[
https://issues.apache.org/jira/browse/HTRACE-303?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15018834#comment-15018834
]
Colin Patrick McCabe commented on HTRACE-303:
---------------------------------------------
There are a few different cases where the client may have to drop spans. One
case is that there is not enough buffer space, or the buffer space can't be
flushed quickly enough. Another case is that the htraced server is not
reachable. Still another case is that the htraced server is overloaded
(because too many spans are being sent in too short of a time).
In some of these cases, it makes sense for the client to notify the server that
spans were dropped; in others, it's not possible (like if the server isn't
reachable). This patch removes the half-implemented server-side
clientDroppedSpans notification mechanism since I think it's too complex to add
to the 4.1 release at this point. Instead, I am adding a simple mechanism that
writes to a local log file (default {{/tmp/htraceDropped}}). This should be a
simple and robust mechanism that should let us know if we have to do some
tuning on our clusters to reduce the rate of htraced span ingestion.
The size of the local log file is limited by
{{htraced.dropped.spans.log.max.size}} so that it will not grow too big. While
it's true that we also log information about dropped spans to log4j, log4j
logging is often turned way down or even off, especially for short-lived
clients like FsShell and so forth. It's much easier to search through a single
file in {{/tmp}} to know that we have to do some tuning.
> Add client-side htraceDropped log file to track dropped spans
> -------------------------------------------------------------
>
> Key: HTRACE-303
> URL: https://issues.apache.org/jira/browse/HTRACE-303
> Project: HTrace
> Issue Type: Improvement
> Components: htraced
> Affects Versions: 4.0
> Reporter: Colin Patrick McCabe
> Assignee: Colin Patrick McCabe
>
> Add a client-side htraceDropped log file to track dropped spans. This is
> more robust than trying to notify the server of dropped spans, since the
> server may not be reachable.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
