For example, please take a look at https://issues.apache.org/jira/browse/CALCITE-4013. I just tested it with tpch q7 which is much slower than non-top-down style optimizter.
top-down : about 1 minute non top-down: < 20 seconds Thanks - Chang Alex Plehanov <plehanov.a...@gmail.com> 于2022年1月11日周二 17:25写道： > Hello, > > Why do you think CALCITE-4052 is a blocker for top-down optimization > in Ignite? As far as I understand, CALCITE-4052 is an umbrella ticket to > implement trait propagation for all physical relational operators of the > "enumerable convention" in Calcite. Ignite doesn't use enumerable > convention and has its own convention with its own physical relational > operators. Can you please explain your concerns? > > сб, 8 янв. 2022 г. в 16:14, Chang Chen <baibaic...@gmail.com>: > > > Hi, ignite guys > > > > I thought that Apache Ignite does what apache drill does[1], but it makes > > me surprised that Ignite uses a top-down calcite optimizer [2]. But I > > believe some issues blocks top-down [3]. > > > > Except for apace drill, both hazelcast[4] and Flink[5] split optimization > > into two independent phases - logical and physical. > > > > Why did Ignite make such a technical decision? > > > > Thanks > > - Chang > > > > [1] > > > > > https://cwiki.apache.org/confluence/display/IGNITE/IEP-37%3A+New+query+execution+engine > > [2] https://issues.apache.org/jira/browse/IGNITE-13117 > > [3] https://issues.apache.org/jira/browse/CALCITE-4052 > > [4] https://github.com/hazelcast/hazelcast > > [5] https://github.com/apache/flink > > >
