airflow-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Albertus Kelvin (Jira)" <j...@apache.org>
Subject [jira] [Created] (AIRFLOW-6214) Spark driver status tracking for standalone, YARN, Mesos and K8s with cluster deploy mode
Date Tue, 10 Dec 2019 09:09:00 GMT
Albertus Kelvin created AIRFLOW-6214:
----------------------------------------

             Summary: Spark driver status tracking for standalone, YARN, Mesos and K8s with
cluster deploy mode
                 Key: AIRFLOW-6214
                 URL: https://issues.apache.org/jira/browse/AIRFLOW-6214
             Project: Apache Airflow
          Issue Type: Improvement
          Components: hooks, operators
    Affects Versions: 1.10.6
            Reporter: Albertus Kelvin


Based on the following code snippet:

{code:python}
def _resolve_should_track_driver_status(self):
        return ('spark://' in self._connection['master'] and
                self._connection['deploy_mode'] == 'cluster')
{code}
 
It seems that the above code will always return *False* because the master address for standalone
cluster doesn't contain *spark://* as shown from the below code snippet.

{code:python}
conn = self.get_connection(self._conn_id)
if conn.port:
    conn_data['master'] = "{}:{}".format(conn.host, conn.port)
else:
    conn_data['master'] = conn.host
{code}

Additionally, I think this driver status tracker should also be enabled for mesos and kubernetes
with cluster mode since the *--status* argument supports all of these cluster managers. Refer
to [this|https://github.com/apache/spark/blob/be867e8a9ee8fc5e4831521770f51793e9265550/core/src/main/scala/org/apache/spark/deploy/SparkSubmitArguments.scala#L543].

For YARN cluster mode, I think we can use built-in commands from yarn itself, such as *yarn
application -status <ApplicationID>*.

Therefore, the below code snippet should be updated similarly to accommodate such a need.

{code:python}
def _build_track_driver_status_command(self):
        # The driver id so we can poll for its status
        if not self._driver_id:
            raise AirflowException(
                    "Invalid status: attempted to poll driver " +
                    "status but no driver id is known. Giving up.")

        if self._connection['master'].startswith("spark://") or 
           self._connection['master'].startswith("mesos://") or 
           self._connection['master'].startswith("k8s://"): 
            # standalone, mesos, kubernetes
            connection_cmd = self._get_spark_binary_path()
            connection_cmd += ["--master", self._connection['master']]
            connection_cmd += ["--status", self._driver_id]
        else:
            # yarn
            connection_cmd = ["yarn application -status"]
            connection_cmd += [self._driver_id]

        self.log.debug("Poll driver status cmd: %s", connection_cmd)

        return connection_cmd
{code}




--
This message was sent by Atlassian Jira
(v8.3.4#803005)

Mime
View raw message