Log non-zero expvars periodically #1955
Conversation
|
Not ready for merging, but discussion/reviews welcome. |
|
Example log line: |
| @@ -21,8 +24,18 @@ type Logging struct { | |||
| ToSyslog *bool `config:"to_syslog"` | |||
| ToFiles *bool `config:"to_files"` | |||
| Level string | |||
| Metrics LoggingMetricsConfig `config:"metrics"` | |||
There was a problem hiding this comment.
+1. Should we enable this by default?
There was a problem hiding this comment.
It is enabled by default with a 30s interval.
There was a problem hiding this comment.
Now I see it on line 170. Would probably make sense to also use a ucfg defaultConfig.
There was a problem hiding this comment.
@tsg For the above: I didn't mean as part of this PR.
There was a problem hiding this comment.
I tried initially but didn't work out due to circular imports (cfgfile imports logp).
There was a problem hiding this comment.
:-( @urso interesting issue.
|
LGTM |
tsg
force-pushed
the
log_expvars_periodically
branch
from
6d38f6a to
5746349
Compare
|
Added docs and a basic system test. @ruflin, can you have a look before I squash? |
|
LGTM |
Prints the delta since the last period, if it's non-zero. The idea is to help with troubleshooting while still creating relatively little output in the logs. 30s seems like a good compromise. Part of/related to elastic#1931
tsg
force-pushed
the
log_expvars_periodically
branch
from
2b881d2 to
b378ed8
Compare
tsg
removed
discuss
| func snapshotExpvars(varsMap map[string]int64) { | ||
| expvar.Do(func(kv expvar.KeyValue) { | ||
| switch kv.Value.(type) { | ||
| case *expvar.Int: |
There was a problem hiding this comment.
I think this PR covers the most important metrics for the time being so no action is required at the moment IMO. We can improve this when we add "metrics" within beats. We have some expvar metrics that are nested within a map that won't be printed by this. For example, in Metricbeat we have a fetches.mysql-info.failures metric that is nested two levels deep in maps.
Prints the delta since the last period, if it's non-zero. The idea is
to help with troubleshooting while still creating relatively little output
in the logs. 30s seems like a good compromise.
Part of/related to #1931