[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Flink Query Optimizer

Hi Albert,

Calcite provides a rule-based optimizer (as a framework), which means users can customize it by adding rules. That’s exactly what Flink did. From the logical plan to the physical plan, the translations are triggered by different sets of rules, according to which the relational expressions are replaced, reordered or optimized.

However, IMO, the current optimization rules in Flink Table API are quite primal. Some SQL statements (e.g., multiple joins) are just translated to feasible execution plans, instead of optimized ones, since it’s much more difficult to conduct query optimization on large datasets or dynamic streams. You could first start from the Calcite query optimizer, and then try to make your own rules.


> On Jul 14, 2018, at 11:55 AM, vino yang <yanghua1127@xxxxxxxxx> wrote:
> Hi Albert,
> First I guess the query optimizer you mentioned is about Flink table & sql
> (for batch API there is another optimizer which is implemented by Flink).
> Yes, now for table & sql, Flink use Apache Calcite's query optimizer to
> translate into a Calcite plan
> which is then optimized according to Calcite's optimization rules.
> The following rules are applied so far:
> https://github.com/apache/flink/blob/master/flink-libraries/flink-table/src/main/scala/org/apache/flink/table/plan/rules/FlinkRuleSets.scala
> In view of Flink depends on the Calcite to do the optimization, I think
> enhance Flink and Calcite would be the right direction.
> Hope for you provide more idea and details. Flink community welcome your
> idea and contribution.
> Thanks.
> Vino.
> 2018-07-13 23:39 GMT+08:00 Albert Jonathan <albert@xxxxxxxxxx>:
>> Hello,
>> I am just wondering, does Flink use Apache Calcite's query optimizer to
>> generate an optimal logical plan, or does it have its own query optimizer?
>> From what I observed so far, the Flink's query optimizer only groups
>> operator together without changing the order of aggregation operators
>> (e.g., join). Did I miss anything?
>> I am thinking of extending Flink to apply query optimization as in the
>> RDBMS by either integrating it with Calcite or implementing it as a new
>> module.
>> Any feedback or guidelines will be highly appreciated.
>> Thank you,
>> Albert