[
https://issues.apache.org/jira/browse/HBASE-27191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17566082#comment-17566082
]
Andrew Kyle Purtell edited comment on HBASE-27191 at 7/13/22 12:52 AM:
-----------------------------------------------------------------------
Perhaps instead of a system table for every single one of these cases we can do
one named {{hbase:events}} where a bunch of them all go. This concept is
similar to Java's Flight Recorder, which internally uses a ring buffer to hold
events that have a common/generic envelope and are retrieved, parsed, filtered,
and handled case by case by consumer tools.
ResourceLeakDetector warnings should go in for sure.
Events for every significant Exception could go in the new table too. Tools to
list and filter it could come later.
I feel like we should have done this for the slow log too, although it is fine
to have that be a special case, because slow logs are first class things for
some other data storage systems, and especially because it already exists as
such in our case.
was (Author: apurtell):
Perhaps instead of a system table for every single one of these cases we can do
one named {{hbase:events}} where a bunch of them all go. This concept is
similar to Java's Flight Recorder, which internally uses a ring buffer to hold
events that have a common/generic envelope and are retrieved, parsed, filtered,
and handled case by case by consumer tools.
ResourceLeakDetector warnings should go in for sure.
Events for every significant Exception could go in the new table too. Tools to
list and filter it could come later.
I feel like we should have done this for the slow log too, although it is fine
to have that be a special case, especially because it already exists as such.
> ResourceLeakDetector stacktraces can be truncated when logged
> -------------------------------------------------------------
>
> Key: HBASE-27191
> URL: https://issues.apache.org/jira/browse/HBASE-27191
> Project: HBase
> Issue Type: Bug
> Affects Versions: 2.5.0
> Reporter: Andrew Kyle Purtell
> Priority: Minor
>
> Looks like ResourceLeakDetector stacktraces can be truncated by log4j, at
> least with the default configuration. This is unfortunate because what log4j
> will do then is print the first bit of the log line then only the end, which
> omits the crucial top part of the stack trace.
> Should these be given the NamedQueueRecorder treatment like the slow log or
> balancer decisions, i.e. we would also log detected resource leaks and their
> complete stacktrace details into a system table? WDYT [~vjasani] [~zhangduo]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
