陈梓立 created FLINK-9779:
--------------------------
Summary: Remove SlotRequest timeout
Key: FLINK-9779
URL: https://issues.apache.org/jira/browse/FLINK-9779
Project: Flink
Issue Type: Improvement
Components: JobManager, ResourceManager, TaskManager
Reporter: 陈梓立
As is involved in FLINK-8643 and FLINK-8653, we use external timeout to
replace internal timeout of slot request. Follow the question: why not entirely
remove this timeout mechanism? In our industrial case, this timeout mechanism
causes more no-needed fail and makes resource allocation inaccurate.
I would propose to get rid of slot request timeout. Instead, we handle TM fail
in RM where properly cancel pending request and if TM cannot offer slot to JM,
we introduce a blacklist mechanism to nudge RM realloc for pending request.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
