Consider ProjectFilterTransposeRule, a simple rule that matches a Project on top of a Filter. But how to get that Project and Filter?
Run RelOptRulesTest.testPushProjectPastFilter in the debugger, and put a breakpoint in ProjectFilterTransposeRule.onMatch [1]. Look at the Project “origProj” in the debugger. Its input is a HepRelVertex containing a Filter. But Filter “filter” is, sure enough, a Filter. So, to get the filter, don’t call origProj.getInput(), instead call call.rel(1). The RelOptRuleCall will have ensured that this is the correct type, and removed any wrappers. Julian [1] https://insight.io/github.com/apache/calcite/blob/master/core/src/main/java/org/apache/calcite/rel/rules/ProjectFilterTransposeRule.java?line=78 <https://insight.io/github.com/apache/calcite/blob/master/core/src/main/java/org/apache/calcite/rel/rules/ProjectFilterTransposeRule.java?line=78> > On Jun 29, 2017, at 8:47 AM, Muhammad Gelbana <[email protected]> wrote: > > But how can I handle the unexpected variation of the input class type ? > For example, please consider the following *RelOptRule* constructor > > public IncortaLimitRule() { >> super(operand(*DrillLimitRel*.class, operand(*IncortaJdbcDrel*.class, >> operand(*JdbcRel*.class, any()))), "TestRule"); >> } > > > I'm trying to be precise so I specified the RelNode (*DrillLimitRel*), it's > input (*IncortaJdbcDrel*) and it's input's input (*JdbcRel*, this is an > interface but I understand that the planner should match relnodes if the > rule's operand is a parent of examined operand, so for instance, > *JdbcTableScan* should match). > > Is there a way to guarantee the relnode's input type ? Looking into other > rules, I don't remember seeing type checking for input relnodes. Is it a > good sign that I have to do multiple checks for the input relnode type ? > > *---------------------* > *Muhammad Gelbana* > > On Tue, Jun 27, 2017 at 1:36 AM, Julian Hyde <[email protected]> wrote: > >> RelSubset provides a level of indirection that allows Calcite to optimize >> queries. If the input to a relational operator is an equivalence class, not >> a particular relational expression, then Calcite has the freedom to choose >> the member of the equivalence class that has the cheapest cost. >> >> Calcite uses the Volcano algorithm, a form of dynamic programming, to do >> this. >> >> As you have discovered it means that you cannot simply tree-walk over a >> relational expression, looking for, say, a Project on top of a Filter. You >> need to write a RelOptRule that declares that it wants to see a Project on >> top of a Filter, and Volcano will fire that rule whenever that combination >> arises. >> >> Julian >> >> >>> On Jun 26, 2017, at 4:00 PM, Muhammad Gelbana <[email protected]> >> wrote: >>> >>> While debugging through some of the rules I'm trying to get to work, I >> find >>> that the matched relnode's input is *HepRelVertex* or a *RelSubset*. >>> >>> Why wouldn't I be getting a descendant of the following types: >>> >>> - BiRel >>> - MultiJoin >>> - SetOp >>> - SingleRel >>> - TableScan >>> >>> There are other types too but I can assume how to handle the types I >> listed. >>> >>> *RelSubset* is really confusing because, as far as I understand, it's a >>> group of relnodes. So how can I decide which node to get from this group >> so >>> I can use it as an input to the new relnode I'm creating ?! Or should I >>> pass the *RelSubset* itself ? Why am I getting it in the first place and >>> not a single relnode ? >>> >>> *---------------------* >>> *Muhammad Gelbana* >> >>
