[
https://issues.apache.org/jira/browse/PHOENIX-3471?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15687945#comment-15687945
]
Maryann Xue commented on PHOENIX-3471:
--------------------------------------
[~julianhyde], you've made a good point. Right now the Calcite-Phoenix plans
are prone to insignificant changes in textual output. Rather trivial operators
like PhoenixXXXProject can come and go due to even small implementation
adjustment in Calcite rules or Sql2RelConverter or something else. And the
field names in a Project are sometimes not deterministic. We don't care about
these operators though, as long as the more important operators like
PhoenixXXXAggregate or PhoenixXXXSort remain essentially the same as expected
(some attributes may differ depending on where the Project nodes are). So in
this case the textual approach and the programmatic approach would both work
well. But meanwhile, for plans with more complicated structures like joins,
subqueries, or multiple aggregations, the programmatic approach would be
better, for reasons like: 1) join ordering matters, 2) multiple Aggregate or
Sort nodes can be either client-side or server-side operators and we need to
check them respectively, etc.
> Allow accessing full (legacy) Phoenix EXPLAIN information via Calcite
> ---------------------------------------------------------------------
>
> Key: PHOENIX-3471
> URL: https://issues.apache.org/jira/browse/PHOENIX-3471
> Project: Phoenix
> Issue Type: Sub-task
> Reporter: Gabriel Reid
> Assignee: Gabriel Reid
>
> The EXPLAIN syntax in Calcite-Phoenix (either "EXPLAIN <sql>" or "EXPLAIN
> PLAN FOR <sql>") currently returns the Calcite plan for a query. For example:
> {code}
> EXPLAIN SELECT MAX(I) FROM T1
> {code}
> results in the following Calcite explain plan:
> {code}
> PhoenixToEnumerableConverter
> PhoenixServerAggregate(group=[{}], EXPR$0=[MAX($0)])
> PhoenixTableScan(table=[[phoenix, T1]])
> {code}
> and the following (legacy) Phoenix explain plan:
> {code}
> CLIENT PARALLEL 1-WAY FULL SCAN OVER T1
> SERVER FILTER BY FIRST KEY ONLY
> {code}
> There are currently a large number of integration tests which depend on the
> legacy Phoenix format of explain plan, and this format is no longer available
> when running via Calcite. PHOENIX-3105 added support for accessing the
> explain plan via the "EXPLAIN <sql>" syntax, but this update to the syntax
> still only provides the Calcite-specific explain plan.
> There are three main approaches which can be taken here:
> h4. Option 1: Custom EXPLAIN execution
> This approach extends the work done in PHOENIX-3105 to plug in a custom
> SqlPhoenixExplain
> node which returns the legacy Phoenix explain plan, with the "EXPLAIN PLAN
> FOR <sql>"
> syntax still returning the Calcite explain plan.
> h4. Option 2: Add the legacy Phoenix explain plan to the Calcite plan as a
> top-level attribute
> This approach results in an explain plan that looks as follows:
> {code}
> PhoenixToEnumerableConverter(PhoenixExecutionPlan=[CLIENT PARALLEL 1-WAY FULL
> SCAN OVER T1
> SERVER FILTER BY FIRST KEY ONLY])
> PhoenixServerAggregate(group=[{}], EXPR$0=[MAX($0)])
> PhoenixTableScan(table=[[phoenix, T1]])
> {code}
> The disadvantage of this approach is that it's not really "correct" -- we're
> just tacking
> a different representation of the explain plan into the Calcite explain plan.
> The advantage of this approach is that it's very quick and easy to implement
> (i.e. it
> can be done immediately), and it will require minimal changes to the many
> test cases which have
> hard-coded explain plans that things are checked against. All we need to do
> is have a
> utility to extract the PhoenixExecutionPlan value from the full Calcite plan,
> and other
> than that all test cases stay the same.
> h4. Option 3: Add all relevant information to the correct parts of the
> Calcite explain plan
> This approach would result in an explain plan that looks as follows:
> {code}
> PhoenixToEnumerableConverter
> PhoenixServerAggregate(group=[{}], EXPR$0=[MAX($0)])
> PhoenixTableScan(table=[[phoenix, T1]], scanType[CLIENT PARALLEL 1-WAY
> FULL ])
> {code}
> This is undoubtedly the "right" way to do things. However, it has the major
> disadvantage
> that it will require a large amount of work to do the following:
> * add all relevant information into various implementations of
> {{AbstractRelNode.explainTerms}}
> * rework all test cases which verify things against an expected explain plan
> It is of course also an option is to start with option 2 here, and eventually
> migrate to option 3.
> If we go for option 2 or option 3, we should probably remove the custom
> EXPLAIN parsing.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
