I have been following this discussion thread as part of the storm-core-ui migration. I would like to bring up a couple of points:
* The names of the packages "storm-client" and "storm-server" are a bit misleading to me. Isn’t what we really mean here "storm-workers" and "storm-daemons” ? Even if not these names, we should pick names that as close as possible to the “physical system”. * storm-client-misc * I noticed that this module only has two classes [1]. They are currently used in the module storm-starter and nowhere else. If that is the case, we should just put the classes in the module storm-starter. The concern is if some users may be using them in their deployments. Do you know of any users using these classes? Perhaps we could poll us...@storm.apache.org<mailto:us...@storm.apache.org> and find out. * the -misc extension is also very confusing to me. My first thought was that it was some sort of library dependency placeholder, or something like that. If at all possible, my suggestion would be for us to eliminate this module altogether. * Since we Storm 2.0 is a major release, if we find out that not many users (maybe none) are using the classes [1] we could probably just put the classes HttpForwardingMetricsConsumer, HttpForwardingMetricsServer in storm-starter. As for the concern of breaking backwards compatibility, document a workaround using storm-starter. Thanks, Hugo [1] - HttpForwardingMetricsConsumer, HttpForwardingMetricsServer On Jul 31, 2017, at 6:51 AM, Bobby Evans <ev...@yahoo-inc.com.INVALID<mailto:ev...@yahoo-inc.com.INVALID>> wrote: Those look reasonable to me. - Bobby On Monday, July 31, 2017, 2:22:47 AM CDT, Jungtaek Lim <kabh...@gmail.com<mailto:kabh...@gmail.com>> wrote: I agreed to minimize the target of shade & relocation artifacts minimal as possible, but as we shaded almost everything (meaning non-relocation will affect user experience) so may need to find exhaustive set of troublesome artifacts and relocate at least them. (Maybe union of everyone's lists?) For me Guava, HttpClient, Netty (maybe no need to shade for now if we don't plan to upgrade to 4.x: package name differs) is in my list. Would be better to initiate poll or discussion with separate thread? - Jungtaek Lim (HeartSaVioR) 2017년 7월 20일 (목) 오전 2:27, Bobby Evans <ev...@yahoo-inc.com.invalid<mailto:ev...@yahoo-inc.com.invalid>>님이 작성: I am fine with a separate project for relocated dependencies (or even just separate packages, you do a maven install of them and not include them in the IDE at all). Shading still has some drawbacks, but I think in a few cases it makes since. I would prefer it if we picked a very small number of dependencies that cause people issues and just shade those. Guava is the big one that I worry about. Netty is a possibility and I think asm would be another, but it is a transitive dependency so it would require us with our own version of kryo exposing the kryo API but pulling in a shaded asm. The servlet-api concerns me, but it looks like it is tied to the IHttpCredentialsPlugin which should move to the server package anyways. The rest I am not concerned about, are things that are exposed to end users, or are for test and not actually shipped. $ mvn dependecy:tree... [INFO] --- maven-dependency-plugin:2.8:tree (default-cli) @ storm-client --- [INFO] org.apache.storm:storm-client:jar:2.0.0-SNAPSHOT [INFO] +- uk.org.lidalia:sysout-over-slf4j:jar:1.0.2:compile [INFO] +- org.slf4j:slf4j-api:jar:1.7.21:compile [INFO] +- org.apache.logging.log4j:log4j-api:jar:2.8.2:compile [INFO] +- org.apache.logging.log4j:log4j-core:jar:2.8.2:compile [INFO] +- org.apache.logging.log4j:log4j-slf4j-impl:jar:2.8.2:compile [INFO] +- org.slf4j:log4j-over-slf4j:jar:1.6.6:compile [INFO] +- com.google.guava:guava:jar:16.0.1:compile [INFO] +- org.apache.thrift:libthrift:jar:0.9.3:compile [INFO] | \- org.apache.httpcomponents:httpcore:jar:4.4.1:compile [INFO] +- commons-io:commons-io:jar:2.5:compile [INFO] +- commons-lang:commons-lang:jar:2.5:compile [INFO] +- commons-collections:commons-collections:jar:3.2.2:compile [INFO] +- com.lmax:disruptor:jar:3.3.2:compile [INFO] +- com.googlecode.json-simple:json-simple:jar:1.1:compile [INFO] +- org.yaml:snakeyaml:jar:1.11:compile [INFO] +- io.netty:netty:jar:3.9.0.Final:compile [INFO] +- com.esotericsoftware:kryo:jar:3.0.3:compile [INFO] | +- com.esotericsoftware:reflectasm:jar:1.10.1:compile [INFO] | | \- org.ow2.asm:asm:jar:5.0.3:compile [INFO] | +- com.esotericsoftware:minlog:jar:1.3.0:compile [INFO] | \- org.objenesis:objenesis:jar:2.1:compile [INFO] +- org.apache.zookeeper:zookeeper:jar:3.4.6:compile [INFO] | \- jline:jline:jar:0.9.94:compile [INFO] +- org.apache.curator:curator-framework:jar:2.12.0:compile [INFO] +- org.jgrapht:jgrapht-core:jar:0.9.0:compile [INFO] +- javax.servlet:servlet-api:jar:2.5:compile [INFO] +- org.apache.httpcomponents:httpclient:jar:4.3.3:compile [INFO] | +- commons-logging:commons-logging:jar:1.1.3:compile [INFO] | \- commons-codec:commons-codec:jar:1.6:compile [INFO] +- org.apache.curator:curator-client:jar:2.12.0:compile [INFO] +- junit:junit:jar:4.11:test [INFO] | \- org.hamcrest:hamcrest-core:jar:1.3:test [INFO] +- org.mockito:mockito-core:jar:1.9.5:test [INFO] \- org.hamcrest:hamcrest-library:jar:1.3:test - Bobby On Wednesday, July 12, 2017, 9:45:43 AM CDT, Jungtaek Lim < kabh...@gmail.com<mailto:kabh...@gmail.com>> wrote: I'd like to bump on this again, since we have a few huge issues for Storm 2.0.0, and this issue is a kind of regression and effectively blocker. (Please note that current master branch removes shading for some libraries to make IDE happy.) At that time I didn't consider option 2 as possible solution, but now Flink is going with this option, and I can't find reason to not doing this. * Repository: https://github.com/apache/flink-shaded * Discussion thread: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-Changing-Flink-s-shading-model-td17419.html Thought? Thanks, Jungtaek Lim (HeartSaVioR) 2017년 3월 31일 (금) 오후 3:12, Jungtaek Lim <kabh...@gmail.com>님이 작성: Bobby, I've worked on separating worker and daemon classpath. - Issue: STORM-2441: Break down 'storm-core' to extract client (worker) artifacts <http://issues.apache.org/jira/browse/STORM-2441> - PR: https://github.com/apache/storm/pull/2034 I don't address your suggestion about "classpath selection" and "hiding local mode". Please file issues if you would like to address. Btw, I exclude artifacts from shade & relocation list so still need to address dependency issue. Folks, any other ideas or opinions around dependency issue? IMHO Option 2 is clearer but not sure where we can create a new git repo (ASF git or even outside), and also it's not against LICENSEs to repackage shade & relocated artifacts to Maven. Thanks, Jungtaek Lim (HeartSaVioR) 2017년 3월 29일 (수) 오후 10:42, Bobby Evans <ev...@yahoo-inc.com.invalid>님이 작성: I am fine with those changes so long as we finish the separation of worker and daemon classpaths. Otherwise we have made some very big changes for our end users that are going to have a hard time upgrading. If all we support is the option to run an old worker version with a new supervisor/nimbus I think that would be good enough, although I would like to see a full separation of the classpaths. - Bobby On Tuesday, March 28, 2017, 6:03:26 PM CDT, Jungtaek Lim < kabh...@gmail.com> wrote:Just FYI: I've worked with minimal patch for 3, though I still don't like such workaround: https://github.com/HeartSaVioR/storm/commit/d3122faa7ae182915242b979beaac156f91fe3b2 It excludes 'libthrift', 'jetty', 'codahale metrics' from relocation targets. I can see IDEA is OK to build the project, and Maven build passing. - Jungtaek Lim (HeartSaVioR) 2017년 3월 29일 (수) 오전 11:02, Jungtaek Lim <kabh...@gmail.com>님이 작성: Back to origin issue (before breaking down 'storm-core'), turned out IntelliJ doesn't recognize relocated classes within project. That's why build (via Maven) for master branch succeeds but IDEA compile doesn't. There're some issues filed but no action has been made. https://youtrack.jetbrains.com/issue/IDEA-93855 https://youtrack.jetbrains.com/issue/IDEA-126596 So suppose we have two modules A and B within project, and A relocates L to Lr. B relies on A's method which returns a class of Lr or has parameters for a class of LR, B needs to use Lr rather than L, and Lr is not recognized from B. Moving 'storm-drpc-server' to 'storm-core' may help but it's not a nice solution though. (think about why we add new module 'storm-drpc-server') To minimize dependency for worker (which actually affects end users) we should break down 'storm-core' and it will remain to be headache. There seemed to be little workarounds. 1. Guide IDEA users to take hacky workaround. Quoting https://youtrack.jetbrains.com/issue/IDEA-93855#comment=27-1838157 : "A hacky workaround is to make the module in intellij with the dependency depend on the jar explicitly in target/. This at least allows things to compile and tests to run." That is really bad and annoying, but we might have no choice when we don't want to take other workarounds. 2. Maintaining separate project for relocated dependencies. This avoids contributors to take hacky workaround so good to go, but maintaining relocated artifacts might be another headache, and I'm not sure ASF (or LICENSE of relocated targets) allows to do that. 3. Minimize (or remove) relocate targets and/or don't relocate troublesome targets. For 'storm-drpc-server', there seems to be three troublesome targets: - 'thrift' - 'codahale metrics' - 'jetty server' (We may be able to move this to 'storm-drpc-server' when another webapp port is done.) If we are OK to give up relocating those things we might be OK for now. We may want to extend the list when we break down more modules from 'storm-core'. Btw, IMHO relocating is not a good option. Elastic gives up shading anything for 2.0. ( https://www.elastic.co/blog/to-shade-or-not-to-shade) Someone might feel that it's a regression, but we need to decide to do it when it can provide better shape. Please add ideas if you have any, and give your opinions about above options. Thanks, Jungtaek Lim (HeartSaVioR) 2017년 3월 28일 (화) 오후 10:23, Bobby Evans <ev...@yahoo-inc.com.invalid>님이 작성: Sure I am happy to help out how I can. I really would like to spend more time on storm, but sadly work has shifted and my team got 2 new projects recently, but we have not increased the head count to cover it yet, so I am swamped. But if you do need help with some of these let me know and I'll see what I can do in my spare time. - Bobby On Tuesday, March 28, 2017, 2:10:46 AM CDT, Jungtaek Lim < kabh...@gmail.com> wrote:Bobby, I just tried to follow your suggestion and found it's less error-prone compared to my approach, and has lots of benefits. (I am seeing the great chance to minimize dependencies for 'storm-client', say, Worker.) Thanks for the suggestion. I'm working on this now. I'll mention you when I finish working this, or need your help. Thanks, Jungtaek Lim (HeartSaVioR) 2017년 3월 28일 (화) 오전 8:15, Jungtaek Lim <kabh...@gmail.com>님이 작성: I think we could also fix this issue for separating 'storm-core' and 'storm-webapp' (rename from 'storm-drpc-server'), since local cluster doesn't need to have 'storm-webapp', DRPC server (local DRPC will still be in 'storm-core'), UI, Logviewer. That's what I'm working on, which seems to require heavy efforts. Your plan looks really promising, but in other perspective this plan is even much harder to address. Do you have time frame for working on this? If you can finish the work in time frame so that it can be included in 2.0.0, I'll just discard my work and move forward to port other things (logviewer, ui) first. Regarding local mode, exposing local mode provides easy debug functionality with IDE, and hiding it takes away such functionality. We have ConfigurableTopology for 2.0.0 which helps to remove ceremony code, so exposing is not that bad. Thanks, Jungtaek Lim (HeartSaVioR) 2017년 3월 27일 (월) 오후 11:10, Bobby Evans <ev...@yahoo-inc.com.invalid>님이 작성: +1 for fixing dependency/IDE issues, but I am not sure it is as simple as what you describe. The issue is that there is no clean way to get local mode without pulling in almost all of the daemons too. If we are going to go through the pain of separating them out, I would prefer to do it once and do it right. I am happy to help out with this, as it is something I have been thinking about for a while, but just haven't found the time to tackle on my own. First we need a good way to give a control to our users about the base classpath of the worker, ideally the JVM version too. We have been doing a really good job with rolling upgrades and I think it would be great if we could have multiple versions of storm/JVM installed on the worker nodes and the end user can pick what JVM and what version of storm they want their worker to run with. We can argue over details of how that would work later. The point is that it lets us make changes to the classpath in very drastic ways without breaking end users. Second we need a better way to hide local mode. Every example we have supports local mode which means we will ship a copy of the storm daemons in each topology jar if we pull them out of the default classpath. We need to be able to run existing topologies that do not have "local mode support" in local mode. We should be able to make storm-submitter work, there are already stubs for this kind of thing, but we may need to play around with DRPC and a few other APIs to make it transparent. We then create new jars from the existing storm-core and storm-drpc-server. storm-client - Just what the client and worker needs. The only external dependencies are logging and possibly metrics. storm-local - This would pull in local mode dependencies (almost everything in storm core). We might even make it a test jar. storm-daemon - all of our daemon processes (most if not all shading removed). We can subdivide this more if we want to. storm-core would go away or just pull in storm-client. The storm jar command would by default only pull in storm-client and its dependencies. If you wanted local mode you could add in a flag that would adjust the classpath, boot up a local mode cluster, change the client to transparently interact with that instead of a regular cluster, and jump to the end users main. There could also be an option to just include everything on the classpath without the local mode cluster. Ideally if we include everything on the classpath with storm jar, that would also add a flag that would make the supervisor include everything on the classpath when launching the worker. - Bobby On Monday, March 27, 2017, 12:11:44 AM CDT, Jungtaek Lim < kabh...@gmail.com> wrote:Hi devs, I took a first step of finalizing port work via resolving dependency issue with DRPC. Here's what I'm giving a try: - rename 'storm-drpc-server' to 'storm-webapp' - remove 'storm-core' from 'storm-drpc-server' -- 'storm-drpc-server' will have its own library directory or shaded jar - create 'storm-common' and extract all the things used for both 'storm-core' and 'storm-webapp' It requires numerous files to be moved to, and huge code block should be moved / modified. A bit painful to work on. Other approach would be separating 'storm-worker' (or 'storm-client') and 'storm-daemon', and link to different libraries directory. (Maybe we could make uber jar for 'storm-daemon'.) This also requires similar work and maybe introduce more big effect to users. Other than above ideas I don't have any other ideas. We're shading libraries which are both needed from 'storm-core' and 'storm-drpc-server' which in turn makes known issue - able to build with maven but IDE can't compile 'storm-drpc-server' project. Please share other ideas if you have one. Thanks, Jungtaek Lim (HeartSaVioR)
