[Job Debugging] Basic Implement Of Job Debugging. (#2272)

docs/job_debugging.md

@@ -0,0 +1,57 @@
+## Job Debugging
+
+When user submit the job, set the following property in the jobEnv of jobConfig. If the job's user command fails, the container will be kept for 1 week. And user could debug the container after ssh to it. After debugging, user should manually stop it to recycle the system resources.
+
+- 1 week is a default value. It could be configured by your cluster admin.
+
+
+### Submit a debugging job 
+
+#### ```Submit job through json file```
+
+If you submit through a json file and you want to enable job debugging feature for this job, you should set following configuration in your job's json file.
+```JSON
+{
+  "jobEnvs": {
+    "isDebug": true
+  }
+}
+```
+
+#### ```Submit job through webportal```
+
+If you submit through webportal and you want to enable job debugging feature for this job, you should set following configuration in ```jobEnvs```.
+
+![webportal_submit_job](./pic/webportal-job-debugging.png)
+
+
+### Debugging your job, after job failure
+
+
+If users' job is failed and the command exits with a none-zero code, the job's container will be reserved for job debugging.
+
+In Webportal, the job's status is running. ```TODO: show the debugging job with a specify tag in webportal.```  
+
+![webportal_reserved_job_status](./pic/webportal-job-debugging-status.png)
+
+You will find the log following in your job.
+
+```text
+job has finished with exit code 2
+=============================================================================
+======   The job container failed, so it will be reserved for 1 week   ======
+======          After debugging, please stop the job manually.         ======
+=============================================================================
+``` 
+
+
+### Stop your job manually after debugging
+
+You should manually stop the reserved job. Or the occupied resource won't be free. 
+
+![webportal_reserved_job_stop](./pic/webportal-job-debugging-stop.png)
+
+
+### Job failed due to system error
+
+The job, which failed due to system error such as too high usage of file system, can't be reserved by OpenPAI. 

docs/job_tutorial.md

@@ -124,6 +124,7 @@ Below please find the detailed explanation for each of the parameters in the con
 | `retryCount`                     | Integer, optional          | Job retry count, no less than 0          |
 | `jobEnvs`                        | Object, optional           | Job env parameters, key-value pairs, available in job container and **no substitution allowed** |
 | `jobEnvs.paiAzRDMA`                        | Boolean, optional           | If you cluster is azure rdma capable, you could specify the parameter to make your container azure rdma capable. How to use azure rdma? Please follow this [job example](../examples/azure-rdma-inte-mpi-benchmark-with-horovod-image) |
+| `jobEnvs.isDebug`                        | Boolean, optional           | after this flag is set as ```true```, if user's command exits with a none-zero value, the failed container will be reserved for job debugging.  [More detail](./job_debugging.md)|
 
 For more details on explanation, please refer to [frameworklauncher usermanual](../subprojects/frameworklauncher/yarn/doc/USERMANUAL.md).
 

examples/cluster-configuration/services-configuration.yaml

@@ -82,10 +82,12 @@ rest-server:
   default-pai-admin-username: your_default_pai_admin_username
   # database admin password
   default-pai-admin-password: your_default_pai_admin_password
-   # rest server would achieve marketplace template from below configed github repository
+  # rest server would achieve marketplace template from below configed github repository
   #github-owner: Microsoft
   #github-repository: pai
   #github-path: marketplace
+  # Job Debugging Reservation Seconds.
+  #debugging-reservation-seconds: 604800
 
 # uncomment following section if you want to customize the port of web portal
 # webportal:

paictl.py

@@ -99,3 +99,4 @@ def main(args):
 
     setup_logging()
     main(sys.argv[1:])
+

src/rest-server/config/rest-server.md

@@ -19,6 +19,7 @@ other config fields are optional, includes:
 - `github-owner: Microsoft` The marketplace repo owner in GitHub
 - `github-repository: pai` The marketplace repo name
 - `github-path: marketplace` The marketpalce path in the repo
+- `debugging-reservation-seconds: 604800` The seconds to reserved a job container to debug.
 
 ## Generated Configuration <a name="G_Config"></a>
 
@@ -34,6 +35,7 @@ rest-server:
     github-owner: Microsoft
     github-repository: pai
     github-path: marketplace
+    debugging-reservation-seconds: 604800
 ```
 
 ## Table <a name="T_Config"></a>
@@ -99,4 +101,10 @@ rest-server:
     <td>cluster_cfg["rest-server"]["etcd-uris"]</td>
     <td>String</td>
 </tr>
+<tr>
+    <td>rest-server.debugging-reservation-seconds</td>
+    <td>com["rest-server"]["debugging-reservation-seconds"]</td>
+    <td>cluster_cfg["rest-server"]["debugging-reservation-seconds"]</td>
+    <td>String</td>
+</tr>
 </table>

src/rest-server/config/rest-server.yaml

@@ -20,3 +20,4 @@ jwt-secret: pai-secret
 github-owner: Microsoft
 github-repository: pai
 github-path: marketplace
+debugging-reservation-seconds: 604800

src/rest-server/config/rest_server.py

@@ -29,6 +29,12 @@ def validation_pre(self):
             return False, '"default-pai-admin-username" is required in rest-server'
         if 'default-pai-admin-password' not in self.service_configuration:
             return False, '"default-pai-admin-password" is required in rest-server'
+        try:
+            reservation_time = int(self.service_configuration['debugging-reservation-seconds'])
+        except ValueError:
+            return False, '"debugging-reservation-seconds" should be a positive integer.'
+        if reservation_time <= 0:
+            return False, '"debugging-reservation-seconds" should be a positive integer.'
 
         return True, None
 
@@ -50,6 +56,7 @@ def run(self):
         service_object_model['github-owner'] = self.service_configuration['github-owner']
         service_object_model['github-repository'] = self.service_configuration['github-repository']
         service_object_model['github-path'] = self.service_configuration['github-path']
+        service_object_model['debugging-reservation-seconds'] = self.service_configuration['debugging-reservation-seconds']
         service_object_model['etcd-uris'] = ','.join('http://{0}:4001'.format(host['hostip'])
                                                      for host in machine_list
                                                      if host.get('k8s-role') == 'master')

src/rest-server/deploy/rest-server.yaml.template

@@ -57,6 +57,8 @@ spec:
           value: {{ cluster_cfg['layout']['kubernetes']['api-servers-url'] }}
         - name: AZ_RDMA
           value: "{{ cluster_cfg['cluster']['common']['az-rdma']}}"
+        - name: DEBUGGING_RESERVATION_SECONDS
+          value: "{{ cluster_cfg['rest-server']['debugging-reservation-seconds']}}"
 {% if cluster_cfg['rest-server']['github-owner'] %}
         - name: GITHUB_OWNER
           value: {{ cluster_cfg['rest-server']['github-owner'] }}

src/rest-server/src/config/paiConfig.js

@@ -26,18 +26,20 @@ try {
     paiMachineList = yaml.safeLoad(fs.readFileSync('/pai-cluster-config/layout.yaml', 'utf8'))['machine-list'];
 } catch (err) {
     paiMachineList = [];
-    logger.info('Unable to load machine list from cluster-configuration.');
-    logger.info(err.stack);
+    logger.warn('Unable to load machine list from cluster-configuration.');
+    logger.warn('The machine list will be initialized as an empty list.');
 }
 
 let paiConfigData = {
     machineList: paiMachineList,
+    debuggingReservationSeconds: Number(process.env.DEBUGGING_RESERVATION_SECONDS || '604800'),
 };
 
 
 // define the schema for pai configuration
 const paiConfigSchema = Joi.object().keys({
     machineList: Joi.array(),
+    debuggingReservationSeconds: Joi.number().integer().positive(),
 }).required();
 
 

src/rest-server/src/models/job.js

@@ -445,6 +445,8 @@ class Job {
           'azRDMA': azureEnv.azRDMA === 'false' ? false : true,
           'paiMachineList': paiConfig.machineList,
           'reqAzRDMA': data.jobEnvs && data.jobEnvs.paiAzRDMA === true ? true : false,
+          'isDebug': data.jobEnvs && data.jobEnvs.isDebug === true ? true : false,
+          'debuggingReservationSeconds': paiConfig.debuggingReservationSeconds,
         });
     return dockerContainerScript;
   }

src/rest-server/src/templates/dockerContainerScript.mustache

@@ -202,6 +202,24 @@ else
   wait $user_command_pid
   user_command_exitcode=$?
   echo "job has finished with exit code $user_command_exitcode"
+{{# isDebug }}
+  if [[ $user_command_exitcode -ne 0 ]]; then
+    echo "============================================================================="
+    echo "======   The job container failed, so it will be reserved for 1 week   ======"
+    echo "======          After debugging, please stop the job manually.         ======"
+    echo "============================================================================="
+
+    sleep_time={{ debuggingReservationSeconds }}
+    sleep_count=0
+
+    while [ $(( $(date +%s) - $(stat -c %Y /alive/yarn_$PAI_CONTAINER_ID) )) -lt 30 ] && \
+        [ "$sleep_count" -lt "$sleep_time" ]; do
+        sleep 20
+        sleep_count=$((sleep_count+20))
+    done
+
+  fi
+{{/ isDebug }}
   exit $user_command_exitcode
 fi
 

src/rest-server/test/setup.js

@@ -28,6 +28,7 @@ process.env.DEFAULT_PAI_ADMIN_PASSWORD = 'adminis';
 process.env.YARN_URI = 'http://yarn.test.pai:8088';
 process.env.K8S_APISERVER_URI = 'http://kubernetes.test.pai:8080';
 process.env.AZ_RDMA = 'false';
+process.env.DEBUGGING_RESERVATION_SECONDS = '604800';
 
 
 // module dependencies