Qifan Chen created IMPALA-11163:
-----------------------------------
Summary: To scan small dimensional tables, the number of nodes
selected by FE can be less
Key: IMPALA-11163
URL: https://issues.apache.org/jira/browse/IMPALA-11163
Project: IMPALA
Issue Type: Improvement
Reporter: Qifan Chen
In Impala, FE determines the # of exec nodes to use for scan based on the # of
local/remote nodes hosting data blocks. For example for a dimensional table,
assume its #local nodes = 3, and its #remote nodes = 17. Then # of exec nodes
for scan is 20. The final value is min(20, #exec nodes in cluster).
In the case of a partitioned join(f, d) where f is the fact table and d is the
dimensional table, the # of network opens from join to table d can be made less
(say 2 instead of 20). Therefore, the system can handle more # of queries.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
