Hi,
Adding my +1 for this proposal. It should make blueprint inheritance
much
more useful.
I think Aled is right that 'merge' is going to be the most common use
case,
and optimising for that seems like the right idea. Then, overwrite (i.e.
replacing the entire map) and deletion of individual keys can be
implemented using the more complex DSL syntax. Regarding the aside,
that's
an interesting point. There's a few ways of treating it that might be
possible to include as part of this functionality. For example,
providing a
base environment map, to which separate maps for
install/customize/launch
commands have their own maps that are then merged (or overwritten)
individually, would be useful, as well as just being able to supply
different map configuration. Some of this can be handled using YAML
templating functionality, defining the base map and including it in
various
places, but that seems a bit messy; intrinsic syntax that is used
explicitly for handling maps of data would be better.
Another thing to think about when dealing with environment maps would be
the ability to auto-populate it with a set of attributes (with some
mapping
from the dot-separated lowercase sensor name to an underscore-separated
uppercase name) and their current values. This is a common pattern in
many
blueprints, so a simple bit of DSL syntax would be handy.
The templateOptions handling does seem more complicated, but again,
optimising for the common case is the right thing to do, at least
initially. So, merging the leaf nodes of the tree of maps. This also
applies to location inheritance, as well, but I guess that will be
handled
automatically if the logic happens in all `Configurable` brooklyn
objects?
Actually, there will be lists as well as maps, and maps of maps (as
well as
lists of maps, maps of maps of maps, and maps of lists and so on...!)
structures in the entity config as well, so making merge work in the
general case of leaf nodes in an arbitrary tree structure is going to be
required anyway.
Basically, if we follow the principle of least surprise when
implementing,
it should also be reasonably intuitive.
Andrew.
On Wed, 25 May 2016 at 21:56 Aled Sage <[email protected]> wrote:
Alex, all,
I wondered about (1) as well. I concluded that we should optimise for
what I believe is the most common case for these config keys: being
able
to add additional values and override specific values. For example, if
someone defines a Tomcat entity with default environment variables,
then
I want to be able to add to those and override specific values easily.
The alternative approach you mentioned (below) is technically elegant,
but feels overly complicated YAML for the common case.
shell.env:
$brooklyn:merge:
- overridden_key: value
new_key: value
If folk agree that the common case to optimise for is the "merge" for
config keys like shell.env, then we could limit the less-obvious
approach for the overwrite use-case:
shell.env:
$brooklyn:overwrite:
- new_key1: value
new_key2: value
/(As an aside, at some point I'd like to completely revisit shell.env:
we should better support supplying different environment variables to
each of the install/launch/checkRunning scripts.)/
---
For (2), I lean towards treating templateOptions (within
provisioningProperties) as a special case. Really we need a major
overhaul of our JcloudsLocation code.
My long-term ideal would be that we don't need to put those key-values
within a templateOptions sub-map. That is really a consequence of our
implementation.
I think merging of templateOptions is the behaviour that a user
(unfamiliar with the underlying implementation) would expect. Until one
explains about how this maps to the jclouds TemplateOptions class,
there
is little logic for what is a top-level key-value and what goes inside
templateOptions. So I think a user would want them both merged.
Aled
On 25/05/2016 17:07, Alex Heneveld wrote:
two difficulties i see:
1) how do i clear inherited map values on a MERGEd config?
2) how do we specify that "templateOptions" in a
"provisioningProperties" is to be merged?
(followed by conclusion -- outlining an alternative but overall
unsure)
*1) how do i clear inherited map values on a MERGEd config?*
e.g. say we have
parent with
shell.env: { X: 1 }
and child wants to ensure shell.env has *nothing* for X. previously
child could just say
shell.env: {}
however with this proposal i think the child now requires:
shell.env: { X: null }
listing every key it inherits and hoping that the shell to-string
excludes nulls.
*2) how do we specify that "templateOptions" in a
"provisioningProperties" is to be merged?*
for this proposal we'll write code to understand ConfigInheritance at
entities and locations. i don't think we yet have discussed any way
to do this one level deeper. specifically if i've got
# parent
provisioningProperties:
templateOptions:
floatingIpPoolNames: pool
# child
type: parent
provisioningProperties:
templateOptions:
networks: xyz
we know from the definition of PROVISIONING_PROPERTIES on
SoftwareProcess that that map should be merged with its super. however
when merging the actual provisioning properties we have no way to
understand the semantics, do we? specifically there is no indication
that PROVISIONING_PROPERTIES is:
* a map of config
* usually containing config keys from JcloudsLocationConfig.
without that knowledge we can't do the "depth 2" merge illustrated in
the example, can we? in other words we'll lose "floatingIpPoolNames:
pool".
*conclusion*
these two issues aren't showstoppers but they are a little bit
smelly. apart from them the proposal is very good: it solves an
irritation around maps in a fairly simple elegant way and only
impacting opt-in config-keys in a cleanly defined way.
using $brooklyn:super() with a proposed $brooklyn:merge is an
alternative solution which lets us solve (2) and avoids both of these
issues:
shell.env:
$brooklyn:merge:
- $brooklyn:super()
- overridden_key: value
new_key: value
this could also work for lists. however it requires the user to
explicitly write this, it's uglier, and it might be hard to implement.
if we introduced a `$brooklyn:overwrite` we could combine aled's
proposal with dsl solutions to problems (1) and (2) described here.
but it makes behaviours more complicated.
in short not yet sure what is best...
--a
On 25/05/2016 07:36, Geoff Macartney wrote:
+1
This sounds like a good proposal. At the same time it’s fairly
complex, so I think an important part of the change for this would be
1. to test it comprehensively, so each of the scenarios below would
require at least one test case, and then
2. to document it equally comprehensively - a new subsection could be
added in the User Manual under YAML blueprints, with content taken
from the email below and beefed up for general readership
At the moment I don’t think the documentation is comprehensive enough
about all these details (as they work today), this could be a good
opportunity to improve it.
cheers
Geoff
————————————————————
Gnu PGP key - http://is.gd/TTTTuI
On 25 May 2016, at 12:44, Svetoslav Neykov
<[email protected]> wrote:
+1 for the proposal, definitely makes sense.
One thing that's not clear to me is how deep the merge should be.
Having templateOptions as an example I think it should be a shallow
merge. Can't think of deep complex structures passed in yaml that
would favour deep merge.
Re generalizing "$brooklyn:super()" - could have it as a string key
in maps that we want to merge. That is the owner of the map that's
doing the override can define whether he prefers merge or override.
It makes sense when developing blueprints because you know what the
catalog items being inherited are and can decide which way to go.
Svet.
On 25.05.2016 г., at 14:12, Aled Sage <[email protected]> wrote:
Hi all,
TL;DR: we should merge config when overriding entities/locations,
where it's obvious that such behaviour is desired. For example,
where an entity type defines shell.env, then a new entity extending
this type should inherit and add to those values.
_*REQUIREMENTS*_
_*shell.env in entities*_
When extending an existing entity type in YAML, it is not possible
to extend the set of environment variables. Instead, if the
sub-type declares shell.env it will override the inherited values.
For example, consider the catalog items below:
# Catalog
brooklyn.catalog:
items:
- id: machine-with-env
item:
type:
org.apache.brooklyn.entity.software.base.VanillaSoftwareProcess
brooklyn.config:
shell.env:
ENV1: myEnv1
# Blueprint
location: ...
services:
- type: machine-with-env
brooklyn.config:
shell.env:
ENV2: myEnv2
launch.command: echo "ENV1=$ENV1, ENV2=$ENV2"
A user might well expect the launch.command to have myEnv1 and
myEnv2. However, it does not get the ENV1 environment variable.
This is a real pain when trying to customize stock blueprints.
We propose that the shell.env map should be *merged*.
_*provisioning.properties*_
An entity can be configured with provisioning.properties. These are
passed to the location when obtaining a new machine. They
supplement and override the values configured on the location.
However, for templateOptions the expected/desired behaviour would
be to merge the options.
Consider the blueprint below:_*
*_
location:
minCores: 1
templateOptions:
networks: myNetwork
services:
- type: org.apache.brooklyn.entity.machine.MachineEntity
brooklyn.config:
provisioning.properties:
minRam: 2G
templateOptions:
tags: myTag
A user might well expect the VM to be created with the given
networks and tags. However, currently the templateOptions in
provisoining.properties will override the existing value, rather
than being merged with it.
We propose that the templateOptions map should be *merged*.
Valentin made a start to fix this in
https://github.com/apache/brooklyn-server/pull/151.
_*_*provisioning.properties in sub-entities*_
*_
A similar argument holds for when extending an entity-type in YAML.
If the super-type declares template options, then any additional
provisioning.properties declared on the entity sub-type should be
*merged* (including merging the templateOptions map contained
within it).
_*files.preinstall, templates.preinstall, etc*_
The same applies for the map config for: files.preinstall,
templates.preinstall, files.install, templates.install,
files.runtime and templates.runtime.
We propose that these maps get *merged* with the value defined in
the super-type.
_*Overriding default values*_
For default values in the super-type, we propose that this value
*does* get overridden, rather than merged.
For example, in the blueprint below we suggest that the
launch-command in the sub-type should have ENV2 but not
ENV_IN_DEFAULT.
brooklyn.catalog:
items:
- id: machine-with-env
version: 1.0.0
item:
type:
org.apache.brooklyn.entity.software.base.VanillaSoftwareProcess
brooklyn.parameters:
- name: shell.env
default:
ENV_IN_DEFAULT: myEnvInDefault
- id: machine-with-env-2
version: 1.0.0
item:
type: machine-with-env
brooklyn.config:
shell.env:
ENV2: myEnv2
launch.command: echo "ENV_IN_DEFAULT=$ENV_IN_DEFAULT,
ENV2=$ENV2"
(Interestingly, the current behaviour of machine-with-env is that
it gets the value for ENV_IN_DEFAULT but not for ENV2, so sometime
strange is going on with re-defining the shell.env config key!)
_*Extending commands: deferred*_
Another scenario is where a super-type declares a value for
`install.command`, and the sub-type wants to augment this by adding
additional commands. Currently that is not possible. Instead the
sub-type needs to use pre.install.command and/or
post.install.command. But that leads to the same problem if a
super-type also has a value defined for that key.
Svet suggested we could perhaps introduce something like
$brooklyn:super().
Unless we can generalise that approach to also solve the merging of
`shell.env` etc, then I suggest we defer the `install.command`
use-case. That can be proposed and discussed in a different thread.
However, if we can solve these problems with clever explicit use of
$brooklyn:super(), then that could provide an elegant solution to
all of these problems!
_*Inheritance from parent entities*_
Things are made yet more complicated by the fact we inherit config
from parent entities, in the entity hierarchy.
We propose that this behaviour is also configurable for the config
key, but that the defaults stay as they are. The existing logic is
applied to find the config value that applies to the given entity.
That value is then merged with its super-type, as appropriate.
For example, in the blueprint below... machine1 would get ENV1 and
ENV2 (i.e. the ENV1 definition overrides the ENV_IN_APP
definition). However, machine2 would get ENV1 and ENV_IN_APP (i.e.
it inherits ENV_IN_APP from the parent, and this is meged with the
super-type).
services:
- type: org.apache.brooklyn.entity.stock.BasicApplication
brooklyn.config:
shell.env:
ENV_IN_APP: myEnvInApp
brooklyn.children:
- type: machine-with-env
id: machine1
brooklyn.config:
shell.env:
ENV2: myEnv2
- type: machine-with-env
id: machine2
The reasoning behind this is to figure out the inheritance/override
rules incrementally. We leave the parent-inheritance as-is, and
just focus on the sub-typing inheritance.
Note that there is already a ConfigInheritance defined on ConfigKey
for controlling this kind of inheritance from the parent. The legal
values for ConfigInheritance are currently just ALWAYS and NONE.
_*IMPLEMENTATION*_
Clearly we do not want to implement this piecemeal. We'll add a way
to declare that a config key should be merged with that value from
the super-type.
We'll change the Java ConfigKey code to be:
public interface ConfigKey {
/**
* @since 0.10.0
*/
@Nullable ConfigInheritance getParentInheritance();
/**
* @since 0.10.0
*/
@Nullable ConfigInheritance getTypeInheritance();
/**
* @deprecated since 0.10.0; instead use {@link
#getParentInheritance()}
*/
@Nullable ConfigInheritance getInheritance();
}
We'll add to ConfigInheritance support for MERGE. We'll change the
name "ALWAYS" to OVERRIDE (deprecating the old value).
We'll change EntityConfigMap.getConfig to handle this new merge
behaviour. And same for locations, policies and enrichers.
Aled
--
Andrew Kennedy ; Founder clocker.io project ; @grkvlt ; Cloudsoft
