Hi Yang, I agree that these two of works would benefit from single assignee. My concern is as below
1. Both share libs & remote flink dist/libs are remote ship files. I don't think we have to implement multiple codepath/configuration. 2. So, for concept clarification, there are (1) an option to disable shipping local libs (2) flink-dist supports multiple schema at least we said "hdfs://" (3) an option for registering remote shipfiles with path & visibility. I think new configuration system helps. the reason we have to special handling (2) instead of including it in (3) is because when shipping flink-dist to TM container, we specially detect flink-dist. Of course we can merge it into general ship files and validate shipfiles finally contain flink-dist, which is an alternative. The *most important* difference is (1) and (3) which we don't have an option for only remote libs. Is this clarification satisfy your proposal? Best, tison. Till Rohrmann <trohrm...@apache.org> 于2020年4月17日周五 下午7:49写道： > Hi Yang, > > from what I understand it sounds reasonable to me. Could you sync with > Tison on FLINK-14964 on how to proceed. I'm not super deep into these > issues but they seem to be somewhat related and Tison already did some > implementation work. > > I'd say it be awesome if we could include this kind of improvement into > the release. > > Cheers, > Till > > On Thu, Apr 16, 2020 at 4:43 AM Yang Wang <danrtsey...@gmail.com> wrote: > >> Hi All, thanks a lot for reviving this discussion. >> >> I think we could unify the FLINK-13938 and FLINK-14964 since they have >> the similar >> purpose, avoid unnecessary uploading and downloading jars in YARN >> deployment. >> The difference is FLINK-13938 aims to support the flink system lib >> directory only, while >> FLINK-14964 is trying to support arbitrary pre-uloaded jars(including >> user and system jars). >> >> >> So i suggest to do this feature as following. >> 1. Upload the flink lib directory or users to hdfs, e.g. >> "hdfs://hdpdev/flink/release/flink-1.x" >> "hdfs://hdpdev/user/someone/mylib" >> 2. Use the -ypl argument to specify the shared lib, multiple directories >> could be specified >> 3. YarnClusterDescriptor will use the pre-uploaded jars to avoid >> unnecessary uploading, >> both for system and user jars >> 4. YarnClusterDescriptor needs to set the system jars to public >> visibility so that the distributed >> cache in the YARN nodemanager could be reused by multiple applications. >> This is to avoid >> unnecessary downloading, especially for the "flink-dist-*.jar". For the >> user shared lib, the >> visibility is still set to "APPLICATION" level. >> >> >> For our past internal use case, the shared lib could help with >> accelerating the submission a lot. >> Also it helps to reduce the pressure of HDFS when we want to launch many >> applications together. >> >> @tison @Till Rohrmann <trohrm...@apache.org> @Hailu, Andreas >> <andreas.ha...@gs.com> If you guys thinks the suggestion makes sense. I >> will try to find some time to work on this and hope it could catch up >> with release-1.1 cycle. >> >> >> Best, >> Yang >> >> Hailu, Andreas [Engineering] <andreas.ha...@gs.com> 于2020年4月16日周四 >> 上午8:47写道： >> >>> Okay, I’ll continue to watch the JIRAs. Thanks for the update, Till. >>> >>> >>> >>> *// *ah >>> >>> >>> >>> *From:* Till Rohrmann <trohrm...@apache.org> >>> *Sent:* Wednesday, April 15, 2020 10:51 AM >>> *To:* Hailu, Andreas [Engineering] <andreas.ha...@ny.email.gs.com> >>> *Cc:* Yang Wang <danrtsey...@gmail.com>; tison <wander4...@gmail.com>; >>> user@flink.apache.org >>> *Subject:* Re: Flink Conf "yarn.flink-dist-jar" Question >>> >>> >>> >>> Hi Andreas, >>> >>> >>> >>> it looks as if FLINK-13938 and FLINK-14964 won't make it into the 1.10.1 >>> release because the community is about to start the release process. Since >>> FLINK-13938 is a new feature it will be shipped with a major release. There >>> is still a bit of time until the 1.11 feature freeze and if Yang Wang has >>> time to finish this PR, then we could ship it. >>> >>> >>> >>> Cheers, >>> >>> Till >>> >>> >>> >>> On Wed, Apr 15, 2020 at 3:23 PM Hailu, Andreas [Engineering] < >>> andreas.ha...@gs.com> wrote: >>> >>> Yang, Tison, >>> >>> >>> >>> Do we know when some solution for 13938 and 14964 will arrive? Do you >>> think it will be in a 1.10.x version? >>> >>> >>> >>> *// *ah >>> >>> >>> >>> *From:* Hailu, Andreas [Engineering] >>> *Sent:* Friday, March 20, 2020 9:19 AM >>> *To:* 'Yang Wang' <danrtsey...@gmail.com> >>> *Cc:* tison <wander4...@gmail.com>; user@flink.apache.org >>> *Subject:* RE: Flink Conf "yarn.flink-dist-jar" Question >>> >>> >>> >>> Hi Yang, >>> >>> >>> >>> This is good to know. As a stopgap measure until a solution between >>> 13938 and 14964 arrives, we can automate the application staging directory >>> cleanup from our client should the process fail. It’s not ideal, but will >>> at least begin to manage our users’ quota. I’ll continue to watch the two >>> tickets. Thank you. >>> >>> >>> >>> *// *ah >>> >>> >>> >>> *From:* Yang Wang <danrtsey...@gmail.com> >>> *Sent:* Monday, March 16, 2020 9:37 PM >>> *To:* Hailu, Andreas [Engineering] <andreas.ha...@ny.email.gs.com> >>> *Cc:* tison <wander4...@gmail.com>; user@flink.apache.org >>> *Subject:* Re: Flink Conf "yarn.flink-dist-jar" Question >>> >>> >>> >>> Hi Hailu, >>> >>> >>> >>> Sorry for the late response. If the Flink cluster(e.g. Yarn application) >>> is stopped directly >>> >>> by `yarn application -kill`, then the staging directory will be left >>> behind. Since the jobmanager >>> >>> do not have any change to clean up the staging directly. Also it may >>> happen when the >>> >>> jobmanager crashed and reached the attempts limit of Yarn. >>> >>> >>> >>> For FLINK-13938, yes, it is trying to use the Yarn public cache to >>> accelerate the container >>> >>> launch. >>> >>> >>> >>> >>> >>> Best, >>> >>> Yang >>> >>> >>> >>> Hailu, Andreas <andreas.ha...@gs.com> 于2020年3月10日周二 上午4:38写道： >>> >>> Also may I ask what causes these application ID directories to be left >>> behind? Is it a job failure, or can they persist even if the application >>> succeeds? I’d like to know so that I can implement my own cleanup in the >>> interim to prevent exceeding user disk space quotas. >>> >>> >>> >>> *// *ah >>> >>> >>> >>> *From:* Hailu, Andreas [Engineering] >>> *Sent:* Monday, March 9, 2020 1:20 PM >>> *To:* 'Yang Wang' <danrtsey...@gmail.com> >>> *Cc:* tison <wander4...@gmail.com>; user@flink.apache.org >>> *Subject:* RE: Flink Conf "yarn.flink-dist-jar" Question >>> >>> >>> >>> Hi Yang, >>> >>> >>> >>> Yes, a combination of these two would be very helpful for us. We have a >>> single shaded binary which we use to run all of the jobs on our YARN >>> cluster. If we could designate a single location in HDFS for that as well, >>> we could also greatly benefit from FLINK-13938. >>> >>> >>> >>> It sounds like a general public cache solution is what’s being called >>> for? >>> >>> >>> >>> *// *ah >>> >>> >>> >>> *From:* Yang Wang <danrtsey...@gmail.com> >>> *Sent:* Sunday, March 8, 2020 10:52 PM >>> *To:* Hailu, Andreas [Engineering] <andreas.ha...@ny.email.gs.com> >>> *Cc:* tison <wander4...@gmail.com>; user@flink.apache.org >>> *Subject:* Re: Flink Conf "yarn.flink-dist-jar" Question >>> >>> >>> >>> Hi Hailu, tison, >>> >>> >>> >>> I created a very similar ticket before to accelerate Flink submission on >>> Yarn[1]. However, >>> >>> we do not get a consensus in the PR. Maybe it's time to revive the >>> discussion and try >>> >>> to find a common solution for both the two tickets[1][2]. >>> >>> >>> >>> >>> >>> [1]. https://issues.apache.org/jira/browse/FLINK-13938 >>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__issues.apache.org_jira_browse_FLINK-2D13938&d=DwMFaQ&c=7563p3e2zaQw0AB1wrFVgyagb2IE5rTZOYPxLxfZlX4&r=hRr4SA7BtUvKoMBP6VDhfisy2OJ1ZAzai-pcCC6TFXM&m=rlD0F8Cr4H0aPlN6O2_K13Q76RFOERSWuJANh4q6X_8&s=njA3vGYTf0g7Zsog8AiwS4bbXxblOxepBEWUV9W3E0s&e=> >>> >>> [2]. https://issues.apache.org/jira/browse/FLINK-14964 >>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__issues.apache.org_jira_browse_FLINK-2D14964&d=DwMFaQ&c=7563p3e2zaQw0AB1wrFVgyagb2IE5rTZOYPxLxfZlX4&r=hRr4SA7BtUvKoMBP6VDhfisy2OJ1ZAzai-pcCC6TFXM&m=rlD0F8Cr4H0aPlN6O2_K13Q76RFOERSWuJANh4q6X_8&s=9kT1RZkGwWh3MAbc_ZUrsEsmRRfw6VK4rlNIeNxs6GU&e=> >>> >>> >>> >>> >>> >>> Best, >>> >>> Yang >>> >>> >>> >>> Hailu, Andreas <andreas.ha...@gs.com> 于2020年3月7日周六 上午11:21写道： >>> >>> Hi Tison, thanks for the reply. I’ve replied to the ticket. I’ll be >>> watching it as well. >>> >>> >>> >>> *// *ah >>> >>> >>> >>> *From:* tison <wander4...@gmail.com> >>> *Sent:* Friday, March 6, 2020 1:40 PM >>> *To:* Hailu, Andreas [Engineering] <andreas.ha...@ny.email.gs.com> >>> *Cc:* user@flink.apache.org >>> *Subject:* Re: Flink Conf "yarn.flink-dist-jar" Question >>> >>> >>> >>> FLINK-13938 seems a bit different than your requirement. The one totally >>> matches is FLINK-14964 >>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__issues.apache.org_jira_browse_FLINK-2D14964&d=DwMFaQ&c=7563p3e2zaQw0AB1wrFVgyagb2IE5rTZOYPxLxfZlX4&r=hRr4SA7BtUvKoMBP6VDhfisy2OJ1ZAzai-pcCC6TFXM&m=9sMjDI0I_9Yni5ZWqV8GScK_KBTaA65yK9kBG-LE5_4&s=X1ZoN456fuc5mNxO6fBzDboEhrI0EHL873LzOd6tnN8&e=>. >>> I'll appreciate it if you can share you opinion on the JIRA ticket. >>> >>> >>> >>> Best, >>> >>> tison. >>> >>> >>> >>> >>> >>> tison <wander4...@gmail.com> 于2020年3月7日周六 上午2:35写道： >>> >>> Yes your requirement is exactly taken into consideration by the >>> community. We currently have an open JIRA ticket for the specific >>> feature[1] and works for loosing the constraint of flink-jar schema to >>> support DFS location should happen. >>> >>> >>> >>> Best, >>> >>> tison. >>> >>> >>> >>> [1] https://issues.apache.org/jira/browse/FLINK-13938 >>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__issues.apache.org_jira_browse_FLINK-2D13938&d=DwMFaQ&c=7563p3e2zaQw0AB1wrFVgyagb2IE5rTZOYPxLxfZlX4&r=hRr4SA7BtUvKoMBP6VDhfisy2OJ1ZAzai-pcCC6TFXM&m=9sMjDI0I_9Yni5ZWqV8GScK_KBTaA65yK9kBG-LE5_4&s=ediMPoQtcPX7K-5fjXJxE2cPp5OySkzwXYfYj8mDWO0&e=> >>> >>> >>> >>> >>> >>> Hailu, Andreas <andreas.ha...@gs.com> 于2020年3月7日周六 上午2:03写道： >>> >>> Hi, >>> >>> >>> >>> We noticed that every time an application runs, it uploads the >>> flink-dist artifact to the /user/<user>/.flink HDFS directory. This causes >>> a user disk space quota issue as we submit thousands of apps to our cluster >>> an hour. We had a similar problem with our Spark applications where it >>> uploaded the Spark Assembly package for every app. Spark provides an >>> argument to use a location in HDFS its for applications to leverage so they >>> don’t need to upload them for every run, and that was our solution (see >>> “spark.yarn.jar” configuration if interested.) >>> >>> >>> >>> Looking at the Resource Orchestration Frameworks page >>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__ci.apache.org_projects_flink_flink-2Ddocs-2Dstable_ops_config.html-23yarn-2Dflink-2Ddist-2Djar&d=DwMFaQ&c=7563p3e2zaQw0AB1wrFVgyagb2IE5rTZOYPxLxfZlX4&r=hRr4SA7BtUvKoMBP6VDhfisy2OJ1ZAzai-pcCC6TFXM&m=9sMjDI0I_9Yni5ZWqV8GScK_KBTaA65yK9kBG-LE5_4&s=3SPuvZu9nPph-qnE3TtbTngG-k3XDBLQGyk9I_tjNtI&e=>, >>> I see there’s might be a similar concept through a “yarn.flink-dist-jar” >>> configuration option. I wanted to place the flink-dist package we’re using >>> in a location in HDFS and configure out jobs to point to it, e.g. >>> >>> >>> >>> yarn.flink-dist-jar: hdfs:////user/delp/.flink/flink-dist_2.11-1.9.1.jar >>> >>> >>> >>> Am I correct in that this is what I’m looking for? I gave this a try >>> with some jobs today, and based on what I’m seeing in the >>> launch_container.sh in our YARN application, it still looks like it’s being >>> uploaded: >>> >>> >>> >>> export >>> _FLINK_JAR_PATH="hdfs://d279536/user/delp/.flink/application_1583031705852_117863/flink-dist_2.11-1.9.1.jar" >>> >>> >>> >>> How can I confirm? Or is this perhaps not config I’m looking for? >>> >>> >>> >>> Best, >>> >>> Andreas >>> >>> >>> ------------------------------ >>> >>> >>> Your Personal Data: We may collect and process information about you >>> that may be subject to data protection laws. For more information about how >>> we use and disclose your personal data, how we protect your information, >>> our legal basis to use your information, your rights and who you can >>> contact, please refer to: www.gs.com/privacy-notices >>> >>> >>> ------------------------------ >>> >>> >>> Your Personal Data: We may collect and process information about you >>> that may be subject to data protection laws. For more information about how >>> we use and disclose your personal data, how we protect your information, >>> our legal basis to use your information, your rights and who you can >>> contact, please refer to: www.gs.com/privacy-notices >>> >>> >>> ------------------------------ >>> >>> >>> Your Personal Data: We may collect and process information about you >>> that may be subject to data protection laws. For more information about how >>> we use and disclose your personal data, how we protect your information, >>> our legal basis to use your information, your rights and who you can >>> contact, please refer to: www.gs.com/privacy-notices >>> >>> >>> ------------------------------ >>> >>> >>> Your Personal Data: We may collect and process information about you >>> that may be subject to data protection laws. For more information about how >>> we use and disclose your personal data, how we protect your information, >>> our legal basis to use your information, your rights and who you can >>> contact, please refer to: www.gs.com/privacy-notices >>> >>> >>> ------------------------------ >>> >>> Your Personal Data: We may collect and process information about you >>> that may be subject to data protection laws. For more information about how >>> we use and disclose your personal data, how we protect your information, >>> our legal basis to use your information, your rights and who you can >>> contact, please refer to: www.gs.com/privacy-notices >>> >>
