Table cost through workload-based learning

[elsa0520]

Summary

Table cost learning is the first workload based learning feature we build. It measures each table's contribution to workload cost, enabling tools to assess table significance relative to user workloads. For example, "auto analyze" can use this cost to rank tables in its priority queue. The write path is based on a async kernel process that analyze the workload (statement summary) and record the "cost" of each table used in the workload in a ew metadata table. The read path involves caching and updating the new metadata table and consuming the information by the priority queue system.

Catagory

TableCost is one kind of feedback feature in the whole workload-based learning picture

TableCost Overview

Producer part —— Tablecost worker

1. Load workload data from storage "statement summary table"
2. Analyze and compute the tablecost value
3. Write the tablecost into storage "workload values table"
   Consumer part —— Feedback cache worker
4. Cache the feedback value —— table cost from storage into memory
5. Stats priority queue uses the table cost directly
   Storage —— New Workload table
6. Used to save all of the workload values(table cost now) in this table

Only data dependency between the producer part and the consumer part.