Prompted by David Lehman's patch set, I wanted to discuss how we're building what we're installing in Fedora and RHEL, and how we might need to represent that both in the media and in the UI.
What we've got now exposed at the lowest level (yum/kickstart) and supported as first-level objects throughout the system: 1. packages 2. groups (of packages) 3. repos (which contain group lists, and packages)
What we additionally expose only in anaconda, and lose post-install: 4. tasks in install classes (lists of groups)
What we expose in Fedora: 5. spins (live ISO via kickstart file) 6. 'the install DVD' (a package set via a kickstart file)
What we expose in RHEL: 7. variants (essentially, a spin that's a package set instead of a live ISO) 8. add-ons (repos, with groups, that apply to specific variants)
All of these use the comps file, which is a list of groups.
Now, as we rework anaconda's UI (and backend, per dlehman's patches), we need to make sure not to break everything horribly.
In Fedora, spins and the install DVD are driven off of kickstart files. This was done because it's the only thing pungi and livecd-creator supports, and because each spin/install DVD is a distinct object that can have its own kickstart. Anaconda doesn't need to worry about separately consuming them, as there's only one repo to ever pull from, and the spins are live images.
In RHEL, variants and add-ons are driven off of a different set of files, for assorted historical reasons. Kickstart files don't directly work here because they don't allow you to define intermediate objects (add-ons) that are presented in the installer. Anaconda processes variants and add-ons via code that reads .treeinfo.
PROPOSAL GIVEN THE NEW UI
For reference, the proposed new UI is at: http://linuxgrrl.com/fedora-ux/Projects/Anaconda/Live%20Prototypes/07-softwa...
As you'll notice, it contains something that looks like a spin or variant on the left, and things that could be considered 'add-ons' on the right. These apply to the choices on the left.
This would seem to fit nicely into the variants + add-ons model that RHEL uses. However, there are a couple of problems:
- The code that parses the variants/add-on information and creates the appropriate .treeinfo bits does not exist in Fedora
This is obviously a simple matter of not-so-simple code.
- It requires creating a separate set of metadata (the variants file) that does not currently exist.
This is relatively trivial.
- It can cause skew between the live and not-live spins if this variants file falls out of sync with the kickstarts in spin-kickstarts.
This is ugly - how can we get these to use a single set of data? Unfortunately, kickstart's %packages model doesn't give us an easy way to define a variant and its add-ons in a kickstart file. Ideas?
Bill
On Fri, 2012-02-24 at 16:55 -0500, Bill Nottingham wrote:
Prompted by David Lehman's patch set, I wanted to discuss how we're building what we're installing in Fedora and RHEL, and how we might need to represent that both in the media and in the UI.
What we've got now exposed at the lowest level (yum/kickstart) and supported as first-level objects throughout the system:
- packages
- groups (of packages)
- repos (which contain group lists, and packages)
What we additionally expose only in anaconda, and lose post-install: 4. tasks in install classes (lists of groups)
This should go away, whatever else we do.
What we expose in Fedora: 5. spins (live ISO via kickstart file) 6. 'the install DVD' (a package set via a kickstart file)
What we expose in RHEL: 7. variants (essentially, a spin that's a package set instead of a live ISO) 8. add-ons (repos, with groups, that apply to specific variants)
All of these use the comps file, which is a list of groups.
Now, as we rework anaconda's UI (and backend, per dlehman's patches), we need to make sure not to break everything horribly.
In Fedora, spins and the install DVD are driven off of kickstart files. This was done because it's the only thing pungi and livecd-creator supports, and because each spin/install DVD is a distinct object that can have its own kickstart. Anaconda doesn't need to worry about separately consuming them, as there's only one repo to ever pull from, and the spins are live images.
In RHEL, variants and add-ons are driven off of a different set of files, for assorted historical reasons. Kickstart files don't directly work here because they don't allow you to define intermediate objects (add-ons) that are presented in the installer. Anaconda processes variants and add-ons via code that reads .treeinfo.
PROPOSAL GIVEN THE NEW UI
For reference, the proposed new UI is at: http://linuxgrrl.com/fedora-ux/Projects/Anaconda/Live%20Prototypes/07-softwa...
As you'll notice, it contains something that looks like a spin or variant on the left, and things that could be considered 'add-ons' on the right. These apply to the choices on the left.
This would seem to fit nicely into the variants + add-ons model that RHEL uses. However, there are a couple of problems:
- The code that parses the variants/add-on information and creates the appropriate .treeinfo bits does not exist in Fedora
This is obviously a simple matter of not-so-simple code.
- It requires creating a separate set of metadata (the variants file) that does not currently exist.
This is relatively trivial.
- It can cause skew between the live and not-live spins if this variants file falls out of sync with the kickstarts in spin-kickstarts.
This is ugly - how can we get these to use a single set of data? Unfortunately, kickstart's %packages model doesn't give us an easy way to define a variant and its add-ons in a kickstart file. Ideas?
In theory, what we need is two new meta-group types: variant (aka spin) and addon. Each is a set of groups. You can only have one variant selected or installed at any given time. Each variant defines a set of addons, of which any number may be selected/installed. Regular groups and packages are only exposed via kickstart, if at all.
Once we have the groups set up and accompanying kickstart support, the spins should end up being exactly the same package set as you'd get by choosing the same variant+addons in the anaconda software selection ui. The addition of meta-packages and/or kickstart script snippets (perhaps per variant?) could bridge the gap so that the spins end up identical to the corresponding anaconda variant/addon selections.
How does this fit with your proposal and your concerns?
Dave
David Lehman (dlehman@redhat.com) said:
In theory, what we need is two new meta-group types: variant (aka spin) and addon. Each is a set of groups. You can only have one variant selected or installed at any given time. Each variant defines a set of addons, of which any number may be selected/installed. Regular groups and packages are only exposed via kickstart, if at all.
Once we have the groups set up and accompanying kickstart support, the spins should end up being exactly the same package set as you'd get by choosing the same variant+addons in the anaconda software selection ui. The addition of meta-packages and/or kickstart script snippets (perhaps per variant?) could bridge the gap so that the spins end up identical to the corresponding anaconda variant/addon selections.
How does this fit with your proposal and your concerns?
I can see this working. Initial concerns:
- Would we need to drive the variants and add-ons into first class objects in yum as well (such that, if intstalled, they're marked on the system in some way (yumdb, etc.), and that they can be installed post-install, etc?) The simplest solution is add these as additional sections in the comps file, and parse them with yum's comps module - the question is whether this info is used in yum itself outside of that, or if it's left for higher-level tools.
- In RHEL, due to fun with entitlements & certs, add-ons are segmented as separate repos. Your proposal doesn't obviously prevent this, but I think we need to explicitly keep that working.
Bill