This message was deleted.

You can disable the middle manager, await for all the tasks to complete, and then continue the process. See: https://druid.apache.org/docs/latest/operations/api-reference.html#middlemanager for the API to disable them. See: https://druid.apache.org/docs/latest/operations/api-reference.html#overlord to query for task status Finally more information can be found on the rolling update documentation which explains this process further: https://druid.apache.org/docs/latest/operations/rolling-updates.html#middle-managersindexers

Thx Adam, we saw this option, but will have to provide specific time for patching and our load jobs will run for longer time. Assume this may be the common case for most of the orgs. we may have to have customized scripts to accommodate this only if no other good options.

With batch job, I usually just let the job failed and rerun later via Airflow. With real time job, can’t you pause the supervisor, let the tasks finished, and then kill the MM?

also, if you use replicas on the ingestion job, you can tolerate one MM failed without issues on the query side

Didip with hundred's of nodes with multiple clusters, wanted to see this process can be done with out manual interaction, i believe with above options we manually need to run/pause. Renato any idea how to use replica on ingestion job? Haven't got a choice to test but what i am thinking is on unix have kill script on rc3.d to hold the reboot until the existing process completes on MM, so the patching team issues reboot but the node waits until the completion of th eprocess.

if the ingestion is Kafka, it's just a setting in the ingestion json

we have both kafka and index_parallel

1 thing that helped us tremendously is that we actually segregate Kafka MM’s per datasource. And batch ingestion uses a completely separate pool of MM. This allows us to pick and choose which MM pool to perform maintenance.

Got it, will check these options also.