Change Details

In order to have confidence in our migration plan for https://phabricator.wikimedia.org/T255973, we will create a cluster of a few Kafka nodes and use Kafka MirrorMaker to replicate the kafka-jumbo cluster data to the new cluster. We can then add nodes to the mirror cluster and test rebalancing the partitions to include the new nodes, and test that this process is smooth. For simplicity, we'll want to mirror a subset of topics, making sure to include one of the highest-traffic topics, like `webrequest_text`. One candidate set of nodes is analytics1051-analytics1057, which are former hadoop workers and are not currently in use. These apparently only have 1GB/s network, whereas the kafka-jumbo nodes have 10GB/s, so if the migration works on these less-networked nodes, it should be just fine on the production cluster.

In order to have confidence in our migration plan for https://phabricator.wikimedia.org/T255973, we will create a cluster of a few Kafka nodes and use Kafka MirrorMaker to replicate the kafka-jumbo cluster data to the new cluster. We can then add nodes to the mirror cluster and test rebalancing the partitions to include the new nodes, and test that this process is smooth. For simplicity, we'll want to mirror a subset of topics, making sure to include one of the highest-traffic topics, like `webrequest_text`. One candidate set of nodes is analytics1051-analytics1056, which are former hadoop workers and are not currently in use. These apparently only have 1GB/s network, whereas the kafka-jumbo nodes have 10GB/s, so if the migration works on these less-networked nodes, it should be just fine on the production cluster.

In order to have confidence in our migration plan for https://phabricator.wikimedia.org/T255973, we will create a cluster of a few Kafka nodes and use Kafka MirrorMaker to replicate the kafka-jumbo cluster data to the new cluster. We can then add nodes to the mirror cluster and test rebalancing the partitions to include the new nodes, and test that this process is smooth. For simplicity, we'll want to mirror a subset of topics, making sure to include one of the highest-traffic topics, like `webrequest_text`. One candidate set of nodes is analytics1051-analytics10576, which are former hadoop workers and are not currently in use. These apparently only have 1GB/s network, whereas the kafka-jumbo nodes have 10GB/s, so if the migration works on these less-networked nodes, it should be just fine on the production cluster.