[jira] [Created] (CALCITE-1843) Convert a RelOptCost to a totally-ordered value

["JD Zheng (JIRA)" <ji...@apache.org>]

JD Zheng created CALCITE-1843:
---------------------------------

             Summary: Convert a RelOptCost to a totally-ordered value
                 Key: CALCITE-1843
                 URL: https://issues.apache.org/jira/browse/CALCITE-1843
             Project: Calcite
          Issue Type: Improvement
          Components: core
            Reporter: JD Zheng
            Assignee: Julian Hyde


Currently, we have dead code  in VolcanoCost.java

{code:java}
  public boolean isLe(RelOptCost other) {
    VolcanoCost that = (VolcanoCost) other;
    if (true) {
      return this == that
          || this.rowCount <= that.rowCount;
    }
    return (this == that)
        || ((this.rowCount <= that.rowCount)
        && (this.cpu <= that.cpu)
        && (this.io <= that.io));
  }

  public boolean isLt(RelOptCost other) {
    if (true) {
      VolcanoCost that = (VolcanoCost) other;
      return this.rowCount < that.rowCount;
    }
    return isLe(other) && !equals(other);
  }
{code}

The reason for the “if (false)” is that we found that costs don’t work well if they form only a partial order, not a total order. If you have two RelNodes R1 and R2 in an equivalent set, and they have costs C1 and C2, and neither C1 <= C2 nor C2 <= C1 is true, which is the Volcano planner to pick? It will tend to pick the one that it saw first, and that is bad news because it is arbitrary and non-deterministic.

So, we should probably find a way to convert a RelOptCost to a totally-ordered value, such as by applying weights to cpu, io and memory cost and returning a double. 



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)