Github user IgorBerman commented on the issue:
https://github.com/apache/spark/pull/20640
@squito I'm with debug logging and see declines. Overall it's working.
However there are different possible reasons for declines not necessary due to
blacklisted node(Actually there is jira that talks about enriching this
information about declines)
Regarding "much effect" - I mean that killing executor manually is not
necessary causing BlacklistTracker to add it to the blacklist(driver usually
accepts offer from other slave), so chances that I'll cause enough failures on
same executor are pretty low. Maybe I'll reduce all blacklist.max* to be 1,
thus aggressively blacklisting only 1 failure. WDYT?
Regarding "other bug", which is probably not a bug...My understanding was
that in coarse grained mode only 1 executor possible per application per node,
but might be I'm wrong.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
