Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On Wed, Jul 09, 2014 at 10:33:26PM +, Randall Burt wrote: On Jul 9, 2014, at 4:38 PM, Zane Bitter zbit...@redhat.com wrote: On 08/07/14 17:17, Steven Hardy wrote: Regarding forcing deployers to make a one-time decision, I have a question re cost (money and performance) of the Swift approach vs just hitting the Heat API - If folks use the Swift resource and it stores data associated with the signal in Swift, does that incurr cost to the user in a public cloud scenario? Good question. I believe the way WaitConditions work in AWS is that it sets up a pre-signed URL in a bucket owned by CloudFormation. If we went with that approach we would probably want some sort of quota, I imagine. Just to clarify, you suggest that the swift-based signal mechanism use containers that Heat owns rather than ones owned by the user? I guess it's probably best to just make this configurable, so the swift data can be either owned by the stack owner (same as all other resources, probably the default), or put in a container owned by the heat service user. The other approach is to set up a new container, owned by the user, every time. In that case, a provider selecting this implementation would need to make it clear to customers if they would be billed for a WaitCondition resource. I'd prefer to avoid this scenario though (regardless of the plug-point). Why? If we won't let the user choose, then why wouldn't we let the provider make this choice? I don't think its wise of us to make decisions based on what a theoretical operator may theoretically do. If the same theoretical provider were to also charge users to create a trust, would we then be concerned about that implementation as well? What if said provider decides charges the user per resource in a stack regardless of what they are? Having Heat own the container(s) as suggested above doesn't preclude that operator from charging the stack owner for those either. While I agree that these examples are totally silly, I'm just trying to illustrate that we shouldn't deny an operator an option so long as its understood what that option entails from a technical/usage perspective. I don't really get why this question is totally silly - I made a genuine request for education based on near-zero knowledge of public cloud provider pricing models. The reason for the question was simply that we're discussing the two alternate WaitCondition implementations which may, possibly, have very different implications from a cost perspective. I'm not saying we shouldn't let the provider make that decision on behalf of the user, just pointing out that if their particular use-case demands sending a gadzillion signals, they might like the option to choose the lightweight API signal approach. - What sort of overhead are we adding, with the signals going to swift, then in the current implementation being copied back into the heat DB[1]? I wasn't aware we were doing that, and I'm a bit unsure about it myself. I don't think it's a big overhead, though. In the current implementation, I think it is minor as well, just a few extra Swift API calls which should be pretty minor overhead considering the stack as a whole. Plus, it minimizes the above concern around potentially costly user containers in that it gets rid of them as soon as its done. It seems to me at the moment that the swift notification method is good if you have significant data associated with the signals, but there are advantages to the simple API signal approach I've been working on when you just need a simple one shot low overhead way to get data back from an instance. FWIW, the reason I revived these patches was I found that SoftwareDeployments did not meet my needs for a really simple signalling mechanism when writing tempest tests: https://review.openstack.org/#/c/90143/16/tempest/scenario/orchestration/test_volumes_create_from_backup.yaml These tests currently use the AWS WaitCondition resources, and I wanted a native alternative, without the complexity of using SoftwareDeployments (which also won't work with minimal cirros images without some pretty hacky workarounds[2]) Yep, I am all for this. I think that Swift is the best way when we have it, but not every cloud has Swift (and the latest rumours from DefCore are that it's likely to stay that way), so we need operators ( developers!) to be able to plug in an alternative implementation. Very true, but not every cloud has trusts either. Many may have trusts, but they don't employ the EC2 extensions to Keystone and therefore can't use the native signals either (as I understand them anyway). Point being that either way, we already impose requirements on a cloud you want to run Heat against. I think it in our interest to make the effort to provide choices with obvious trade-offs. The whole point of the resources I've
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On 10/07/14 05:34, Steven Hardy wrote: The other approach is to set up a new container, owned by the user, every time. In that case, a provider selecting this implementation would need to make it clear to customers if they would be billed for a WaitCondition resource. I'd prefer to avoid this scenario though (regardless of the plug-point). Why? If we won't let the user choose, then why wouldn't we let the provider make this choice? I don't think its wise of us to make decisions based on what a theoretical operator may theoretically do. If the same theoretical provider were to also charge users to create a trust, would we then be concerned about that implementation as well? What if said provider decides charges the user per resource in a stack regardless of what they are? Having Heat own the container(s) as suggested above doesn't preclude that operator from charging the stack owner for those either. While I agree that these examples are totally silly, I'm just trying to illustrate that we shouldn't deny an operator an option so long as its understood what that option entails from a technical/usage perspective. I don't really get why this question is totally silly - I made a genuine request for education based on near-zero knowledge of public cloud provider pricing models. The way I read it Randall was not saying that the question was silly, he was acknowledging that his own examples (like charging per-resource) were contrived (to the point of absurdity) to illustrate his argument. - ZB ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On Jul 10, 2014, at 9:21 AM, Zane Bitter zbit...@redhat.com wrote: On 10/07/14 05:34, Steven Hardy wrote: The other approach is to set up a new container, owned by the user, every time. In that case, a provider selecting this implementation would need to make it clear to customers if they would be billed for a WaitCondition resource. I'd prefer to avoid this scenario though (regardless of the plug-point). Why? If we won't let the user choose, then why wouldn't we let the provider make this choice? I don't think its wise of us to make decisions based on what a theoretical operator may theoretically do. If the same theoretical provider were to also charge users to create a trust, would we then be concerned about that implementation as well? What if said provider decides charges the user per resource in a stack regardless of what they are? Having Heat own the container(s) as suggested above doesn't preclude that operator from charging the stack owner for those either. While I agree that these examples are totally silly, I'm just trying to illustrate that we shouldn't deny an operator an option so long as its understood what that option entails from a technical/usage perspective. I don't really get why this question is totally silly - I made a genuine request for education based on near-zero knowledge of public cloud provider pricing models. The way I read it Randall was not saying that the question was silly, he was acknowledging that his own examples (like charging per-resource) were contrived (to the point of absurdity) to illustrate his argument. Yes. I didn't mean to imply the questions or any of the responses were silly, only my contrived examples. ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On 08/07/14 17:13, Angus Salkeld wrote: -BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 08/07/14 09:14, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're what's wrong with using the environment for that? Just have two resources and you do something like this: https://github.com/openstack/heat/blob/master/etc/heat/environment.d/default.yaml#L7 It doesn't cover other things that need signals, like software deployments (third-party plugin authors are also on their own). We only want n implementations not n*(number of resources that use signals) implementations. trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) Totally depends if we want this to be operator configurable (config file or plugin) or end user configurable (use their environment to choose the implementation). I just noticed, however, that there is an already-partially-implemented blueprint[3] and further pending patches[4] to use stevedore for *all* types of plugins - particularly resource plugins[5] - in Heat. I feel very strongly that stevedore is _not_ a good fit for all of those use cases. (Disclaimer: obviously I _would_ think that, since I implemented the current system instead of using stevedore for precisely that reason.) haha. The stated benefit of switching to stevedore is that it solves issues like https://launchpad.net/bugs/1292655 that are caused by the current convoluted layout of /contrib. I think the layout stems at least in part I think another great reason is consistency with how all other plugins are openstack are written (stevedore). Sure, consistency is nice, sometimes even at the expense of being not quite the right tool for the job. But there are limits to that trade-off. Also I *really* don't think we should optimize for our contrib plugins but for: 1) our built in plugins 2) out of tree plugins I completely agree, which is why I was surprised by this change. It seems to be deprecating a system that is working well for built-in and out-of-tree plugins in order to make minor improvements to how we handle contrib. from a misunderstanding of how the current plugin_manager works. The point of the plugin_manager is that each plugin directory does *not* have to be a Python package - it can be any directory. Modules in the directory then appear in the package heat.engine.plugins once imported. So there is no need to do what we are currently doing, creating a resources package, and then a parent package that contains the tests package as well, and then in the tests doing: from ..resources import docker_container ## noqa All we really need to do is throw the resources in any old directory, add that directory to the plugin_dirs list, stick the tests in any old package, and from the tests do from heat.engine.plugins import docker_container The main reason we haven't done this seems to be to avoid having to list the various contrib plugin dirs separately. Stevedore solves this by forcing us to list not only each directory but each class in each module in each directory separately. The tricky part of fixing the current layout is ensuring the contrib plugin directories get added to the plugin_dirs list during the unit tests and only during the unit tests. However, I'm confident that could be fixed with no more difficulty than the stevedore changes and with far less disruption to existing operators using custom plugins. Stevedore is ideal for configuring an implementation for a small number of well known plug points. It does not appear to be ideal for managing an application like Heat that comprises a vast collection of implementations of the same interface, each bound to its own plug point. I wouldn't call our resources vast. I count 73 in your patch, not including contrib and assuming you didn't miss any ;). It's seems clear to me that we're well past the point of what the Extensions API was designed for. When everything is an extension you need different tools to manage it. Quantity has a quality all it's own ;) Really, I think it works great. As discussed on IRC yesterday, we could potentially make the plugin the existing
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
I think my reply to Angus covers most of your points, except this one: On 08/07/14 17:39, Steve Baker wrote: On 09/07/14 07:08, Zane Bitter wrote: Constraints, I feel, are very similar to resources in this respect. I am less concerned about template formats, since there are so few of them... although it would be really nice to be able to install these as subpackages too, and using stevedore appears to eliminate that as an option :( To me constraints are more like client plugins. They make API calls and need specific knowledge of client exceptions. That is why I have been moving them into the client plugin modules. It would be preferable if they used the same plugin loading mechanism too. OK, cool, I'm persuaded by that. So they are like client plugin plugins? Should we consider doing something like with the intrinsic functions, where we just tie them to the client plugin? cheers, Zane. ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On Jul 9, 2014, at 3:15 PM, Zane Bitter zbit...@redhat.com wrote: On 08/07/14 17:13, Angus Salkeld wrote: -BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 08/07/14 09:14, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're what's wrong with using the environment for that? Just have two resources and you do something like this: https://github.com/openstack/heat/blob/master/etc/heat/environment.d/default.yaml#L7 It doesn't cover other things that need signals, like software deployments (third-party plugin authors are also on their own). We only want n implementations not n*(number of resources that use signals) implementations. trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) Totally depends if we want this to be operator configurable (config file or plugin) or end user configurable (use their environment to choose the implementation). I just noticed, however, that there is an already-partially-implemented blueprint[3] and further pending patches[4] to use stevedore for *all* types of plugins - particularly resource plugins[5] - in Heat. I feel very strongly that stevedore is _not_ a good fit for all of those use cases. (Disclaimer: obviously I _would_ think that, since I implemented the current system instead of using stevedore for precisely that reason.) haha. The stated benefit of switching to stevedore is that it solves issues like https://launchpad.net/bugs/1292655 that are caused by the current convoluted layout of /contrib. I think the layout stems at least in part I think another great reason is consistency with how all other plugins are openstack are written (stevedore). Sure, consistency is nice, sometimes even at the expense of being not quite the right tool for the job. But there are limits to that trade-off. Also I *really* don't think we should optimize for our contrib plugins but for: 1) our built in plugins 2) out of tree plugins I completely agree, which is why I was surprised by this change. It seems to be deprecating a system that is working well for built-in and out-of-tree plugins in order to make minor improvements to how we handle contrib. FWIW, when it comes to deploying Heat with non-built-in, there's no substantive difference in the experience between contrib and out-of-tree plugins, so neither system is more or less optimized for either. However, with the current system, there's no easy way to get rid of the built-in ones you don't want. from a misunderstanding of how the current plugin_manager works. The point of the plugin_manager is that each plugin directory does *not* have to be a Python package - it can be any directory. Modules in the directory then appear in the package heat.engine.plugins once imported. So there is no need to do what we are currently doing, creating a resources package, and then a parent package that contains the tests package as well, and then in the tests doing: from ..resources import docker_container ## noqa All we really need to do is throw the resources in any old directory, add that directory to the plugin_dirs list, stick the tests in any old package, and from the tests do from heat.engine.plugins import docker_container The main reason we haven't done this seems to be to avoid having to list the various contrib plugin dirs separately. Stevedore solves this by forcing us to list not only each directory but each class in each module in each directory separately. The tricky part of fixing the current layout is ensuring the contrib plugin directories get added to the plugin_dirs list during the unit tests and only during the unit tests. However, I'm confident that could be fixed with no more difficulty than the stevedore changes and with far less disruption to existing operators using custom plugins. Stevedore is ideal for configuring an implementation for a small number of well known plug points. It does not appear to be ideal for managing an application like Heat that comprises a vast collection of implementations of the same interface, each bound to its own plug point.
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 09/07/14 10:17, Zane Bitter wrote: On 08/07/14 17:13, Angus Salkeld wrote: -BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 08/07/14 09:14, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're what's wrong with using the environment for that? Just have two resources and you do something like this: https://github.com/openstack/heat/blob/master/etc/heat/environment.d/default.yaml#L7 It doesn't cover other things that need signals, like software deployments (third-party plugin authors are also on their own). We only want n implementations not n*(number of resources that use signals) implementations. trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) Totally depends if we want this to be operator configurable (config file or plugin) or end user configurable (use their environment to choose the implementation). I just noticed, however, that there is an already-partially-implemented blueprint[3] and further pending patches[4] to use stevedore for *all* types of plugins - particularly resource plugins[5] - in Heat. I feel very strongly that stevedore is _not_ a good fit for all of those use cases. (Disclaimer: obviously I _would_ think that, since I implemented the current system instead of using stevedore for precisely that reason.) haha. The stated benefit of switching to stevedore is that it solves issues like https://launchpad.net/bugs/1292655 that are caused by the current convoluted layout of /contrib. I think the layout stems at least in part I think another great reason is consistency with how all other plugins are openstack are written (stevedore). Sure, consistency is nice, sometimes even at the expense of being not quite the right tool for the job. But there are limits to that trade-off. Also I *really* don't think we should optimize for our contrib plugins but for: 1) our built in plugins 2) out of tree plugins I completely agree, which is why I was surprised by this change. It seems to be deprecating a system that is working well for built-in and out-of-tree plugins in order to make minor improvements to how we handle contrib. from a misunderstanding of how the current plugin_manager works. The point of the plugin_manager is that each plugin directory does *not* have to be a Python package - it can be any directory. Modules in the directory then appear in the package heat.engine.plugins once imported. So there is no need to do what we are currently doing, creating a resources package, and then a parent package that contains the tests package as well, and then in the tests doing: from ..resources import docker_container ## noqa All we really need to do is throw the resources in any old directory, add that directory to the plugin_dirs list, stick the tests in any old package, and from the tests do from heat.engine.plugins import docker_container The main reason we haven't done this seems to be to avoid having to list the various contrib plugin dirs separately. Stevedore solves this by forcing us to list not only each directory but each class in each module in each directory separately. The tricky part of fixing the current layout is ensuring the contrib plugin directories get added to the plugin_dirs list during the unit tests and only during the unit tests. However, I'm confident that could be fixed with no more difficulty than the stevedore changes and with far less disruption to existing operators using custom plugins. Stevedore is ideal for configuring an implementation for a small number of well known plug points. It does not appear to be ideal for managing an application like Heat that comprises a vast collection of implementations of the same interface, each bound to its own plug point. I wouldn't call our resources vast. I count 73 in your patch, not including contrib and assuming you didn't miss any ;). It's seems clear to me that we're well past the point of what the Extensions API was designed for. When everything is an extension you need different tools to manage it.
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 09/07/14 11:03, Randall Burt wrote: On Jul 9, 2014, at 3:15 PM, Zane Bitter zbit...@redhat.com wrote: On 08/07/14 17:13, Angus Salkeld wrote: -BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 08/07/14 09:14, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're what's wrong with using the environment for that? Just have two resources and you do something like this: https://github.com/openstack/heat/blob/master/etc/heat/environment.d/default.yaml#L7 It doesn't cover other things that need signals, like software deployments (third-party plugin authors are also on their own). We only want n implementations not n*(number of resources that use signals) implementations. trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) Totally depends if we want this to be operator configurable (config file or plugin) or end user configurable (use their environment to choose the implementation). I just noticed, however, that there is an already-partially-implemented blueprint[3] and further pending patches[4] to use stevedore for *all* types of plugins - particularly resource plugins[5] - in Heat. I feel very strongly that stevedore is _not_ a good fit for all of those use cases. (Disclaimer: obviously I _would_ think that, since I implemented the current system instead of using stevedore for precisely that reason.) haha. The stated benefit of switching to stevedore is that it solves issues like https://launchpad.net/bugs/1292655 that are caused by the current convoluted layout of /contrib. I think the layout stems at least in part I think another great reason is consistency with how all other plugins are openstack are written (stevedore). Sure, consistency is nice, sometimes even at the expense of being not quite the right tool for the job. But there are limits to that trade-off. Also I *really* don't think we should optimize for our contrib plugins but for: 1) our built in plugins 2) out of tree plugins I completely agree, which is why I was surprised by this change. It seems to be deprecating a system that is working well for built-in and out-of-tree plugins in order to make minor improvements to how we handle contrib. FWIW, when it comes to deploying Heat with non-built-in, there's no substantive difference in the experience between contrib and out-of-tree plugins, so neither system is more or less optimized for either. However, with the current system, there's no easy way to get rid of the built-in ones you don't want. drop a file in /etc/environment.d/ that has this in: OS::one_I_dont_want: You are saying that the resource should have no implementation. This disables the resource here. https://github.com/openstack/heat/blob/master/heat/engine/environment.py#L186-L202 - -A from a misunderstanding of how the current plugin_manager works. The point of the plugin_manager is that each plugin directory does *not* have to be a Python package - it can be any directory. Modules in the directory then appear in the package heat.engine.plugins once imported. So there is no need to do what we are currently doing, creating a resources package, and then a parent package that contains the tests package as well, and then in the tests doing: from ..resources import docker_container ## noqa All we really need to do is throw the resources in any old directory, add that directory to the plugin_dirs list, stick the tests in any old package, and from the tests do from heat.engine.plugins import docker_container The main reason we haven't done this seems to be to avoid having to list the various contrib plugin dirs separately. Stevedore solves this by forcing us to list not only each directory but each class in each module in each directory separately. The tricky part of fixing the current layout is ensuring the contrib plugin directories get added to the plugin_dirs list during the unit tests and only during the unit tests. However, I'm confident that could be fixed with no more difficulty than the stevedore changes and with far less
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On 08/07/14 17:17, Steven Hardy wrote: On Tue, Jul 08, 2014 at 03:08:32PM -0400, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) To clarify, they're not competing as such - Jason and I have chatted about the two approaches and have been working to maintain a common interface, such that they would be easily substituted based on deployer or user preferences. Yes, poor choice of words on my part :) My initial assumption was that this substitution would happen via resource mappings in the global environment, but I now see that you are proposing the configurable part to be at a lower level, subsituting the transport behind a common resource implementation. Yeah, exactly. Regarding forcing deployers to make a one-time decision, I have a question re cost (money and performance) of the Swift approach vs just hitting the Heat API - If folks use the Swift resource and it stores data associated with the signal in Swift, does that incurr cost to the user in a public cloud scenario? Good question. I believe the way WaitConditions work in AWS is that it sets up a pre-signed URL in a bucket owned by CloudFormation. If we went with that approach we would probably want some sort of quota, I imagine. The other approach is to set up a new container, owned by the user, every time. In that case, a provider selecting this implementation would need to make it clear to customers if they would be billed for a WaitCondition resource. I'd prefer to avoid this scenario though (regardless of the plug-point). - What sort of overhead are we adding, with the signals going to swift, then in the current implementation being copied back into the heat DB[1]? I wasn't aware we were doing that, and I'm a bit unsure about it myself. I don't think it's a big overhead, though. It seems to me at the moment that the swift notification method is good if you have significant data associated with the signals, but there are advantages to the simple API signal approach I've been working on when you just need a simple one shot low overhead way to get data back from an instance. FWIW, the reason I revived these patches was I found that SoftwareDeployments did not meet my needs for a really simple signalling mechanism when writing tempest tests: https://review.openstack.org/#/c/90143/16/tempest/scenario/orchestration/test_volumes_create_from_backup.yaml These tests currently use the AWS WaitCondition resources, and I wanted a native alternative, without the complexity of using SoftwareDeployments (which also won't work with minimal cirros images without some pretty hacky workarounds[2]) Yep, I am all for this. I think that Swift is the best way when we have it, but not every cloud has Swift (and the latest rumours from DefCore are that it's likely to stay that way), so we need operators ( developers!) to be able to plug in an alternative implementation. I'm all for making things simple, avoiding duplication and confusion for users, but I'd like to ensure that making this a one-time deployer level decision definitely makes sense, vs giving users some choice over what method is used. Agree, this is an important question to ask. The downside to leaving the choice to the user is that it reduces interoperability between clouds. (In fact, it's unclear whether operators _would_ give users a choice, or just deploy one implementation anyway.) It's not insurmountable (thanks to environments), but it does add friction to the ecosystem so we have to weigh up the trade-offs. cheers, Zane. [1] https://review.openstack.org/#/c/96947/ [2] https://review.openstack.org/#/c/91475/ Steve ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On 09/07/14 17:10, Angus Salkeld wrote: If we could make them separate Python packages within a single Git repo, I would be +2 on that. I don't know if that's feasible with our current tooling (though I guess it's not dissimilar to what you're doing with the contrib stuff in this patch series?). Another option is we could move away from contrib and have: heat/resources/builtin/all native resources here heat/resources/optional/{aws,contrib} The ones under optional have their own setup.cfg and need to be packaged (by the distro) separately. I'm not familiar enough with Python packaging to know if that works, but I'm fine with it if it does ;) Given the choice I'd probably marginally prefer to move the plugins to a separate tree (i.e. keep them in the same repo but not have them under heat/) though. We can default to resources are loaded but not enabled by default the each distro package can drop a file into/etc/environment.d/ that enables it's resources. Sounds like a recipe for distro bugs. What I was getting at was, we should separate the mechanism for loading the plugins and what is enabled/visible to the user. And that logic should really live in the resource registry in the environment. +1 One of the main design goals of the current resource plugins was to move the mapping from resource names to classes away from one central point (where all of the modules were imported) and place the configuration alongside the code it applies to. I am definitely not looking forward to having to go look up a config file to find out what each resource is every time I open the autoscaling module (and I do need to remind myself _every_ time I open it), to say nothing of the constant merge conflicts that we used to have to deal with when there was a central registry. They are grouped by name, so you will only run into an issue when someone else is working on the same area as you. A central registry is also problematic for operators that modify it, who will have a difficult, manual and potentially error-prone merge task to perform on the config file every time they upgrade. I don't see why an operator will be editing this, they should be using the environment to disable plugins/rename things. You don't have to touch this if you are adding your own plugin. Constraints, I feel, are very similar to resources in this respect. I am less concerned about template formats, since there are so few of them... although it would be really nice to be able to install these as subpackages too, and using stevedore appears to eliminate that as an option :( Do we want to move constraints to Hooks as well? Guessing yes, to make it consistent. I'm not too familiar with the constraints stuff, but I think yes - I doubt we want to be listing all of these in setup.cfg either. We can probably have one hook plugin for each client plugin to cover the built-in stuff, while still allowing out-of-tree plugins to join at the same hook point. cheers, Zane. ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On Jul 9, 2014, at 4:38 PM, Zane Bitter zbit...@redhat.com wrote: On 08/07/14 17:17, Steven Hardy wrote: Regarding forcing deployers to make a one-time decision, I have a question re cost (money and performance) of the Swift approach vs just hitting the Heat API - If folks use the Swift resource and it stores data associated with the signal in Swift, does that incurr cost to the user in a public cloud scenario? Good question. I believe the way WaitConditions work in AWS is that it sets up a pre-signed URL in a bucket owned by CloudFormation. If we went with that approach we would probably want some sort of quota, I imagine. Just to clarify, you suggest that the swift-based signal mechanism use containers that Heat owns rather than ones owned by the user? The other approach is to set up a new container, owned by the user, every time. In that case, a provider selecting this implementation would need to make it clear to customers if they would be billed for a WaitCondition resource. I'd prefer to avoid this scenario though (regardless of the plug-point). Why? If we won't let the user choose, then why wouldn't we let the provider make this choice? I don't think its wise of us to make decisions based on what a theoretical operator may theoretically do. If the same theoretical provider were to also charge users to create a trust, would we then be concerned about that implementation as well? What if said provider decides charges the user per resource in a stack regardless of what they are? Having Heat own the container(s) as suggested above doesn't preclude that operator from charging the stack owner for those either. While I agree that these examples are totally silly, I'm just trying to illustrate that we shouldn't deny an operator an option so long as its understood what that option entails from a technical/usage perspective. - What sort of overhead are we adding, with the signals going to swift, then in the current implementation being copied back into the heat DB[1]? I wasn't aware we were doing that, and I'm a bit unsure about it myself. I don't think it's a big overhead, though. In the current implementation, I think it is minor as well, just a few extra Swift API calls which should be pretty minor overhead considering the stack as a whole. Plus, it minimizes the above concern around potentially costly user containers in that it gets rid of them as soon as its done. It seems to me at the moment that the swift notification method is good if you have significant data associated with the signals, but there are advantages to the simple API signal approach I've been working on when you just need a simple one shot low overhead way to get data back from an instance. FWIW, the reason I revived these patches was I found that SoftwareDeployments did not meet my needs for a really simple signalling mechanism when writing tempest tests: https://review.openstack.org/#/c/90143/16/tempest/scenario/orchestration/test_volumes_create_from_backup.yaml These tests currently use the AWS WaitCondition resources, and I wanted a native alternative, without the complexity of using SoftwareDeployments (which also won't work with minimal cirros images without some pretty hacky workarounds[2]) Yep, I am all for this. I think that Swift is the best way when we have it, but not every cloud has Swift (and the latest rumours from DefCore are that it's likely to stay that way), so we need operators ( developers!) to be able to plug in an alternative implementation. Very true, but not every cloud has trusts either. Many may have trusts, but they don't employ the EC2 extensions to Keystone and therefore can't use the native signals either (as I understand them anyway). Point being that either way, we already impose requirements on a cloud you want to run Heat against. I think it in our interest to make the effort to provide choices with obvious trade-offs. I'm all for making things simple, avoiding duplication and confusion for users, but I'd like to ensure that making this a one-time deployer level decision definitely makes sense, vs giving users some choice over what method is used. Agree, this is an important question to ask. The downside to leaving the choice to the user is that it reduces interoperability between clouds. (In fact, it's unclear whether operators _would_ give users a choice, or just deploy one implementation anyway.) It's not insurmountable (thanks to environments), but it does add friction to the ecosystem so we have to weigh up the trade-offs. Agreed that this is an important concern, but one of mine is that no other resource has selectable back-ends. The way an operator controls this today is via the global environment where they have the option to disable one or more of these resources or even alias one to the other. Seems a large change for something an operator already has the ability to deal with.
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
Excerpts from Randall Burt's message of 2014-07-09 15:33:26 -0700: On Jul 9, 2014, at 4:38 PM, Zane Bitter zbit...@redhat.com wrote: On 08/07/14 17:17, Steven Hardy wrote: Regarding forcing deployers to make a one-time decision, I have a question re cost (money and performance) of the Swift approach vs just hitting the Heat API - If folks use the Swift resource and it stores data associated with the signal in Swift, does that incurr cost to the user in a public cloud scenario? Good question. I believe the way WaitConditions work in AWS is that it sets up a pre-signed URL in a bucket owned by CloudFormation. If we went with that approach we would probably want some sort of quota, I imagine. Just to clarify, you suggest that the swift-based signal mechanism use containers that Heat owns rather than ones owned by the user? +1, don't hide it. The other approach is to set up a new container, owned by the user, every time. In that case, a provider selecting this implementation would need to make it clear to customers if they would be billed for a WaitCondition resource. I'd prefer to avoid this scenario though (regardless of the plug-point). Why? If we won't let the user choose, then why wouldn't we let the provider make this choice? I don't think its wise of us to make decisions based on what a theoretical operator may theoretically do. If the same theoretical provider were to also charge users to create a trust, would we then be concerned about that implementation as well? What if said provider decides charges the user per resource in a stack regardless of what they are? Having Heat own the container(s) as suggested above doesn't preclude that operator from charging the stack owner for those either. This is a nice use case for preview. A user should be able to preview a stack and know what will be consumed. Wait conditions will show a swift container if preview is worth anything. ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On 09/07/14 18:33, Randall Burt wrote: On Jul 9, 2014, at 4:38 PM, Zane Bitter zbit...@redhat.com wrote: On 08/07/14 17:17, Steven Hardy wrote: Regarding forcing deployers to make a one-time decision, I have a question re cost (money and performance) of the Swift approach vs just hitting the Heat API - If folks use the Swift resource and it stores data associated with the signal in Swift, does that incurr cost to the user in a public cloud scenario? Good question. I believe the way WaitConditions work in AWS is that it sets up a pre-signed URL in a bucket owned by CloudFormation. If we went with that approach we would probably want some sort of quota, I imagine. Just to clarify, you suggest that the swift-based signal mechanism use containers that Heat owns rather than ones owned by the user? I'm suggesting that's one possible implementation, yes. The other approach is to set up a new container, owned by the user, every time. In that case, a provider selecting this implementation would need to make it clear to customers if they would be billed for a WaitCondition resource. I'd prefer to avoid this scenario though (regardless of the plug-point). Why? If we won't let the user choose, then why wouldn't we let the provider make this choice? I don't think its wise of us to make decisions based on what a theoretical operator may theoretically do. If the same theoretical provider were to also charge users to create a trust, would we then be concerned about that implementation as well? What if said provider decides charges the user per resource in a stack regardless of what they are? Having Heat own the container(s) as suggested above doesn't preclude that operator from charging the stack owner for those either. While I agree that these examples are totally silly, I'm just trying to illustrate that we shouldn't deny an operator an option so long as its understood what that option entails from a technical/usage perspective. Fair enough, I'm not suggesting that I want to deny the operator the option of charging, more that I'd prefer to avoid pushing them into a corner where they feel like they'd _have_ to charge. In fact, if we adopt the plugin system I am suggesting, we could in theory implement _both_ of the above options ;) - What sort of overhead are we adding, with the signals going to swift, then in the current implementation being copied back into the heat DB[1]? I wasn't aware we were doing that, and I'm a bit unsure about it myself. I don't think it's a big overhead, though. In the current implementation, I think it is minor as well, just a few extra Swift API calls which should be pretty minor overhead considering the stack as a whole. Plus, it minimizes the above concern around potentially costly user containers in that it gets rid of them as soon as its done. One of the nice things about this is that it largely negates the need for charging, by making the containers fairly useless for nefarious purposes. Unfortunately there are some problems with it too, which I've noted in the review. It seems to me at the moment that the swift notification method is good if you have significant data associated with the signals, but there are advantages to the simple API signal approach I've been working on when you just need a simple one shot low overhead way to get data back from an instance. FWIW, the reason I revived these patches was I found that SoftwareDeployments did not meet my needs for a really simple signalling mechanism when writing tempest tests: https://review.openstack.org/#/c/90143/16/tempest/scenario/orchestration/test_volumes_create_from_backup.yaml These tests currently use the AWS WaitCondition resources, and I wanted a native alternative, without the complexity of using SoftwareDeployments (which also won't work with minimal cirros images without some pretty hacky workarounds[2]) Yep, I am all for this. I think that Swift is the best way when we have it, but not every cloud has Swift (and the latest rumours from DefCore are that it's likely to stay that way), so we need operators ( developers!) to be able to plug in an alternative implementation. Very true, but not every cloud has trusts either. Many may have trusts, but they don't employ the EC2 extensions to Keystone and therefore can't use the native signals either (as I understand them anyway). Point being that either way, we already impose requirements on a cloud you want to run Heat against. I think it in our interest to make the effort to provide choices with obvious trade-offs. The AWS resources require EC2 extensions; Steve is making the native resources so that we won't require EC2 extensions any longer. When that is done, we should make the AWS resources use the same implementation (the implementation in AWS is closer to the Swift thing that Jason is working on than anything we have ever done). So part of what I'm suggesting here is that both
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 I wouldn't call our resources vast. I count 73 in your patch, not including contrib and assuming you didn't miss any ;). It's seems clear to me that we're well past the point of what the Extensions API was designed for. When everything is an extension you need different tools to manage it. Quantity has a quality all it's own ;) I think you might be on your own here: https://github.com/openstack/python-openstackclient/blob/master/setup.cfg https://github.com/openstack/nova/blob/master/setup.cfg https://github.com/openstack/ceilometer/blob/master/setup.cfg https://github.com/openstack/neutron/blob/master/setup.cfg - -A -BEGIN PGP SIGNATURE- Version: GnuPG v1 Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/ iQEcBAEBAgAGBQJTve08AAoJEFrDYBLxZjWoGoMH+wV5mRUketk5iH9FCVwmwq3o 8HaidJrOpXuK3vDXyO+DPBBw2dx3cvuXDuJIT6q1UnwP66+G5xQRaz/mqr6UZDLp WowL84ChrEcbwNJf5pMqXIpvG2sv9RZwz69MOQrwfodweTfNaVe3qBeapY1AIYvB 3f7EGuepmw8pIvQWMpVhmA/xCkrzfJ8Yj9VXwvUGYaLEADoyubNRp/EZG29nPP0/ sgdN4+Y+w+uG3AAbcXDP6M3ydiJAp7lswbOqsRQ+Lejc9UaKIZ03eis2WZQ24WzI qrO4mTbwBKlRXBk8BlSbwF0hQe0gz58pq80iZYQKRMMa515jwhUa9hPzrm8J0SI= =FY5b -END PGP SIGNATURE- ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
[openstack-dev] [Heat] stevedore plugins (and wait conditions)
I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) I just noticed, however, that there is an already-partially-implemented blueprint[3] and further pending patches[4] to use stevedore for *all* types of plugins - particularly resource plugins[5] - in Heat. I feel very strongly that stevedore is _not_ a good fit for all of those use cases. (Disclaimer: obviously I _would_ think that, since I implemented the current system instead of using stevedore for precisely that reason.) The stated benefit of switching to stevedore is that it solves issues like https://launchpad.net/bugs/1292655 that are caused by the current convoluted layout of /contrib. I think the layout stems at least in part from a misunderstanding of how the current plugin_manager works. The point of the plugin_manager is that each plugin directory does *not* have to be a Python package - it can be any directory. Modules in the directory then appear in the package heat.engine.plugins once imported. So there is no need to do what we are currently doing, creating a resources package, and then a parent package that contains the tests package as well, and then in the tests doing: from ..resources import docker_container ## noqa All we really need to do is throw the resources in any old directory, add that directory to the plugin_dirs list, stick the tests in any old package, and from the tests do from heat.engine.plugins import docker_container The main reason we haven't done this seems to be to avoid having to list the various contrib plugin dirs separately. Stevedore solves this by forcing us to list not only each directory but each class in each module in each directory separately. The tricky part of fixing the current layout is ensuring the contrib plugin directories get added to the plugin_dirs list during the unit tests and only during the unit tests. However, I'm confident that could be fixed with no more difficulty than the stevedore changes and with far less disruption to existing operators using custom plugins. Stevedore is ideal for configuring an implementation for a small number of well known plug points. It does not appear to be ideal for managing an application like Heat that comprises a vast collection of implementations of the same interface, each bound to its own plug point. For example, there's a subtle difference in how plugin_manager loads external modules - by searching a list of plugin directories for Python modules - and how stevedore does it, by loading a specified module already in the Python path. The latter is great for selecting one of a number of implementations that already exist in the code, but not so great for dropping in an additional external module, which now needs to be wrapped in a package that has to be installed in the path *and* there's still a configuration file to edit. This is way harder for a packager and/or operator to set up. This approach actually precludes a number of things we know we want to do in the future - for example it would be great if the native and AWS resource plugins were distributed as separate subpackages so that yum install heat-engine installed only the native resources, and a separate yum install heat-cfn-plugins added the AWS-compatibility resources. You can't (safely) package things that way if the installation would involve editing a config file. One of the main design goals of the current resource plugins was to move the mapping from resource names to classes away from one central point (where all of the modules were imported) and place the configuration alongside the code it applies to. I am definitely not looking forward to having to go look up a config file to find out what each resource is every time I open the autoscaling module (and I do need to remind myself _every_ time I open it), to say nothing of the constant merge conflicts that we used to have to deal with when there was a central registry. A central registry is also problematic for operators that modify it, who will have a difficult, manual and potentially
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 08/07/14 09:14, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're what's wrong with using the environment for that? Just have two resources and you do something like this: https://github.com/openstack/heat/blob/master/etc/heat/environment.d/default.yaml#L7 trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) Totally depends if we want this to be operator configurable (config file or plugin) or end user configurable (use their environment to choose the implementation). I just noticed, however, that there is an already-partially-implemented blueprint[3] and further pending patches[4] to use stevedore for *all* types of plugins - particularly resource plugins[5] - in Heat. I feel very strongly that stevedore is _not_ a good fit for all of those use cases. (Disclaimer: obviously I _would_ think that, since I implemented the current system instead of using stevedore for precisely that reason.) haha. The stated benefit of switching to stevedore is that it solves issues like https://launchpad.net/bugs/1292655 that are caused by the current convoluted layout of /contrib. I think the layout stems at least in part I think another great reason is consistency with how all other plugins are openstack are written (stevedore). Also I *really* don't think we should optimize for our contrib plugins but for: 1) our built in plugins 2) out of tree plugins from a misunderstanding of how the current plugin_manager works. The point of the plugin_manager is that each plugin directory does *not* have to be a Python package - it can be any directory. Modules in the directory then appear in the package heat.engine.plugins once imported. So there is no need to do what we are currently doing, creating a resources package, and then a parent package that contains the tests package as well, and then in the tests doing: from ..resources import docker_container ## noqa All we really need to do is throw the resources in any old directory, add that directory to the plugin_dirs list, stick the tests in any old package, and from the tests do from heat.engine.plugins import docker_container The main reason we haven't done this seems to be to avoid having to list the various contrib plugin dirs separately. Stevedore solves this by forcing us to list not only each directory but each class in each module in each directory separately. The tricky part of fixing the current layout is ensuring the contrib plugin directories get added to the plugin_dirs list during the unit tests and only during the unit tests. However, I'm confident that could be fixed with no more difficulty than the stevedore changes and with far less disruption to existing operators using custom plugins. Stevedore is ideal for configuring an implementation for a small number of well known plug points. It does not appear to be ideal for managing an application like Heat that comprises a vast collection of implementations of the same interface, each bound to its own plug point. I wouldn't call our resources vast. Really, I think it works great. For example, there's a subtle difference in how plugin_manager loads external modules - by searching a list of plugin directories for Python modules - and how stevedore does it, by loading a specified module already in the Python path. The latter is great for selecting one of a number of implementations that already exist in the code, but not so great for dropping in an additional external module, which now needs to be wrapped in a package that has to be installed in the path *and* there's still a configuration file to edit. This is way harder for a packager and/or operator to set up. I think you have this the wrong way around. With stevedore you don't need to edit a config file and with pluginmanager you do if that dir isn't in the list already. stevedore relies on namespaces, so you add your plugins into the heat.resources namespace and then heat will load them (no editing of config files). You do *not* need
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On Tue, Jul 08, 2014 at 03:08:32PM -0400, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) To clarify, they're not competing as such - Jason and I have chatted about the two approaches and have been working to maintain a common interface, such that they would be easily substituted based on deployer or user preferences. My initial assumption was that this substitution would happen via resource mappings in the global environment, but I now see that you are proposing the configurable part to be at a lower level, subsituting the transport behind a common resource implementation. Regarding forcing deployers to make a one-time decision, I have a question re cost (money and performance) of the Swift approach vs just hitting the Heat API - If folks use the Swift resource and it stores data associated with the signal in Swift, does that incurr cost to the user in a public cloud scenario? - What sort of overhead are we adding, with the signals going to swift, then in the current implementation being copied back into the heat DB[1]? It seems to me at the moment that the swift notification method is good if you have significant data associated with the signals, but there are advantages to the simple API signal approach I've been working on when you just need a simple one shot low overhead way to get data back from an instance. FWIW, the reason I revived these patches was I found that SoftwareDeployments did not meet my needs for a really simple signalling mechanism when writing tempest tests: https://review.openstack.org/#/c/90143/16/tempest/scenario/orchestration/test_volumes_create_from_backup.yaml These tests currently use the AWS WaitCondition resources, and I wanted a native alternative, without the complexity of using SoftwareDeployments (which also won't work with minimal cirros images without some pretty hacky workarounds[2]) I'm all for making things simple, avoiding duplication and confusion for users, but I'd like to ensure that making this a one-time deployer level decision definitely makes sense, vs giving users some choice over what method is used. [1] https://review.openstack.org/#/c/96947/ [2] https://review.openstack.org/#/c/91475/ Steve ___ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
Re: [openstack-dev] [Heat] stevedore plugins (and wait conditions)
On 09/07/14 07:08, Zane Bitter wrote: I see that the new client plugins are loaded using stevedore, which is great and IMO absolutely the right tool for that job. Thanks to Angus Steve B for implementing it. Now that we have done that work, I think there are more places we can take advantage of it too - for example, we currently have competing native wait condition resource types being implemented by Jason[1] and Steve H[2] respectively, and IMHO that is a mistake. We should have *one* native wait condition resource type, one AWS-compatible one, software deployments and any custom plugin that require signalling; and they should all use a common SignalResponder implementation that would call an API that is pluggable using stevedore. (In summary, what we're trying to make configurable is an implementation that should be invisible to the user, not an interface that is visible to the user, and therefore the correct unit of abstraction is an API, not a resource.) I just noticed, however, that there is an already-partially-implemented blueprint[3] and further pending patches[4] to use stevedore for *all* types of plugins - particularly resource plugins[5] - in Heat. I feel very strongly that stevedore is _not_ a good fit for all of those use cases. (Disclaimer: obviously I _would_ think that, since I implemented the current system instead of using stevedore for precisely that reason.) The stated benefit of switching to stevedore is that it solves issues like https://launchpad.net/bugs/1292655 that are caused by the current convoluted layout of /contrib. I think the layout stems at least in part from a misunderstanding of how the current plugin_manager works. The point of the plugin_manager is that each plugin directory does *not* have to be a Python package - it can be any directory. Modules in the directory then appear in the package heat.engine.plugins once imported. So there is no need to do what we are currently doing, creating a resources package, and then a parent package that contains the tests package as well, and then in the tests doing: from ..resources import docker_container ## noqa All we really need to do is throw the resources in any old directory, add that directory to the plugin_dirs list, stick the tests in any old package, and from the tests do from heat.engine.plugins import docker_container The main reason we haven't done this seems to be to avoid having to list the various contrib plugin dirs separately. Stevedore solves this by forcing us to list not only each directory but each class in each module in each directory separately. The tricky part of fixing the current layout is ensuring the contrib plugin directories get added to the plugin_dirs list during the unit tests and only during the unit tests. However, I'm confident that could be fixed with no more difficulty than the stevedore changes and with far less disruption to existing operators using custom plugins. There is a design document for stevedore which does a good job of covering all the options for designing a plugin system: http://stevedore.readthedocs.org/en/latest/essays/pycon2013.html Stevedore is ideal for configuring an implementation for a small number of well known plug points. It does not appear to be ideal for managing an application like Heat that comprises a vast collection of implementations of the same interface, each bound to its own plug point. Resource plugins seems to match stevedore's Extensions pattern reasonably well http://stevedore.readthedocs.org/en/latest/patterns_loading.html For example, there's a subtle difference in how plugin_manager loads external modules - by searching a list of plugin directories for Python modules - and how stevedore does it, by loading a specified module already in the Python path. The latter is great for selecting one of a number of implementations that already exist in the code, but not so great for dropping in an additional external module, which now needs to be wrapped in a package that has to be installed in the path *and* there's still a configuration file to edit. This is way harder for a packager and/or operator to set up. Just dropping in a file is convenient, but maybe properly packaging resource plugins is a discipline we should be encouraging third parties to adopt. This approach actually precludes a number of things we know we want to do in the future - for example it would be great if the native and AWS resource plugins were distributed as separate subpackages so that yum install heat-engine installed only the native resources, and a separate yum install heat-cfn-plugins added the AWS-compatibility resources. You can't (safely) package things that way if the installation would involve editing a config file. Yes, patching a single setup.cfg is a non-starter. We would need to do something like move the AWS resources to contrib with their own setup.cfg; maybe that wouldn't be such a