Ahmed Hussein created HADOOP-17346:
--------------------------------------
Summary: Fair call queue is defeated by abusive service principals
Key: HADOOP-17346
URL: https://issues.apache.org/jira/browse/HADOOP-17346
Project: Hadoop Common
Issue Type: Bug
Components: common, ipc
Reporter: Ahmed Hussein
Assignee: Ahmed Hussein
[~daryn] reported that the FCQ prioritizes based on the full kerberos
principal (ie. "user/host@realm") rather than short name (ie. "user") to
prevent service principals like the DNs and NMs being de-prioritized since
service principals are expected to be well behaved. Notably the DNs contribute
a significant but important load so the intent is not to de-prioritize all DNs
because their sum total load is high relative to users.
This has the unfortunate side effect of allowing misbehaving & non-critical
service principals to abuse the FCQ. The gstorm/* principals are a prime
example. Each server is spamming opens as fast as possible which ensures that
none of the gstorm servers can be de-prioritized because each principal is a
fraction of the total load from all principals.
The secondary and more devasting problem is other abusive non-service
principals cannot be effectively de-prioritized. The sum total of all gstorm
load prevents other principals from surpassing the priority thresholds.
Principals stay in the highest priority queues which allows the abusive
principals to overflow the entire call queue for extended periods of time.
Notably it prevents the FCQ from moderating the heavy create loads from p_gup @
DB which cause significant performance degradation.
Prioritization should be based on short name with configurable exemptions for
services like the DN/NM.
[~daryn] suggested a solution that we applied on our clusters.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: common-dev-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-dev-h...@hadoop.apache.org
