Sebastian Toader created AMBARI-17248:
-----------------------------------------
Summary: Reduce the idle time before first command from next stage
is executed on a host
Key: AMBARI-17248
URL: https://issues.apache.org/jira/browse/AMBARI-17248
Project: Ambari
Issue Type: Improvement
Components: ambari-agent, ambari-server
Reporter: Sebastian Toader
Assignee: Sebastian Toader
Fix For: 2.2-next, 2.5.0
Commands to be executed by ambari-agents are being sent down by the server in
the response message to agent heartbeat messages.
The server processes the received heartbeat, it checks if there are next
commands scheduled to be executed by ambari-agent and adds those to the
heartbeat response for the ambari-agent.
The server organises the commands that can be executed in parallel into stages.
Ambari server ensures that only the commands of a single stage is scheduled to
be executed by the agent and starts scheduling the commands of the next stage
only after all commands of current stage has finished successfully.
The processing of command status received with the heartbeat message happens
asynchronously to heartbeat response in HeartBeatProcessor and ActionScheduler
creation thus when the heartbeat response is created the commands for the next
stage are not scheduled yet. This means that the next commands will be sent to
agent only with the next heartbeat.
Agents currently sends a heartbeat to the server on command a completion or at
a timeout = self.netutil.HEARTBEAT_IDDLE_INTERVAL_SEC –
self.netutil.MINIMUM_INTERVAL_BETWEEN_HEARTBEATS interval which is ~10 seconds
if there are no commands to be executed.
This means that when the server receives a heartbeat triggered by the
completion of the last command from the current stage the server will send the
commands for the next stage only 10 seconds later when the next heartbeat is
received. This leads to agents spending considerable amount of time idle when
there are multiple stages to be executed.
Agents should heartbeat at a higher rate while there are still pending stages
to be executed.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
