Sorry Longda, but I can't help telling that I also disagree about changing codebase.
Feature matrix shows us how far Apache Storm and JStorm are diverged, just in point of feature's view. We can't be safe to change although feature matrixes are identical, because feature matrix doesn't contain the details. I mean, users could be scared when expected behaviors are not in place although they're small. User experience is the one of the most important part of the project, and if UX changes are huge, barrier for upgrading their Storm cluster to 2.0 is not far easier than migrating to Heron. It should be the worst scenario I can imagine after merging. The safest way to merge is applying JStorm's great features to Apache Storm. I think porting language of Apache Storm to Java is not tightly related to merge JStorm. I agree that merging becomes a trigger, but Apache Storm itself can port to other languages like Java, Scala, or something else which are more popular than Clojure. And I'm also not scary about Flink, Heron, Spark, etc. It doesn't mean other projects are not greater then Storm. Just I'm saying each projects have their own strength. For example, all conferences are saying about Spark, and as one of users of Spark, Spark is really great. If you are a little bit familiar with Scala, you can just apply Scala-like functional methods to RDD. Really easy to use. But it doesn't mean that Spark can replace Storm in all kind of use cases. Recently I've seen some articles that why Storm is more preferred in realtime streaming processing. Competition should give us a positive motivation. I hope that our roadmap isn't focused to defeat competitors, but is focused to present great features, better performance, and better UX to Storm community. It's not commercial product, it's open source project! tl;dr. Please don't change codebase unless we plan to release a brand new project. It breaks UX completely which could make users leave. I'm also open to other opinions as well. Best, Jungtaek Lim (HeartSaVioR) 2015-11-20 0:00 GMT+09:00 Bobby Evans <[email protected]>: > I disagree completely. You claim that JStorm is compatible with storm > 1.0. I don't believe that it is 100% compatible. There has been more then > 2 years of software development happening on both sides. Security was not > done in a day, and porting it over to JStorm is not going to happen > quickly, and because of the major architectural changes between storm and > JStorm I believe we would have to make some serious enhancements to fully > support a secure TopologyMaster, but I need to look into it more. The blob > store is another piece of code that has taken a very long time to develop. > There are numberous others. The big features are not the ones that make me > nervous because we can plan for them, it is the hundreds of small JIRA and > features that will result in minor incompatibilities. If we start with > storm itself, and follow the same process that we have been doing up until > now, if there is a reason to stop the port add in an important feature and > do a release, we can. We will know that we have compatibility vs starting > with JStorm we know from the start that we do not without adding feature X, > Y, Z, .... > > I personally am not scared of Flink, Heron, Spark, Apex, IBM Streams, > etc... We just did some major performance enhancements that will be > released with STORM 1.0. We now have up to 6x the throughput that we had > before with minimal changes to the latency (20 ms vs 5 ms). We have > automatic back-pressure so if someone was running with acking enabled just > for flow control they can now process close to 16x the throughput they > could before with the same hardware. This puts our throughput very much on > par with flink and Spark, but with a much lower latency compared to either > of them. Plus from what I have heard Flink is still calling the streaming > API beta, and their storm API compatibility is very rudimentary. They are > also going to have more and more problems maintaining compatibility as we > add in new features and functionality. > > Spark only really works well when it is running with several seconds of > latency. Not every one needs sub-second processing, but when your platform > is completely unable to handle it, locks you out of a lot of use cases. > Their throughput is decent and can scale very high when you are willing to > tolerate similarly very high latencies. > Who knows about Heron until they actually release their code, but it is > missing lots of critical features, and the one they touted, better > performance, is a moot point with storm 1.0. The only thing we really are > lacking is advertising, we don't have a big company really pushing storm > and getting it in the news all the time (Sorry Hortonworks, but I really > have not seen much about it in the news). I am trying to do more, but > there is only so much I can do. > Longda I very much agree with you about moving quickly to make the > transition, but I do not believe in any way that starting with JStorm is > going to reduce that transition time. > My proposal is to give everyone about 2 weeks to finish merging new > features into Storm 1.0. On Dec 1st we create a storm-1.x branch and call > for a release. At the same time development work to port storm to java > begins. You said it took 4 developers 1 year to port storm to java the > first time for JStorm. We have 14+ active developers and over one hundred > contributors not including those from the JStorm community. If numbers > scale linearly, I know they don't completely, we should be able to do a > complete port with no JStorm reference in around 100 days. With a copy and > paste for a lot of this from the JStorm codebase, I would expect to be able > to do it in 1 month of development, possibly less if the JStorm community > can really help out too. So by January we should be ready to begin pulling > in features from JStorm that make since. Looking at the feature matrix in > https://github.com/apache/storm/pull/877 there are a few potentially big > improvements that we would want to pull in, but they require architectural > changes in some cases that I don't want to just do lightly. I would > propose that one the code has been ported to java we reopen for all new > features in parallel with the JStorm feature migration, but I am open to > others opinions as well. > - Bobby > > > On Thursday, November 19, 2015 12:14 AM, Longda Feng < > [email protected]> wrote: > > > Sorry for changing the Subject. > > > I am +1 for releasing Storm 2.0 with java core, which is merged with JStorm. > > > I think the change of this release will be the biggest one in history. It > will probably take a long time to develop. At the same time, Heron is going > to open source, and the latest release of Flink provides the compatibility to > Storm’s API. These might be the threat to Storm. So I suggest we start the > development of Storm 2.0 as quickly as possible. In order to accelerate the > development cycle, I proposed to take JStorm 2.1.0 core and UI as the base > version since this version is stable and compatible with API of Storm 1.0. > Please refer to the phases below for the detailed merging plan. > > Note: We provide a demo of JStorm’s web UI. Please refer to > storm.taobao.org > . I think JStorm will give a totally different view to you. > > > I would like to share the experience of initial development of JStorm > (Migrate from clojure core to java core). > > Our team(4 developers) have spent almost one year to finish the migration. We > took 4 months to release the first JStorm version, and 6 months to make > JStorm stable. During this period, we tried to switch more than online 100 > applications with different scenarios from Storm to JStorm, and many bugs > were fixed. Then more and more applications were switched to JStorm in > Alibaba. > > Currently, there are 7000+ nodes of JStorm clusters in Alibaba and 2000+ > applications are running on them. The JStorm Clusters here can handle 1.5 > PB/2 Trillion messages per day. The use cases are not only in BigData field > but also in many other online scenarios. > > Besides it, we have experienced the November 11th Shopping Festival of > Alibaba for last three years. At that day, the computation in our cluster > increased several times than usual. All applications worked well during the > peak time. I can say the stability of JStorm is no doubt today. Actually, > besides Alibaba, the most powerful Chinese IT company are also using JStorm. > > > Phase 1: > > Define the target of Storm 2.0. List the requirement of Storm 2.0 > 1. Open a new Umbrella Jira ( > https://issues.apache.org/jira/browse/STORM-717) > 2. Create one 2.0 branch, > 2.1 Copy modules from JStorm, one module from one module > > 2.2 The sequence is extern modules/client/utils/nimbus/supervisor/drpc/worker > & task/web ui/others > 3. Create jira for all differences between JStorm 2.1.0 and Storm 1.0 > 3.1 Discuss solution for each difference(jira) > > 3.2 Once the solution is finalized, we can start the merging. (Some issues > could be start concurrently. It depends on the discussion.) > > > The phase mainly try to define target and finalize the solution. Hopefully > this phase could be finished in 2 month(before 2016/1/31). . > > > Phase 2: > Release Storm 2.0 beta > 1. Based on phrase 1's discussion, finish all features of Storm 2.0 > > 2. Integrate all modules, make the simplest storm example can run on the > system. > 3. Test with all example and modules in Storm code base. > 4. All daily test can be passed. > > Hopefully this phase could be finished in 2 month(before 2016/3/31) > > > Phase 3: > Persuade some user to have a try. > Alibaba will try to run some online applications on the beta version > > Hopefully this phase could be finished in 1 month(before 2016/4/31). > > > Any comments are welcome. > > > Thanks > Longda------------------------------------------------------------------From:P. > Taylor Goetz <[email protected]>Send Time:2015年11月19日(星期四) 06:23To:dev < > [email protected]>，Bobby Evans <[email protected]>Subject:Re: > [DISCUSS] Plan for Merging JStorm Code > > All I have at this point is a placeholder wiki entry [1], and a lot of local > notes that likely would only make sense to me. > > > Let me know your wiki username and I’ll give you permissions. The same goes > for anyone else who wants to help. > > -Taylor > > [1] > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=61328109 > > > On Nov 18, 2015, at 2:08 PM, Bobby Evans <[email protected] > > wrote: > > > > > Taylor and others I was hoping to get started filing JIRA and planning on > > how we are going to do the java migration + JStorm merger. Is anyone else > > starting to do this? If not would anyone object to me starting on it? - > > Bobby > > > > > > On Thursday, November 12, 2015 12:04 PM, P. Taylor Goetz < > [email protected]> wrote: > > > > > > Thanks for putting this together Basti, that comparison helps a lot. > > > > > And thanks Bobby for converting it into markdown. I was going to just > > attach the spreadsheet to JIRA, but markdown is a much better solution. > > > > -Taylor > > > >> On Nov 12, 2015, at 12:03 PM, Bobby Evans <[email protected] > > wrote: > >> > > >> I translated the excel spreadsheet into a markdown file and put up a pull > >> request for it. > >> https://github.com/apache/storm/pull/877 > > >> I did a few edits to it to make it work with Markdown, and to add in a few > >> of my own comments. I also put in a field for JIRAs to be able to track > >> the migration. > > >> Overall I think your evaluation was very good. We have a fair amount of > >> work ahead of us to decide what version of various features we want to go > >> forward with. > >> - Bobby > >> > >> > >> On Thursday, November 12, 2015 9:37 AM, 刘键(Basti Liu) < > [email protected]> wrote: > >> > >> > >> Hi Bobby & Jungtaek, > >> > >> Thanks for your replay. > > >> I totally agree that compatibility is the most important thing. Actually, > >> JStorm has been compatible with the user API of Storm. > > >> As you mentioned below, we indeed still have some features different > >> between Storm and JStorm. I have tried to list them (minor update or > >> improvements are not included). > > >> Please refer to attachment for details. If any missing, please help to > >> point out. (The current working features are probably missing here.) > > >> Just have a look at these differences. For the missing features in JStorm, > >> I did not see any obstacle which will block the merge to JStorm. > > >> For the features which has different solution between Storm and JStorm, we > >> can evaluate the solution one by one to decision which one is appropriate. > > >> After the finalization of evaluation, I think JStorm team can take the > >> merging job and publish a stable release in 2 months. > > >> But anyway, the detailed implementation for these features with different > >> solution is transparent to user. So, from user's point of view, there is > >> not any compatibility problem. > >> > > >> Besides compatibility, by our experience, stability is also important and > >> is not an easy job. 4 people in JStorm team took almost one year to finish > >> the porting from "clojure core" > > >> to "java core", and to make it stable. Of course, we have many devs in > >> community to make the porting job faster. But it still needs a long time > >> to run many online complex topologys to find bugs and fix them. So, that > >> is the reason why I proposed to do merging and build on a stable "java > >> core". > >> > >> -----Original Message----- > >> From: Bobby Evans [mailto:[email protected]] > >> Sent: Wednesday, November 11, 2015 10:51 PM > >> To: [email protected] > >> Subject: Re: [DISCUSS] Plan for Merging JStorm Code > >> > > >> +1 for doing a 1.0 release based off of the clojure 0.11.x code. > >> Migrating the APIs to org.apache.storm is a big non-backwards compatible > >> move, and a major version bump to 2.x seems like a good move there. > >> +1 for the release plan > >> > > >> I would like the move for user facing APIs to org.apache to be one of the > >> last things we do. Translating clojure code into java and moving it to > >> org.apache I am not too concerned about. > >> > >> Basti, > > >> We have two code bases that have diverged significantly from one another > >> in terms of functionality. The storm code now or soon will have A > >> Heartbeat Server, Nimbus HA (Different Implementation), Resource Aware > >> Scheduling, a distributed cache like API, log searching, security, massive > >> performance improvements, shaded almost all of our dependencies, a REST > >> API for programtically accessing everything on the UI, and I am sure I am > >> missing a few other things. JStorm also has many changes including cgroup > >> isolation, restructured zookeeper layout, classpath isolation, and more > >> too. > > >> No matter what we do it will be a large effort to port changes from one > >> code base to another, and from clojure to java. I proposed this initially > >> because it can be broken up into incremental changes. It may take a > >> little longer, but we will always have a working codebase that is testable > >> and compatible with the current storm release, at least until we move the > >> user facing APIs to be under org.apache. This lets the community continue > >> to build and test the master branch and report problems that they find, > >> which is incredibly valuable. I personally don't think it will be much > >> easier, especially if we are intent on always maintaining compatibility > >> with storm. - Bobby > >> > >> > >> On Wednesday, November 11, 2015 5:42 AM, 刘键(Basti Liu) < > [email protected]> wrote: > >> > >> > >> Hi Taylor, > >> > >> > >> > >> Thanks for the merge plan. I have a question about “Phase 2.2”. > >> > > >> Do you mean community plan to create a fresh new “java core” based on > >> current “clojure core” firstly, and then migrate the features from JStorm? > >> > > >> If so, it confused me. It is really a huge job which might require a long > >> developing time to make it stable, while JStorm is already a stable > >> version. > >> > > >> The release planned to be release after Nov 11th has already run online > >> stably several month in Alibaba. > >> > > >> Besides this, there are many valuable internal requirements in Alibaba, > >> the fast evolution of JStorm is forseeable in next few months. > >> > > >> If the “java core” is totally fresh new, it might bring many problems for > >> the coming merge. > >> > > >> So, from the point of this view, I think it is much better and easier to > >> migrate the features of “clojure core” basing on JStorm for the “java > >> core”. > >> > >> Please correct me, if any misunderstanding. > >> > >> > >> > >> Regards > >> > >> Basti > >> > >> > >> > >> 发件人: P. Taylor Goetz [mailto:[email protected]] > >> 发送时间: 2015年11月11日 5:32 > >> 收件人: [email protected] > >> 主题: [DISCUSS] Plan for Merging JStorm Code > >> > >> > >> > > >> Based on a number of discussions regarding merging the JStorm code, I’ve > >> tried to distill the ideas presented and inserted some of my own. The > >> result is below. > >> > >> > >> > > >> I’ve divided the plan into three phases, though they are not necessarily > >> sequential — obviously some tasks can take place in parallel. > >> > >> > >> > > >> None of this is set in stone, just presented for discussion. Any and all > >> comments are welcome. > >> > >> > >> > >> ------- > >> > >> > >> > >> Phase 1 - Plan for 0.11.x Release > >> > >> 1. Determine feature set for 0.11.x and publish to wiki [1]. > >> > >> 2. Announce feature-freeze for 0.11.x > >> > >> 3. Create 0.11.x branch from master (Phase 2.4 can begin.) > >> > >> 4. Release 0.11.0 (or whatever version # we want to use) > >> > >> 5. Bug fixes and subsequent releases from 0.11.x-branch > >> > >> > >> > >> > >> > >> Phase 2 - Prepare for Merge ("master" and "jstorm-import" branches) > >> > > >> 1. Determine/document unique features in JStorm (e.g. classpath isolation, > >> cgroups, etc.) and create JIRA for migrating the feature. > >> > > >> 2. Create JIRA for migrating each clojure component (or logical group of > >> components) to Java. Assumes tests will be ported as well. > >> > > >> 3. Discuss/establish style guide for Java coding conventions. Consider > >> using Oracle’s or Google’s Java conventions as a base — they are both > >> pretty solid. > >> > > >> 4. align package names (e.g backtype.storm --> org.apache.storm / > >> com.alibaba.jstorm --> org.apache.storm) (Phase 3 can begin) > >> > >> > >> > >> > >> > >> Phase 3 - Migrate Clojure --> Java > >> > > >> 1. Port code/tests to Java, leveraging existing JStorm code wherever > >> possible (core functionality only, features distinct to JStorm migrated > >> separately). > >> > >> 2. Port JStorm-specific features. > >> > >> 3. Begin releasing preview/beta versions. > >> > > >> 4. Code cleanup (across the board) and refactoring using established > >> coding conventions, and leveraging PMD/Checkstyle reports for reference. > >> (Note: good oportunity for new contributors.) > >> > > >> 5. Release 0.12.0 (or whatever version # we want to use) and lift feature > >> freeze. > >> > >> > >> > >> > >> > >> Notes: > >> > > >> We should consider bumping up to version 1.0 sometime soon and then > >> switching to semantic versioning [3] from then on. > >> > >> > >> > >> > >> > > >> With the exception of package name alignment, the "jstorm-import" branch > >> will largely be read-only throughout the process. > >> > >> > >> > > >> During migration, it's probably easiest to operate with two local clones > >> of the Apache Storm repo: one for working (i.e. checked out to working > >> branch) and one for reference/copying (i.e. checked out to > >> "jstorm-import"). > >> > >> > >> > > >> Feature-freeze probably only needs to be enforced against core > >> functionality. Components under "external" can likely be exempt, but we > >> should figure out a process for accepting and releasing new features > >> during the migration. > >> > >> > >> > > >> Performance testing should be continuous throughout the process. Since we > >> don't really have ASF infrastructure for performance testing, we will need > >> a volunteer(s) to host and run the performance tests. Performance test > >> results can be posted to the wiki [2]. It would probably be a good idea to > >> establish a baseline with the 0.10.0 release. > >> > >> > >> > > >> I’ve attached an analysis document Sean Zhong put together a while back to > >> the JStorm merge JIRA [4]. The analysis was against the 0.9.3 release but > >> is still relevant and has a lot of good information. > >> > >> > >> > >> > >> > >> [1] > https://cwiki.apache.org/confluence/display/STORM/Release+0.11.0+Feature+Set > >> > >> [2] https://cwiki.apache.org/confluence/display/STORM/Storm+Home > >> > >> [3] http://semver.org > >> > >> [4] https://issues.apache.org/jira/browse/STORM-717 > >> > >> > >> > >> > >> > >> -Taylor > >> > >> > >> > >> > >> > >> > > > > > > > > > > -- Name : 임 정택 Blog : http://www.heartsavior.net / http://dev.heartsavior.net Twitter : http://twitter.com/heartsavior LinkedIn : http://www.linkedin.com/in/heartsavior
