Actually we already supported maintenance mode in 0.8.0. My bad.
You can give it a try now, with "MAX_OFFLINE_INSTANCES_ALLOWED" set in ClusterConfig, once the # of offline instance reaches to the limit, Helix will put the cluster into maintenance mode. For now, you have to call HelixAdmin.enableMaintenanceMode() manually to exit the maintenance mode. Support of auto existing maintenance mode is on our road-map. Lei Lei Xia Data Infra/Helix [email protected]<mailto:[email protected]> www.linkedin.com/in/lxia1<http://www.linkedin.com/in/lxia1> ________________________________ From: Bo Liu <[email protected]> Sent: Monday, March 19, 2018 6:33:10 PM To: [email protected] Subject: Re: protect a cluster during broad range outage Hi Lei, Thank you so much for the detailed information. We were almost about to implement our own logic to generate ideal state to handle this disaster case (That's why we were looking at the code in BestPossibleStateCalcStage.java). Are you saying the pausing mode is already implemented in 0.8.0? I looked at the code in validateOfflineInstancesLimit(), which only set the maintenance flag not the pause flag when MAX_OFFLINE_INSTANCES_ALLOWED is hit. Did I miss anything? If pausing mode is supported in 0.8.0, we'd like to try it out. Thanks, Bo On Mon, Mar 19, 2018 at 6:18 PM, Lei Xia <[email protected]<mailto:[email protected]>> wrote: Sorry, I totally missed this email thread. Yes, we do have such feature in 0.8 to protect the cluster in case of disasters happening. A new config option "MAX_OFFLINE_INSTANCES_ALLOWED" can be set in ClusterConfig. If it is set, and the number of offline instances reach to the set limit in the cluster, Helix will automatically pause (disable) the cluster, i.e, Helix will not react to any cluster changes anymore. You have to manually re-enable cluster (via HelixAdmin.enableCluster()) though. Keep in mind, once a cluster is disabled, no cluster event will be handled at all by the controller. For example, if an instance went offline and came back, Helix will not bring up any partitions on that instance if the cluster is disabled. This is somewhat a little coarse-grained. For that reason, we are going to introduce a new cluster mode, called "Maintenance mode". Once a cluster is in maintenance mode, it will not actively move partitions across instances, i.e, it will not bootstrap new partitions to any live instances. However, it will still maintain existing partitions to its desired states as it can. For instance, if an instance comes back, Helix will still bring all existing partitions on this instance to its desired states. Another example is, under maintenance mode, if there are only 1 replica for a given partition left active in the cluster, Helix will not try to bring additional new replicas, but Helix will still transition the only replica to its desired state (for example, master). Once we have this "Maintenance mode" support, we will put the cluster into maintenance mode during disaster, instead of totally disabling it, which leaves more automation here for Helix to recover the cluster from disaster. This feature will be included in our next release (0.8.1), which should be out in a couple of weeks. Lei On Mon, Mar 19, 2018 at 4:07 PM, Bo Liu <[email protected]<mailto:[email protected]>> wrote: Just noticed that we have a cluster config "MAX_OFFLINE_INSTANCES_ALLOWED", which is used in https://github.com/apache/helix/blob/master/helix-core/src/main/java/org/apache/helix/controller/stages/BestPossibleStateCalcStage.java#L70-L71 "If the offline/disabled instance number is above this threshold, the rebalancer will be paused." I am wondering if the FULL_AUTO mode has BestPossibleStateCalcStage? Will it help us with the case when a large portion or even the whole cluster disconnect to zk? On Tue, Mar 6, 2018 at 10:51 PM, Bo Liu <[email protected]<mailto:[email protected]>> wrote: I agree semi-auto is a safer mode for stateful service. But we will have to compute ideal state by ourselves (either manually triggered or triggered by live instance change events). That means we need to implement logic for delayed shard move and a shard placement algorithm. Not sure if there is any building blocks exposed by Helix that we could leverage for semi-auto mode. On Tue, Mar 6, 2018 at 7:12 PM, kishore g <[email protected]<mailto:[email protected]>> wrote: This was one of the reasons we came up with the semi-auto mode. It's non-trivial to handle edge cases in full auto mode, especially for stateful services. Having said that, let's see what we can do in catastrophic scenarios. Having a check on the live instances changes is a good check but its hard to compute this reliably in some scenarios - for e.g. lets controllers also went down at the same time and came up back, they would have missed all the changes from ZK. I think it's better to limit the number of changes a controller would trigger in the cluster. This is where throttling and constraints can be used. Helix already has the ability limit the number of transitions in the cluster at once. But this limits the number of concurrent transitions not the number of transitions triggered in a time period. We can probably enhance this functionality to keep track of the number of transitions in last X minutes and limit that number. Any thoughts on that? On Tue, Mar 6, 2018 at 4:30 PM, Bo Liu <[email protected]<mailto:[email protected]>> wrote: Hi, We are using delayed rebalancer to manage a Master-Slave cluster. In the event when a large portion of a cluster disconnect from ZK (network partition, or service crash due to a bug), helix controller will try hard to move shards to the rest of the cluster. This could make the thing worse if it's very expensive to rebuild a replica or there is no live replica left in the rest of the cluster. I am wondering what's the suggested way to handle this case? Is there a way to let Helix controller pause when the change of live instances is more than a threshold? -- Best regards, Bo -- Best regards, Bo -- Best regards, Bo -- Best regards, Bo
