> On Sept. 20, 2017, 4:55 p.m., Bill Farner wrote: > > src/main/java/org/apache/aurora/scheduler/log/mesos/MesosLogStreamModule.java > > Line 148 (original), 150 (patched) > > <https://reviews.apache.org/r/62423/diff/1/?file=1830036#file1830036line150> > > > > This seems to suggest that `zkClientConfig.getServers()` may return the > > wrong addresses. Why is that?
This is because zkClientConfig.getServers() here returns values as specified by -zk_endpoints, which is supposed to be ignored when -zk_in_proc=true. Since ZooKeeperConfig is not injected and we don't know the port of ZooKeeperTestServer when FlaggedZooKeeperConfig.create() is called, it is not trivial to let ZooKeeperConfig#getServer() return a correct value when in-process ZooKeeper server is used. I may make ZooKeeperConfig#getServers() non-public instead. - Keisuke ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/62423/#review185801 ----------------------------------------------------------- On Sept. 19, 2017, 10:49 p.m., Keisuke Nishimoto wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/62423/ > ----------------------------------------------------------- > > (Updated Sept. 19, 2017, 10:49 p.m.) > > > Review request for Aurora. > > > Bugs: AURORA-1947 > https://issues.apache.org/jira/browse/AURORA-1947 > > > Repository: aurora > > > Description > ------- > > MesosLogStreamModule tries to connect to ZooKeeper servers specified by > -zk_endpoints even when -zk_in_proc=true. I updated the module to use > injected server endpoints which will be based on the ephemeral port assigned > to ZooKeeperTestServer if -zk_in_proc=true. This required to make > @ServiceDiscoveryBindings.ZooKeeper public. > > I also tweaked shutdown process of ServiceDiscoveryModule.TestServerService > so that it won't close existing ZooKeeper connections before clients close > the session. While just delaying the execution by 1 second doesn't really > guarantee that behavior, in practice this achieved clean shutdown of the > scheduler with in-process ZooKeeper server. > > > Diffs > ----- > > > src/main/java/org/apache/aurora/scheduler/discovery/ServiceDiscoveryBindings.java > 28cdc4b3b454b3d25008a21c6b12634173e1f878 > > src/main/java/org/apache/aurora/scheduler/discovery/ServiceDiscoveryModule.java > c105dbdbe8339db7f5cca2e5e391fffb4cd87b07 > > src/main/java/org/apache/aurora/scheduler/log/mesos/MesosLogStreamModule.java > 6704a328a4023a178ed8f86ae4772cb04eb2fa8e > > > Diff: https://reviews.apache.org/r/62423/diff/1/ > > > Testing > ------- > > 1. Launch Mesos master and slave on my laptop. > 2. Launch Aurora scheduler with following arguments: > ``` > -backup_dir=/var/lib/aurora/backups > -cluster_name=local > -mesos_master_address=localhost:5050 > -serverset_path=/aurora/scheduler > -ip=127.0.0.1 > -hostname=localhost > -http_port=8081 > -zk_in_proc=true > -zk_endpoints=localhost:2181 > -native_log_zk_group_path=/aurora/replicated-log > -native_log_file_path=/var/db/aurora > ``` > 3. Observe that there are no ZooKeeper error log outputs caused by missing > endpoint. > 4. Create a simple job, observer it launches normally and then kill it. > 5. Stop the scheduler by sending /quitquitquit. > 6. Observe that scheduler process shuts down normally. > > > Thanks, > > Keisuke Nishimoto > >
