This use-case lines up almost exactly with how we deal with the Hadoop client libraries in the Hadoop core charms.
We handle this using the apache-hadoop-plugin subordinate charm. The subordinate charm provides the Hadoop libraries to the client, with the versions supported being locked in the plugin charm, and it does version matching when the plugin is connected to the Hadoop cluster to ensure that it is compatible with the cluster. Thus, the client doesn't need to worry about the Hadoop library version, and it isn't version-locked to a specific Hadoop library version. As long as the client supports the plugin relation and the plugin charm is connected to a compatible Hadoop cluster, the client knows that the provided libraries will work. The plugin charm also gives us a single charm to update with new library versions, while retaining the ability to test those libraries to ensure they work before updating the charm. There is one case of potential incompatibility that wasn't addressed in your original proposal nor in the plugin charm design as described: the case where the client itself has trouble using certain versions of the libraries. Luckily, this is relatively uncommon, and if it arises, we can address it using the same or similar "spec matching" logic that the plugin (and other Hadoop component charms) use to ensure version compatibility, as long as the plugin charm passes through the Hadoop version on the relation. On Thu, Jan 21, 2016 at 12:15 PM, Merlijn Sebrechts < merlijn.sebrec...@gmail.com> wrote: > I'm sorry, I wasn't clear about this. I wasn't consistent enough with my > usage of Charm VS Service. The library is not for the connection between > two charms. The library is for the connection between two services (the two > services that the two Charms deploy). > > I've seen it common for Java services to use a library to connect to > another java service. The actual network protocol is unstable between > versions, but the API of the supplied library is stable. > > 2016-01-21 17:24 GMT+01:00 Nate Finch <nate.fi...@canonical.com>: > >> I actually don't see why you would ever need to distribute a specific >> library for connecting to an API on another charm. The charm using the >> dynamically determined client would still require a static (i.e. backward >> compatible) API on the client library (otherwise the client charm would >> have no way to know how to use the client library)... so why not just put >> that static API on the server? and then you don't need a different client. >> Obviously, you can distribute new clients with new functionality, but by >> definition, you'd need new code in the client charm to understand how to >> use that functionality. >> >> On Thu, Jan 21, 2016 at 8:55 AM Merlijn Sebrechts < >> merlijn.sebrec...@gmail.com> wrote: >> >>> Hi Mark >>> >>> >>> I understand your concern. Leaking operational details is something that >>> should not be done in Juju. Apache proxy relationship cannot be implemented >>> by Nginx because the relation leaks operational details (apache specific >>> config files). This is not 'the Juju way'. >>> >>> The reason for sharing resources actually is interoperability. Say a >>> client wants to connect to a server, then the client may need libraries to >>> do that. What libraries will be used is dependent on both the client and >>> the server. If you ship the libraries with the client, you are essentially >>> hard-coding the server version. This is also not the Juju way; information >>> like this should be passed over relations. The problem here is that both >>> Charms have to make a joint decision about what libraries, if any, to use >>> and where to get them. No Charm has full knowledge to make that decision. >>> >>> I've seen two common patterns in the wild for these kind of libraries: >>> >>> 1. The client uses its own libraries compiled for that specific version >>> of the server. >>> 2. The client uses libraries bundled with the server. >>> >>> The first case is easy: The client has full knowledge about where to get >>> the libraries. It only needs to know the server version. The server tells >>> the client his version and it's done. The second case is a little bit >>> tougher. The client has knowledge about what libraries it needs, but the >>> server has knowledge about where to get those libraries. In this case, the >>> client should request the libraries from the server, and the server should >>> have a way to send them to the client. >>> >>> If both cases are supported by the server, you can swap any client in >>> and out. This also translates nicely to the idea that a Charm gets created >>> by the person who is an expert in that service. The server experts know >>> where to get the libraries if the server has to provide them. The client >>> expert knows how to compile the libraries if the client has to compile them. >>> >>> What do you think? >>> >>> >>> >>> >>> *PS: I don't think this should be done on stack level, since the shared >>> libraries are only relevant to the two related Charms. The relevance of the >>> libraries is on relation-level so they should be handled by the relations.* >>> >>> >>> >>> >>> 2016-01-21 13:58 GMT+01:00 Mark Shuttleworth <m...@ubuntu.com>: >>> >>>> >>>> Two thoughts. >>>> >>>> First, I think it's interesting that you see resources as such a dynamic >>>> thing. I believe the current model accommodates this (we considered >>>> user-provided resources, and what you are describing are essentially >>>> charm-generated resources, but they would behave the same way). >>>> >>>> Second, I am sceptical about cross-charm resource coordination. We go to >>>> great lengths to keep a service encapsulated. Each service is generally >>>> prevented from knowing too much about what is going on in the service >>>> next door. This is deliberate because it leads to the development of >>>> services which can more naturally be swapped out, because the web of >>>> expectations between related services is limited by what they can >>>> possibly know about each other. What you are describing suggests too >>>> intimate an awareness between two services of their exact operational >>>> implementation. >>>> >>>> That said, we do have an idea in the "futures" department, which is a >>>> higher-level charm that would manage a group of charms. While they would >>>> remain encapsulated, the higher-level charm would have admin-like >>>> visibility across all of the services it is supervising. Imagine an >>>> "openstack" charm which can command-and-control events across all the >>>> constituent services. In THAT case, yes, it would make sense for the >>>> higher level charm, which we call a "stack", to be able to coordinate >>>> resources across services under its supervision. >>>> >>>> How does that sound? >>>> >>>> Mark >>>> >>>> On 21/01/16 01:21, Merlijn Sebrechts wrote: >>>> > Hi John, Mark >>>> > >>>> > >>>> > If I understand you correctly, a Charm will be able to decide what >>>> version >>>> > of the resource it needs at runtime? This way, a Charm could tell >>>> related >>>> > Charms what version of the resource they should get? That would solve >>>> my >>>> > use-case almost completely... >>>> > >>>> > The only exception being the case where a Charm compiles a library >>>> during >>>> > installation and wants to distribute that binary to its related >>>> Charms. >>>> > This would require the Charm to be able to push the resource to the >>>> state >>>> > server and then distribute a link to that resource over its relations. >>>> > >>>> > Using the state server instead of rsync would be a lot better >>>> long-term, I >>>> > think. Network spaces and multi-tenancy might make it possible that a >>>> Charm >>>> > cannot ssh to a related Charm... >>>> > >>>> > >>>> > Any thoughts on this? >>>> > >>>> > >>>> > Kind regards >>>> > Merlijn Sebrechts >>>> > >>>> > 2016-01-21 7:42 GMT+01:00 John Meinel <j...@arbash-meinel.com>: >>>> > >>>> >> It does feel like a good fit for resources, with the one caveat that >>>> he >>>> >> wants to maintain a lock-step version of the resource across >>>> services. >>>> >> There is slightly more work with the current designs for resources, >>>> in that >>>> >> each charm will think about its version of the resource >>>> independently. But >>>> >> we will have the fingerprint information to allow for users to >>>> compare and >>>> >> be confident that both services are using the same resource. >>>> >> >>>> >> John >>>> >> =:-> >>>> >> >>>> >> On Wed, Jan 20, 2016 at 8:53 PM, Mark Shuttleworth <m...@ubuntu.com> >>>> >> wrote: >>>> >> >>>> >>> On 20/01/16 14:24, Merlijn Sebrechts wrote: >>>> >>>> So my question is: Is there a way to send large binary files >>>> between >>>> >>>> Charms? Or is this problem better solved by using a subordinate >>>> >>>> kafka-plugin Charm like the Hadoop Charms do? >>>> >>> It sounds like you want the new "Resources" capability coming in >>>> Juju 2.0 >>>> >>> :) >>>> >>> >>>> >>> For shared large blobs (like a JVM or a big ball of libraries) the >>>> charm >>>> >>> can ask the state server to cache the blob and distribute it to all >>>> the >>>> >>> units. There are mechanisms for users to supply the blob if needed, >>>> too. >>>> >>> >>>> >>> Mark >>>> >>> >>>> >>> -- >>>> >>> Juju mailing list >>>> >>> Juju@lists.ubuntu.com >>>> >>> Modify settings or unsubscribe at: >>>> >>> https://lists.ubuntu.com/mailman/listinfo/juju >>>> >>> >>>> >> >>>> >>>> >>>> >>> -- >>> Juju mailing list >>> Juju@lists.ubuntu.com >>> Modify settings or unsubscribe at: >>> https://lists.ubuntu.com/mailman/listinfo/juju >>> >> > > -- > Juju mailing list > Juju@lists.ubuntu.com > Modify settings or unsubscribe at: > https://lists.ubuntu.com/mailman/listinfo/juju > >
-- Juju mailing list Juju@lists.ubuntu.com Modify settings or unsubscribe at: https://lists.ubuntu.com/mailman/listinfo/juju
