The term for "maintenance" nodes

[leventov]

The "maintenance" mode of historical nodes introduced in #6349 confused me recently so raise this issue with 0.14.0 milestone while it's not too late to change the name relatively easily.

I like @glasser's suggestion "draining", I think it reflects the essence of the mode more clearly than "maintenance".

[leventov]

FYI @egor-ryashin @jon-wei @clintropolis

[gianm]

It sounds like the feature in #6349 is a way of telling the coordinator to move all data off a particular historical. "Decommission" seems to be the most popular word to use to describe that kind of operation in other systems I checked just now. It also strikes me as pretty intuitive. So I'd suggest that.

• HDFS calls it decommissioning: https://hadoop.apache.org/docs/stable/hadoop-project-dist/hadoop-hdfs/HdfsDataNodeAdminGuide.html
• So does YARN: https://hadoop.apache.org/docs/current/hadoop-yarn/hadoop-yarn-site/GracefulDecommission.html
• So does Cassandra: http://cassandra.apache.org/doc/4.0/operating/topo_changes.html#removing-nodes
• So does Kafka (if only to mention that it doesn't have a simple tool to do it): https://kafka.apache.org/documentation/#basic_ops_decommissioning_brokers
• Mesos doesn't; it uses terms like "maintenance" and "draining" (possibly the inspiration for the terminology in maintenance mode for Historical #6349): http://mesos.apache.org/documentation/latest/maintenance/

[leventov]

@gianm thanks for the analysis and providing the data.

Now I'm in favor of "decommission".

[jon-wei]

"Decommission" sounds good to me.

[clintropolis]

I don't have strong opinions here, "draining" and "decommission" both do sound a bit more intuitive than "maintenance". "draining" seems more popular with orchestration things like mesos, kubernetes, docker swarm, etc, but I think either are appropriate here, so count this as a vote for changing to one or the other.

[clintropolis]

Actually.. 👍 for "decommission" because it seems more appropriate terminology for this application level function, and we are closer to kafka/cassandra/hadoop than mesos/kubernetes/etc so it seems more consistent.

[glasser]

I'd be interested in making/using a layer on top of this feature where you send an API request to the historical itself and it adds itself to the coordinator's decommission list. (See @clintropolis at #6349 (comment).)

I think "decommission" goes well with this potential feature, especially since one could include an "exit when empty" flag on the API. A decommissioning that ends with exiting feels fitting.

[gianm]

It sounds like people generally like "decommission". Since this is a 0.14.0 blocker we'd want to make this change soon (so we can keep progressing towards a release candidate). Is anyone interested in implementing this change?

[clintropolis]

I can pick this up this afternoon if no one else claims it.