Lei Yang created FLINK-36576:
--------------------------------
Summary: Improving amount-based data balancing distribution
algorithm for DefaultVertexParallelismAndInputInfosDecider
Key: FLINK-36576
URL: https://issues.apache.org/jira/browse/FLINK-36576
Project: Flink
Issue Type: Sub-task
Components: Runtime / Coordination
Reporter: Lei Yang
Currently, the DefaultVertexParallelismAndInputInfosDecider is able to
implement a balanced distribution algorithm based on the amount of data and the
number of subpartitions, however it also has some limitations:
#
Currently, Decider selects the data distribution algorithm via the AllToAll or
Pointwise attribute of the input, which limits the ability of the operator to
dynamically modify the data distribution algorithm.
#
Doesn't support data volume-based balanced distribution for Pointwise inputs.
#
For AllToAll type inputs, it does not support splitting the data corresponding
to the specific key, i.e., it cannot solve the data skewing caused by
single-key hotspot.
For that we plan to introduce the following improvements:
#
Introducing InterInputsKeyCorrelation and IntraInputKeyCorrelation to the input
characterisation which allows the operator to flexibly choose the data balanced
distribution algorithm.
#
Introducing a data volume-based data balanced distribution algorithm for
Pointwise inputs
#
Introducing the ability to split data corresponding to the specific key to
optimise AllToAll's data volume-based data balancing distribution algorithm.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
