage questions)

Reply via email to