Alexander Behm created IMPALA-5977:
--------------------------------------
Summary: Only compute value transfers for materialized slots
Key: IMPALA-5977
URL: https://issues.apache.org/jira/browse/IMPALA-5977
Project: IMPALA
Issue Type: Sub-task
Components: Frontend
Affects Versions: Impala 2.9.0, Impala 2.8.0, Impala 2.7.0, Impala 2.6.0,
Impala 2.5.0, Impala 2.10.0
Reporter: Alexander Behm
Today, we include all slots when computing the transitive closure of the value
transfer graph, but we should really exclude non-materialized slots to speed up
the computation.
To fix this issue we need to do at least these things:
* Mark slots as materialized before computing the value transfer graph (today
these phases are in reverse order)
* When marking slots as materialized, include all slots that are relevant to
the value-transfer computation including the virtual sots of inline views
(today slot materialization only marks base table slots). This means reworking
the existing QueryStmt.materializeRequiredSlots() to not use the baseTblSmap_
and instead rely on the SlotDescriptor.sourceExprs_ to track all source slots
that contribute to the result of an expr.
To get started look at these functions and where they are called:
* QueryStmt.materializeRequiredSlots()
* Analyzer.computeEquivClasses()
* SingleNodePlanner.createSingleNodePlan()
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
