[KUDU-3193] Per-tablet histogram for scan predicate efficiency

• Type: Task
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: metrics, ops-tooling, perf, tablet
• Assignee: Unassigned
• Reporter: Andrew Wong
• Created: 11/Sep/20 21:15
• Updated: 12/Sep/20 01:53

Description

Often times slow queries can be the result of a sub-optimal schema for a given workload, e.g. if a scan's predicate is not on a prefix of the primary key. Diagnosing such issues typically takes some understanding of the workloads that are being run against a given table. It'd be nice if there were something more quantitative to understand whether a table(t)'s schema is to blame for a slow scan.

One thought that comes to mind is maintaining a histogram metric per-tablet of the ratio between the number of rows returned during a given scan and the number of rows iterated through during that scan. A consistently low value of this metric would indicate that predicates applied to the given tablet are doing a lot of IO reading rows that are not in the results set.