[Pulp-dev] the "relative path" problem

Wed Oct 21 18:48:16 UTC 2020

TL;DR: An attempt to propose the least invasive option to solve the case
when remote repository metadata needs to be mirrored. Please provide
feedback if you are interested in the outcome.

There have been multiple attempts and discussions to solve the
relative_path problem in a general way which covers multiple use cases.
They all look very invasive and only possible to be done in Pulp 4+ due to
the amount and significance of changes that needs to be made, to the data
models and/or to the API.

The following proposal solves only this use case: As a user, I can mirror
remote repository metadata as is.
An additional goal is to avoid backward incompatible changes and ideally
leave a way for further improvement to solve the problem in a more general
way.
(The following proposal does NOT solve a use case: As a user, I can have
the same content under different relative paths in any repository.)

Proposal:
- Have a way to distinguish between repositories with managed content and
with the exact mirror (e.g. create a repository with exact_mirror=True or a
new dedicated repository type)
 - For such repos, create a publication at sync time (includes published
artifacts and metadata).
 - For such repos, publish is no-op and always returns the existing
publication for the requested repo version.
 - For such repos, no modifications are allowed except the sync in mirror
mode. (At least for RPM plugin, I believe we can't allow discrepancies
between metadata and content in a repo, especially if some content is
removed.)

Pros:
 - non-invasive, only additive model changes
 - can be implemented in a plugin which needs it or it can be moved to the
pulpcore if it allows plugin input at certain points.
 - leaves a way for further improvement to handle a more general case, see
the full proposal here
https://hackmd.io/02KBjCD3Q0WP7p4ALwzhJw#relative_path-in-PublishedArtifact-only

Cons:
 - doesn't solve the problem of various relative paths for the same content
in general way
 - a separate code path (at times) to handle this type of repositories.

For reference:
 - hackmd doc with all the considered proposals and the summary of the
potentially valid ones https://hackmd.io/02KBjCD3Q0WP7p4ALwzhJw
 - pulpcon video with the discussion of the proposals
https://www.youtube.com/watch?v=7IzxAQYr5-I

Thanks,
Tanya

On Thu, May 7, 2020 at 2:07 PM Brian Bouterse <bmbouter at redhat.com> wrote:

> I agree with that problem statement. pulp_file may want to have the same
> Content at two different paths in different RepositoryVersions (or even the
> same RepositoryVersion). Without this capability a user could never "move"
> where content lives in a RepositoryVersion if its already been placed in
> any other RepositoryVersion.
>
> Additionally pulp_maven may need to sync two repositories in the wild that
> already contain the same content in two locations. I offer this as example
> not to pile-on, but because it's a multi-content artifact which I believe
> we will need to consider also as we work towards a solution.
>
> I've been spending time on developing a solution, but it needs more work
> so it's not ready yet. Also other katello and galaxy_ng work continues to
> pre-empt this, so it could take a while.
>
> On Thu, May 7, 2020 at 3:39 AM Matthias Dellweg <mdellweg at redhat.com>
> wrote:
>
>> > Users need to be able to store the same content unit at different
>> relative paths in different repository versions. This problem is not unique
>> to the RPM plugin. Do we agree about that?
>> Yes, we agree. In pulp_deb relative_path is part of the contents
>> natural_key to circumvent this problem. So this creates two content units
>> that only differ in relativ_path. At least they share the artifact.
>>
>> On Thu, May 7, 2020 at 2:06 AM Dennis Kliban <dkliban at redhat.com> wrote:
>>
>>> I'd like to provide a little bit more context for my previous email by
>>> going back to the original problem statement:
>>>
>>> On Wed, Apr 1, 2020 at 9:23 AM Daniel Alley <dalley at redhat.com> wrote:
>>>
>>>> Problem:
>>>>
>>>> Currently, a relative_path is tied to content in Pulp. This means that
>>>> if a content unit exists in two places within a repository or across
>>>> repositories, it has to be stored as two separate content units. This
>>>> creates redundant data and potential confusion for users.
>>>>
>>>> As a specific example, we need to support mirroring content in pulp_rpm
>>>> <https://pulp.plan.io/issues/6353>. Currently, for each location at
>>>> which a single package is stored, we’ll need to create a content unit. We
>>>> could end up with several records representing a single package. Users may
>>>> be confused about why they see multiple records for a package and they may
>>>> have trouble for example deciding which content unit to copy.
>>>>
>>> Users need to be able to store the same content unit at different
>>> relative paths in different repository versions. This problem is not unique
>>> to the RPM plugin. Do we agree about that?
>>>
>>> I've been working on a potential solution that solves this problem in a
>>> document[0]. It is a complicated change and the document does not fully
>>> capture the plan yet. Feedback and help on the design is welcome.
>>>
>>> [0] https://hackmd.io/02KBjCD3Q0WP7p4ALwzhJw?edit
>>>
>>>
>>> On Mon, May 4, 2020 at 4:11 PM Dennis Kliban <dkliban at redhat.com> wrote:
>>>
>>>> I've reached two conclusions while trying to formulate a solution:
>>>>
>>>> This problem needs to be solved at the repository version level.
>>>> Repository membership needs to be tracked at the artifact level, and not
>>>> content level as it is now.
>>>>
>>>> On Thu, Apr 30, 2020 at 1:11 PM Daniel Alley <dalley at redhat.com> wrote:
>>>>
>>>>> Cool, so the only difference is whether to try to store the
>>>>> relationship in the DB, or leverage the fact that we already have the
>>>>> metadata and just re-parse it.
>>>>>
>>>>> I know the latter approach has yet to be written up, but my concern
>>>>> there is that adding another layer of indirection between "repository
>>>>> version" and "content" is going to have an adverse impact on performance,
>>>>> since it is already the most complex and demanding query we issue to the DB
>>>>> and one of the most common and important.
>>>>>
>>>>> On Thu, Apr 30, 2020 at 12:50 PM David Davis <daviddavis at redhat.com>
>>>>> wrote:
>>>>>
>>>>>> Yes but I was imagining the mapping would be stored not as Content
>>>>>> but as a separate object. So we wouldn't use filename for the mapping
>>>>>> (rather we'd use ContentArtifact pk) and  we wouldn't need to change
>>>>>> ContentArtifact's relative_path at all. That said, I think your solution
>>>>>> captures the idea though and is better in some ways.
>>>>>>
>>>>>> Changing the RepositoryContent model to point to ContentArtifacts and
>>>>>> store relative_paths is probably the best and most correct solution in
>>>>>> theory. However, it's going to be painful to implement for both core and
>>>>>> plugins.
>>>>>>
>>>>>> David
>>>>>>
>>>>>>
>>>>>> On Thu, Apr 30, 2020 at 12:33 PM Daniel Alley <dalley at redhat.com>
>>>>>> wrote:
>>>>>>
>>>>>>> @David Davis <daviddavis at redhat.com>  so this proposal would go
>>>>>>> something like this, correct?:
>>>>>>>
>>>>>>> * For the signed metadata / exact mirror use-case we need to store
>>>>>>> the repository metadata itself as a content unit inside the
>>>>>>> RepositoryVersion anyway (because the hash must be equal)
>>>>>>> * Because we have this metadata lying around, we can reference it at
>>>>>>> publish time to discover the appropriate PublishedArtifact.relative_path
>>>>>>>    * Create a map of "filename" -> "location_href" and look up the
>>>>>>> filename of each RPM package to find the appropriate path
>>>>>>>    * This should be pretty fast for the RPM plugin since
>>>>>>> createrepo_c is doing all the hard work
>>>>>>> * Data migration to ensure ContentArtifact.relative_path is only
>>>>>>> storing the filename (and I would suggest we also change the name to
>>>>>>> "filename")
>>>>>>> * If metadata isn't present in the RepositoryVersion, then just
>>>>>>> tweak the PublishedArtifact.relative_path so that it uses whichever our
>>>>>>> default repo layout is
>>>>>>>
>>>>>>> On Tue, Apr 28, 2020 at 11:41 AM David Davis <daviddavis at redhat.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Yes, that's correct. During our meeting we discussed two options:
>>>>>>>> the first was to extend RepositoryContent to store relative path per
>>>>>>>> ContentArtifact as storing a relative_path per Content won't work for
>>>>>>>> multi-Artifact Content units.
>>>>>>>>
>>>>>>>> An alternative that I pitched was to have plugins (or maybe even
>>>>>>>> core someday) store this information outside RepositoryContent and then use
>>>>>>>> this information during publishing to set relative_path on
>>>>>>>> PublishedArtifacts. We'd have to modify the content app if we wanted to
>>>>>>>> support pass through publications but I think asking plugins to use
>>>>>>>> published artifacts in this case is warranted. That said, I don't think
>>>>>>>> anyone else was keen on this idea though.
>>>>>>>>
>>>>>>>> David
>>>>>>>>
>>>>>>>>
>>>>>>>> On Tue, Apr 28, 2020 at 10:30 AM Matthias Dellweg <
>>>>>>>> mdellweg at redhat.com> wrote:
>>>>>>>>
>>>>>>>>> That is only used for passthrough publication afaik. If you
>>>>>>>>> publish each content unit "by hand", you create a new relative path for
>>>>>>>>> each published artifact. That is, why it can be empty and still the content
>>>>>>>>> can be published.
>>>>>>>>>
>>>>>>>>> On Tue, Apr 28, 2020 at 4:09 PM Daniel Alley <dalley at redhat.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> We realized in our discussion that the original proposal
>>>>>>>>>> described in my email will not work, because "relative_path" ultimately
>>>>>>>>>> describes the path of the published *artifacts* (not content),
>>>>>>>>>> and for content types with multiple artifacts, storing this information in
>>>>>>>>>> a field on RepositoryContent would not be possible.
>>>>>>>>>>
>>>>>>>>>> On Mon, Apr 27, 2020 at 6:08 PM Daniel Alley <dalley at redhat.com>
>>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>>> There is a video call scheduled to discuss this issue tomorrow
>>>>>>>>>>> (Tuesday April 28th) at 13:30 UTC (please convert to your local time).
>>>>>>>>>>> https://meet.google.com/scy-csbx-qiu
>>>>>>>>>>>
>>>>>>>>>>> On Sat, Apr 25, 2020 at 7:02 AM David Davis <
>>>>>>>>>>> daviddavis at redhat.com> wrote:
>>>>>>>>>>>
>>>>>>>>>>>> I had a chance to think about this some more yesterday and
>>>>>>>>>>>> wanted to email out my thoughts. I also think that this change sounds scary
>>>>>>>>>>>> and will have a big impact on plugin writers so I thought of a couple
>>>>>>>>>>>> alternatives:
>>>>>>>>>>>>
>>>>>>>>>>>> First, we could add a relative_path field to RepositoryContent
>>>>>>>>>>>> instead of moving it there. This would be an optional field. It would be up
>>>>>>>>>>>> to plugins to manage this field and they would still need to populate the
>>>>>>>>>>>> relative_path field on ContentArtifact. But plugins could use this optional
>>>>>>>>>>>> field to store relative paths per repository and then use this field when
>>>>>>>>>>>> generating publications.
>>>>>>>>>>>>
>>>>>>>>>>>> The second alternative is one that is already laid out in the
>>>>>>>>>>>> original email but to call it out again: it would be to not solve this in
>>>>>>>>>>>> pulpcore. RPM would create its own object that would map content in a
>>>>>>>>>>>> repository to relative_paths.
>>>>>>>>>>>>
>>>>>>>>>>>> David
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> On Tue, Apr 21, 2020 at 9:22 AM Quirin Pamp <pamp at atix.de>
>>>>>>>>>>>> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> I am not currently very well versed in the classes involved,
>>>>>>>>>>>>> but moving relative_path around sounds slightly scary with the potential to
>>>>>>>>>>>>> break things.
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> As such, I would be interested to be kept in the loop as this
>>>>>>>>>>>>> moves forward. (Mailing list once there is some movement is entirely
>>>>>>>>>>>>> sufficient 😉)
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>>
>>>>>>>>>>>>> Quirin Pamp
>>>>>>>>>>>>> ------------------------------
>>>>>>>>>>>>> *From:* pulp-dev-bounces at redhat.com <
>>>>>>>>>>>>> pulp-dev-bounces at redhat.com> on behalf of Ina Panova <
>>>>>>>>>>>>> ipanova at redhat.com>
>>>>>>>>>>>>> *Sent:* 21 April 2020 14:07:13
>>>>>>>>>>>>> *To:* Daniel Alley <dalley at redhat.com>
>>>>>>>>>>>>> *Cc:* Pulp-dev <pulp-dev at redhat.com>
>>>>>>>>>>>>> *Subject:* Re: [Pulp-dev] the "relative path" problem
>>>>>>>>>>>>>
>>>>>>>>>>>>> Daniel,
>>>>>>>>>>>>>
>>>>>>>>>>>>> how about setting up a meeting and brainstorm the
>>>>>>>>>>>>> alternatives, pros/cons there?
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> --------
>>>>>>>>>>>>> Regards,
>>>>>>>>>>>>>
>>>>>>>>>>>>> Ina Panova
>>>>>>>>>>>>> Senior Software Engineer| Pulp| Red Hat Inc.
>>>>>>>>>>>>>
>>>>>>>>>>>>> "Do not go where the path may lead,
>>>>>>>>>>>>>  go instead where there is no path and leave a trail."
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> On Fri, Apr 17, 2020 at 5:57 PM Daniel Alley <
>>>>>>>>>>>>> dalley at redhat.com> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>> Bump, this item needs to move forwards soon.  Does anyone have
>>>>>>>>>>>>> any thoughts?
>>>>>>>>>>>>>
>>>>>>>>>>>>> On Wed, Apr 1, 2020 at 9:40 AM Pavel Picka <ppicka at redhat.com>
>>>>>>>>>>>>> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>> I'd like to add one more question to this topic. Do you think
>>>>>>>>>>>>> it is a blocker for PRs [0] & [1] as by testing [2] this features I haven't
>>>>>>>>>>>>> run into real world example where two really same name packages appears.
>>>>>>>>>>>>> I think this is a 'must have' feature but until we
>>>>>>>>>>>>> solve/decide it we can have two features working may with warning in docs
>>>>>>>>>>>>> for users that can happen in some 'special' repositories.
>>>>>>>>>>>>>
>>>>>>>>>>>>> To follow topic directly I like proposed move to
>>>>>>>>>>>>> 'RepositoryContent' and add it to its uniqueness constraint (if I
>>>>>>>>>>>>> understand well).
>>>>>>>>>>>>>
>>>>>>>>>>>>> [0] https://github.com/pulp/pulp_rpm/pull/1657
>>>>>>>>>>>>> [1] https://github.com/pulp/pulp_rpm/pull/1642
>>>>>>>>>>>>> [2] tested with centos 7, 8, opensuse and SLE repositories
>>>>>>>>>>>>>
>>>>>>>>>>>>> On Wed, Apr 1, 2020 at 3:22 PM Daniel Alley <dalley at redhat.com>
>>>>>>>>>>>>> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>> We'd like to start a discussion on the "relative path problem"
>>>>>>>>>>>>> identified recently.
>>>>>>>>>>>>> Problem:
>>>>>>>>>>>>>
>>>>>>>>>>>>> Currently, a relative_path is tied to content in Pulp. This
>>>>>>>>>>>>> means that if a content unit exists in two places within a repository or
>>>>>>>>>>>>> across repositories, it has to be stored as two separate content units.
>>>>>>>>>>>>> This creates redundant data and potential confusion for users.
>>>>>>>>>>>>>
>>>>>>>>>>>>> As a specific example, we need to support mirroring content
>>>>>>>>>>>>> in pulp_rpm <https://pulp.plan.io/issues/6353>. Currently,
>>>>>>>>>>>>> for each location at which a single package is stored, we’ll need to create
>>>>>>>>>>>>> a content unit. We could end up with several records representing a single
>>>>>>>>>>>>> package. Users may be confused about why they see multiple records for a
>>>>>>>>>>>>> package and they may have trouble for example deciding which content unit
>>>>>>>>>>>>> to copy.
>>>>>>>>>>>>> Proposed Solution:
>>>>>>>>>>>>>
>>>>>>>>>>>>> Move “relative_path” from its current location on
>>>>>>>>>>>>> ContentArtifact, to RepositoryContent. This will require a sizable data
>>>>>>>>>>>>> migration. It is possibly the case that in rare cases, repository versions
>>>>>>>>>>>>> may change slightly due to deduplication.
>>>>>>>>>>>>>
>>>>>>>>>>>>> A repository-version-wide uniqueness constraint will be
>>>>>>>>>>>>> present on “relative_path”, independently of any other repository uniquness
>>>>>>>>>>>>> constraints (repo_key_fields) defined by the plugin writer.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Modify the Stages API so that the relative_path can be
>>>>>>>>>>>>> processed in the correct location – instead of “DeclarativeArtifact” it
>>>>>>>>>>>>> will likely need to go on “DeclarativeContent”
>>>>>>>>>>>>>
>>>>>>>>>>>>> Remove “location_href” from the RPM Package content model – it
>>>>>>>>>>>>> was never a true part of the RPM (file) metadata, it is derived from the
>>>>>>>>>>>>> repository metadata. So storing it as a part of the Content unit doesn’t
>>>>>>>>>>>>> entirely make sense.
>>>>>>>>>>>>> Alternatives
>>>>>>>>>>>>>
>>>>>>>>>>>>> In most cases, a content unit will have a single relative path
>>>>>>>>>>>>> for a content unit. Creating a general solution to solve a one-off problem
>>>>>>>>>>>>> is usually not a good idea. As an alternative, we could look at another
>>>>>>>>>>>>> solution for mirroring content. One example might be to create a new object
>>>>>>>>>>>>> (e.g. RpmRepoMirrorContentMapping) that maps content to specific paths
>>>>>>>>>>>>> within a repo or repo version.
>>>>>>>>>>>>> Questions
>>>>>>>>>>>>>
>>>>>>>>>>>>>    - How do we handle this in pulp_file? How are content
>>>>>>>>>>>>>    units identified in pulp_file without relative_path?
>>>>>>>>>>>>>       - Checksum?
>>>>>>>>>>>>>       - How was this problem handled in Pulp 2?
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> Please weigh in if you have any input on potential problems
>>>>>>>>>>>>> with the proposal, potential alternate solutions, or other insights or
>>>>>>>>>>>>> questions!
>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>> Pulp-dev mailing list
>>>>>>>>>>>>> Pulp-dev at redhat.com
>>>>>>>>>>>>> https://www.redhat.com/mailman/listinfo/pulp-dev
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> --
>>>>>>>>>>>>> Pavel Picka
>>>>>>>>>>>>> Red Hat
>>>>>>>>>>>>>
>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>> Pulp-dev mailing list
>>>>>>>>>>>>> Pulp-dev at redhat.com
>>>>>>>>>>>>> https://www.redhat.com/mailman/listinfo/pulp-dev
>>>>>>>>>>>>>
>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>> Pulp-dev mailing list
>>>>>>>>>>>>> Pulp-dev at redhat.com
>>>>>>>>>>>>> https://www.redhat.com/mailman/listinfo/pulp-dev
>>>>>>>>>>>>>
>>>>>>>>>>>> _______________________________________________
>>>>>>>>>> Pulp-dev mailing list
>>>>>>>>>> Pulp-dev at redhat.com
>>>>>>>>>> https://www.redhat.com/mailman/listinfo/pulp-dev
>>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>>> Pulp-dev mailing list
>>>>> Pulp-dev at redhat.com
>>>>> https://www.redhat.com/mailman/listinfo/pulp-dev
>>>>>
>>>> _______________________________________________
>>> Pulp-dev mailing list
>>> Pulp-dev at redhat.com
>>> https://www.redhat.com/mailman/listinfo/pulp-dev
>>>
>> _______________________________________________
>> Pulp-dev mailing list
>> Pulp-dev at redhat.com
>> https://www.redhat.com/mailman/listinfo/pulp-dev
>>
> _______________________________________________
> Pulp-dev mailing list
> Pulp-dev at redhat.com
> https://www.redhat.com/mailman/listinfo/pulp-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/pulp-dev/attachments/20201021/e2f5e4eb/attachment.htm>