Chris Thomson created FLINK-8445:
------------------------------------
Summary: hostname used in metric names for taskmanager and
jobmanager are not consistent
Key: FLINK-8445
URL: https://issues.apache.org/jira/browse/FLINK-8445
Project: Flink
Issue Type: Bug
Components: Metrics
Affects Versions: 1.3.1
Environment: I think that this problem is present for metrics
reporting enabled configurations that include '<host>' as part of the scope for
the metrics. For example, using Graphite reporting configuration in
flink-conf.yaml below:
{code:java}
metrics.scope.jm: flink.<host>.jobmanager
metrics.scope.jm.job: flink.<host>.jobmanager.<job_name>
metrics.scope.tm: flink.<host>.taskmanager
metrics.scope.tm.job: flink.<host>.taskmanager.<job_name>
metrics.scope.task:
flink.<host>.taskmanager.<job_name>.<task_name>.<subtask_index>
metrics.scope.operator:
flink.<host>.taskmanager.<job_name>.<operator_name>.<subtask_index>
metrics.reporters: graphite
metrics.reporter.graphite.class:
org.apache.flink.metrics.graphite.GraphiteReporter
...{code}
Reporter: Chris Thomson
Enabled Flink metrics reporting using Graphite using system scopes that contain
'<host>' for both the job manager and task manager. The resulting metrics
reported to Graphite use two different representations for '<host>'.
For *Task Manager metrics* it uses the *short hostname* (without the DNS
domain). This is a result of logic in
org.apache.flink.runtime.taskmanager.TaskManagerLocation constructor that tries
to extract the short hostname from the fully qualified domain name looked up
from InetAddress.getCanonicalHostName().
For *Job Manager metrics* it uses the *fully qualified domain name* (with the
DNS domain). This is a result of there being no logic in
org.apache.flink.runtime.jobmanager.JobManagerRunner or
org.apache.flink.runtime.rpc.akka.AkkaRpcService to perform equivalent
normalization of the fully qualified domain name down to the short hostname.
Ideally the '<host>' placeholders in the system scopes for the job manager and
task manager related metrics would be replaced with a consistent value (either
the short hostname or the fully qualified domain name). Even better if there
was a configuration option to decide which one should be used for metric name
generation.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
