+1 to this approach, but I suspect there's a lot more to discuss/plan.
_*Leaking OSGi*_
I agree with Geoff's worry of us "leaking OSGi". However, we should just
"leak" the semantics of bundle dependencies and versioning, which need
to be explained to the user. I think we can avoid exposing the OSGi
syntax anywhere (except for power-users who want it).
_*`brooklyn.libraries` versus `package-imports analogy*_
OSGi recommends the use of package-imports, rather than explicitly
depending on specific bundles.
The analogy for us would be to depend on other catalog item
name:versionRange (e.g. in a `brooklyn.imports` section, or whatever).
We'd then figure out which bundle to get those from using the same rules
as OSGi package-imports.
Is it worth exploring further that option, rather than going down the
`brooklyn.libraries` route?
_*References Between Bundles*_
I don't think long-term the `brooklyn.libraries` should have URLs of
other bundles. It could give the name:versionRange of other bundles
(which assumes they have somehow already been added to Brooklyn).
(Reasoning is that other smart folk don't embed the URLs, presumably
because they want people to chose which repo things will come from -
e.g. enterprise private repo, versus maven central, etc).
_*Deploying Multiple Bundles*_
I lean towards something analogous to the kar file, but maybe lighter
weight rather than introducing anything like "features".
Really we just need a list of bundle URLs. If someone then deletes one
of those bundles, which breaks the other bundles deployed along with it,
then tough!
_*Blueprint Validation*_
Longer term, it would be nice if we further improved our blueprint
validation to ensure all references in a catalog item being added are
actually resolvable.
Adding to Brooklyn the ability to add a list-of-bundles will no doubt
make that harder!
_*Incremental Steps?*_
This might well prove to be a big change under-the-covers. Things that
scare me include:
* do we need to re-implement for catalog-items the same
dependency/version rules as OSGi uses for class-loading? (+)
* how much will backwards compatibility with historic persisted state
make things hard for us.
Are there sensible ways to scope this and de-risk it? Any incremental
milestones, which give us incremental value or de-risk it?
(+) We could potentially hack something where we synthesise a
package/resource per catalog item in our auto-generated bundles, and
then ask OSGi where it would load that from! (But that feels like a bad
idea, because power-users might add their own bundles directly to karaf.)
_*Business Case (aka value to end-users)*_
It would be good to be clearer on the user-value we deliver from this
(but please shout if others think that is already clear enough).
We're making "upgrades" easier - focusing specifically on deploying a
second version of a catalog item, which can then be used (ignoring
upgrade of existing app instances for now).
We're solving something like the DLL-hell problem (currently, if you
don't declare explicit versions within your blueprint, then your
blueprint might break when someone else adds a newer version of your
dependent blueprint for their own usage). Our current workaround for DLL
hell is impractical: hardcode explicit version numbers of your
dependencies in the `type: name:version` (no support for version ranges
etc, and it would be burried deep within the yaml files).
We're making it easier to deploy different versions of a "group of
catalog items" (e.g. a big app that is split into several catalog
items). Currently (if you didn't hardcode versions in every `type: ...`
reference), deploying a 2.0.0 version would likely break a subsequent
deployment of the 1.0.0 version, because it would default to the latest
for each catalog reference. That is, you'd only get the 1.0.0 version of
the main item, and then the 2.0.0 version of each item it referenced.
We'd move away from `brooklyn.libraries` hard-coding URLs to other
bundles, which most likely goes against the security policies of many
enterprises!
Is that a fair summary? Anything to add/change?
Aled
On 21/06/2017 09:41, Geoff Macartney wrote:
hi Alex,
that sounds fair enough. We'll want to add good documentation around what
it means if *everything* is now a bundle (and how we wrap plain catalogs
into bundles on users' behalf). We can use the text of your email as a
starting point for the docs; it would be good to explain some of the
_motivation_ around the choice to make everything a bundle, rather than
just say what we do.
I do like the idea of version ranges in brooklyn.libraries, but agree it's
best to proceed in small steps so that would be a later development.
Some more questions on implementation details
- is it part of this plan to move bundle storage into the persistence
location so that HA Brooklyn nodes can find all bundles?
- if a catalog.bom (symbolic name xyz) includes brooklyn.libraries and
specifies a remote URL (which gets downloaded), what happens if we delete
the bundle xyz - do the downloaded bundles get "cascade deleted"?
Geoff
On Tue, 20 Jun 2017 at 17:22 Alex Heneveld <[email protected]>
wrote:
Hi Geoff
Thanks. Replies to specific points:
> whether we are "leaking" an implementation detail (the fact that we
use an OSGI runtime)
My motivation was to try to minimize this leak. The fact that we use
OSGi internally for
dependency isolation forces us to use OSGi versions with bundles. But
most people
most places use the "-qualifier" suffix not ".qualifier", so to
accommodate that the
proposal suggests a recommended syntax in line with what most people do
("-qualifer")
which we can internally safely convert to an OSGi version when needed.
There is no need for authors ever to see OSGi and the docs PR doesn't
mention OSGi
until an "advanced" section.
> proposal would rely on Brooklyn moving exclusively to Karaf
No. The bundle version format is already a problem with the use of
embedded Felix
(it solves the problem noted in #672 about needing to use ".qualifer" in
your catalog.bom)
> libraries / required-bundle / version ranges
Where users want to indicate dependent bundles they currently have the
choice of
"brooklyn.libraries" (non-osgi) or "require-bundle" (osgi). Both work.
We don't need to disable the
former. I'm in two minds about whether we'd want to, at some point, but
not right now of course.
An argument for disabling is that it sets up a single canonical way to
do things. Yes it
means people have to learn OSGi but this is a fairly advanced use case
and OSGi is the right tool for the job.
If we didn't, it almost certainly means adding version range support in
`brooklyn.libraries` which
is yet more duplication of what OSGi lets us do. However it would be
easier for non-OSGi users
(and a lot of OSGi users I bet!) if we did invest in a nice range syntax
in `brooklyn.libraries`.
Note `brooklyn.libraries` is the only place where you can specify a URL
for a dependent bundle;
the OSGi manifest doesn't let you do this (though if using Karaf you can
configure it to look in a set of repos,
which is possibly).
It boils down to how we expect people to manage the installation of
multiple catalog bundles
(URLs v configured repos v something else ... maybe KAR archives which
could easily be supported
but they're wasteful if the same dependency is included in many KARs).
Not sure where we'll want to go with this but it's not something we need
to solve for this proposal.
HTH
Best
Alex
On 20/06/2017 17:00, Geoff Macartney wrote:
hi Alex,
In general I think this is a good direction to go in, some questions and
thoughts below.
One concern is whether we are "leaking" an implementation detail (the
fact
that we
use an OSGI runtime) into the specification of what Brooklyn is. I think
it is good to
aim for the improved modularity of having a bundle of catalog definitions
and related
resources, all at a given version and added or removed as a unit. I'm
less
sure we
want to advertise that all Brooklyn definitions must be contained in OSGI
bundles,
even if we will wrap them for you. This seems to impose an additional
requirement
on blueprint authors (learn OSGI) that we don't really want to impose?
As an aside, I'm assuming that your proposal would rely on Brooklyn
moving
exclusively to the Karaf launcher, is that the case?
Specific notes inline below.
Geoff
On Mon, 19 Jun 2017 at 17:39 Alex Heneveld <
[email protected]>
wrote:
Hi All-
With bundles now being neatly supported for add and remove, I think it
makes sense to use make the "bundle" a first-class concept in managing a
brooklyn catalog (type registry).
The basic idea is that users will work with ZIPs (or directories if
using the br CLI), where a "catalog.bom" is in the root; this is a
bundle, and operations like add/remove/update apply to the bundle.
Increasingly people are doing this for additions, so it seems pretty
obvious that delete should follow suit, and something like a "/bundles"
REST endpoint is appropriate.
I do like this; as part of this we should re-instate the automatic
scanning
of bundles for catalog.bom (feature FEATURE_LOAD_BUNDLE_CATALOG_BOM).
Now that catalog items can tell you their containing bundle
(getContainingBundle())
it should be possible to make this work with rebind.
You can still upload just a catalog.bom file; in this proposal it will
be wrapped in a bundle (internally, hidden from user). A bundle name
and version will have to be inferred/guessed for backwards
compatibility, and probably we should deprecate supplying a BOM which
doesn't define a bundle name and version.
+1
If a ZIP/folder is supplied, it doesn't have to contain any OSGi bundle
definition; that call all be inferred. If a MANIFEST.MF is supplied,
any bundle name/version given there must of course match that in the
catalog.bom file.
If dependent bundles are declared, they are obeyed at
runtime and can be used to find things (pretty much identically to how
we treat "libraries" in the catalog.bom.)
Not so sure about that last paragraph - again this is related to leaking
the use
of OSGI. This would allow someone to write a catalog.bom that would ONLY
work when supplied along with a related manifest (to specify certain
dependencies
and so set up the search paths). Do we want to allow that, or should all
definitions
required for entities in a catalog be supplied in the catalog.bom (or its
included
catalog definitions).
* Search Paths - Bundles fix an issue where types have implicit version
dependencies. Consider an "acme-cluster" of "acme-node" entities,
defined in a bundle. If I release a 1.0 version of this, and then a 2.0
version, I'd probably expect the 1.0 acme-cluster to load 1.0 acme-node
entities, and same for 2.0. But currently any reference to "acme-node"
takes the latest version, even if there is a different version in the
same bundle. Shifting to bundles, I think we should change this, so
that versions of items in the same bundle (or on the OSGi or library
search path for that bundle) are preferred over more recent versions in
the repository.
+1 this sounds like a better match to what blueprint authors intend
(And since we will follow OSGi loading, we can also, if
we need to, use OSGi config to say that "acme" bundle requires a given
version _range_ for some other bundle, to have more control over
dependency versions when required.)
again, how "use OSGI config" - there may be value in specifying ranges
but
we wouldn't want users to have to use OSGI declarations to achieve it.
For most users and blueprints this will cause no changes. However there
are some changes to note:
* If people have been sloppy with versions or metadata, in most cases
they will get a warning
* In some cases if uploading a ZIP, the catalog.bom may need to be
edited (the upload will give a clear error, and it should not expect
rebind)
* When a type is referenced with no version, the loading semantics may
change if that type is in the same bundle (this is probably what the
user expects)
Does anyone oppose any of the above, or have further questions?
Best
Alex
[1] https://github.com/apache/brooklyn-server/pull/672
