[KUDU-3147] Balance tablets based on range hash buckets

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.12.0
• Fix Version/s: 1.17.0
• Component/s: master, perf
• Labels: balancer performance roadmap-candidate
• Assignee: Ravi Bhanot
• Reporter: Grant Henke
• Created: 09/Jun/20 03:20
• Updated: 08/Apr/22 21:16
• Resolved: 08/Apr/22 21:16

Description

When a user defines a schema that uses range + hash partitioning its is often the case that the tablets in the latest range, based on time or any semi-sequential data, are the only tablets that receive writes. Or even if not the latest, it is common for a single range to receive a burst of writes if backloading.

This is so common, that the default Kudu balancing scheme should consider placing/rebalancing the tablets for the hash buckets within each range on as many servers as possible in order to support the maximum write throughput. In that case, `min(#buckets, #total-cluster-tservers)` tservers will be used to handle the writes if the cluster is perfectly balanced. Today, even if perfectly balanced, it is possible for all the hash buckets to be on a single tserver.

Issue Links

• is related to: KUDU-2823 Place tablet replicas based on dimension
• is related to: KUDU-2974 Make the rebalancer honor dimension-based replica placement
• relates to: KUDU-3061 Balance tablet leaders across TServers