You are viewing a plain text version of this content. The canonical link for it is here.
Posted to solr-dev@lucene.apache.org by Yonik Seeley <yo...@apache.org> on 2007/06/19 16:26:31 UTC
requestsPerSecond, averageResponseTime
requestsPerSecond and averageResponseTime were added to statistics for
each response handler. Are these statistics really useful enough to
keep as-is?
averageResponseTime is cumulative since the server started, so it's
not useful for monitoring purposes, but only benchmarking purposes (it
won't tell you if your queries are getting slower all of a sudden).
(it will also count slower warming queries, not just live queries).
requestsPerSecond is likewise flawed... it won't let you detect a
flood of traffic or a dropoff. Also, if you turned off traffic to the
server yesterday, that will continue to be reflected in the
requestsPerSecond today.
Since it seems like these parameters are only useful for benchmarking
(which can easily be done from log files), perhaps we should defer
adding them until we can come up with versions that are useful for
monitoring?
-Yonik
Re: requestsPerSecond, averageResponseTime
Posted by Ryan McKinley <ry...@gmail.com>.
Will Johnson wrote:
>> Has anyone tried to get solr statistics with cacti/nagios? If it isn't
>
>> too difficult, I would like to set this up.
>
>> Can cacti read & parse a file?
>
> Generally speaking nagios/cacti are as powerful as you are with bash.
> We haven't done it yet but it's a requirement at my company to integrate
> the three in the next couple months. Our current plan is to get as many
> stats into the statistics page and then have a shell script grab the xml
> (possibly with an xsl) and then feed that into the monitoring apps.
>>>From there it's ringing pagers and usage graphs galore.
>
sounds good. make sure to check:
http://localhost:8983/solr/admin/plugins?stats=true
the PluginInfoHandler response format can still be tweaked to make
something like this easier
ryan
RE: requestsPerSecond, averageResponseTime
Posted by Will Johnson <wj...@GETCONNECTED.COM>.
>Has anyone tried to get solr statistics with cacti/nagios? If it isn't
>too difficult, I would like to set this up.
>Can cacti read & parse a file?
Generally speaking nagios/cacti are as powerful as you are with bash.
We haven't done it yet but it's a requirement at my company to integrate
the three in the next couple months. Our current plan is to get as many
stats into the statistics page and then have a shell script grab the xml
(possibly with an xsl) and then feed that into the monitoring apps.
>From there it's ringing pagers and usage graphs galore.
When we get something working I'll make sure to post a write up on the
list unless someone else beats me to it.
- will
Re: requestsPerSecond, averageResponseTime
Posted by Ryan McKinley <ry...@gmail.com>.
>>
>> For the record nagios and/or cacti could both keep track of 'in the last
>> x' type of statistics based on totals but having solr compute that
>> automatically would be nice.
>
> That's a good point about totals... totalResponseTime is actually more
> powerful than averageResponseTime in that a monitoring app can poll at
> any frequency they wish to get averageResponseTime over the last
> timeperiod.
>
Has anyone tried to get solr statistics with cacti/nagios? If it isn't
too difficult, I would like to set this up.
Can cacti read & parse a file?
If we return an accumulated totalTime, what happens with the server
restarts and the totalTime is reset to zero? My guess is it gets an
invalid (negative?) number for that time slice, but hopefully continues
normally.
ryan
Re: requestsPerSecond, averageResponseTime
Posted by Yonik Seeley <yo...@apache.org>.
On 6/19/07, Will Johnson <wj...@getconnected.com> wrote:
> Would it be better to have an option to record traffic for the last 'x
> minutes/seconds/hours' configurable on a per handler basis? The goal is
> to have hooks for nagios/cacti/etc to be able to pull live status info
> for monitoring purposes. If you want fine grained performance history
> then log files are the best approach, I just think a way to have beepers
> go off if a server starts getting huge amounts of traffic is a good
> thing.
>
> For the record nagios and/or cacti could both keep track of 'in the last
> x' type of statistics based on totals but having solr compute that
> automatically would be nice.
That's a good point about totals... totalResponseTime is actually more
powerful than averageResponseTime in that a monitoring app can poll at
any frequency they wish to get averageResponseTime over the last
timeperiod.
Calculating other things such as percentiles is more difficult.
Perhaps a plugin interface there is warranted so people can make
different tradeoffs? Regardless, we should try and have good defaults
w/o further configuration.
-Yonik
RE: requestsPerSecond, averageResponseTime
Posted by Will Johnson <wj...@GETCONNECTED.COM>.
Would it be better to have an option to record traffic for the last 'x
minutes/seconds/hours' configurable on a per handler basis? The goal is
to have hooks for nagios/cacti/etc to be able to pull live status info
for monitoring purposes. If you want fine grained performance history
then log files are the best approach, I just think a way to have beepers
go off if a server starts getting huge amounts of traffic is a good
thing.
For the record nagios and/or cacti could both keep track of 'in the last
x' type of statistics based on totals but having solr compute that
automatically would be nice.
- will
-----Original Message-----
From: yseeley@gmail.com [mailto:yseeley@gmail.com] On Behalf Of Yonik
Seeley
Sent: Tuesday, June 19, 2007 10:27 AM
To: solr-dev@lucene.apache.org
Subject: requestsPerSecond, averageResponseTime
requestsPerSecond and averageResponseTime were added to statistics for
each response handler. Are these statistics really useful enough to
keep as-is?
averageResponseTime is cumulative since the server started, so it's
not useful for monitoring purposes, but only benchmarking purposes (it
won't tell you if your queries are getting slower all of a sudden).
(it will also count slower warming queries, not just live queries).
requestsPerSecond is likewise flawed... it won't let you detect a
flood of traffic or a dropoff. Also, if you turned off traffic to the
server yesterday, that will continue to be reflected in the
requestsPerSecond today.
Since it seems like these parameters are only useful for benchmarking
(which can easily be done from log files), perhaps we should defer
adding them until we can come up with versions that are useful for
monitoring?
-Yonik
Re: requestsPerSecond, averageResponseTime
Posted by Yonik Seeley <yo...@apache.org>.
On 6/20/07, Will Johnson <wj...@getconnected.com> wrote:
> The one thing most people (ie product managers) want to see is the
> number of times that users get 0 hits for a query but that doesn't seem
> to be logged anywhere in solr that's easily accessible in log files. Am
> I missing something very obvious or should we try and fix this somehow?
> I know some other engines will log the number of hits in with the query
> log which seems like a nice way of doing things.
Since users normally don't hit solr directly, this could also be
handled at the application-level tier. It might also choose to
implement other logic such as querying an alternate source if no
results were returned from Solr.
-Yonik
RE: requestsPerSecond, averageResponseTime
Posted by Chris Hostetter <ho...@fucit.org>.
: The one thing most people (ie product managers) want to see is the
: number of times that users get 0 hits for a query but that doesn't seem
: to be logged anywhere in solr that's easily accessible in log files. Am
: I missing something very obvious or should we try and fix this somehow?
it's the kind of information the core Solr application can't log because
it doesn't know, but individual request handlers which focus on a "main"
result set (like StandardRequestHandler and DisMaxRequestHandler)
certainly could certinaly log that at an INFO level ... but it might make
sense forthem to use a customer logger name so it can be confiured
seperately from other "code" related log messages, ie...
Logger statsLog = Logger.getLogger(this.getClass().getName() + ".STATS")
-Hoss
RE: requestsPerSecond, averageResponseTime
Posted by Will Johnson <wj...@GETCONNECTED.COM>.
The one thing most people (ie product managers) want to see is the
number of times that users get 0 hits for a query but that doesn't seem
to be logged anywhere in solr that's easily accessible in log files. Am
I missing something very obvious or should we try and fix this somehow?
I know some other engines will log the number of hits in with the query
log which seems like a nice way of doing things.
Any ideas or pointers?
- will
-----Original Message-----
From: Clay Webster [mailto:webster.lists@gmail.com]
Sent: Wednesday, June 20, 2007 10:33 AM
To: solr-dev@lucene.apache.org
Subject: Re: requestsPerSecond, averageResponseTime
Hey Ian, these version with all the parameter options only shows the
table headers.. no data. (No requests?)
PS: I think there's interest. ;-)
--cw
On 6/19/07, Ian Holsman <li...@holsman.net> wrote:
>
> I've been working on a tool to parse log files to get some of this
kind
> of information as well
>
> it's really alpha, but if your curious the dummy system is here:
>
> http://pyro.holsman.net:9081/top/ -- slightly obfuscated queries (to
> roll them up)
> http://pyro.holsman.net:9081/overall/?period=5m&hours=12 -- # of
> requests, response time, and deviation in that
>
>
http://pyro.holsman.net:9081/overall/?period=5m&hours=12&format=csv&cols
=1,2,5,6,7,8
> - same thing as a CSV file and showing selected columns
>
>
> The aim is to use this as a data source for something like cacti and
> sticking a flash graph on top of it.
>
> If there is enough interest I can contribute this to solr
>
> Yonik Seeley wrote:
> > requestsPerSecond and averageResponseTime were added to statistics
for
> > each response handler. Are these statistics really useful enough to
> > keep as-is?
> >
> > averageResponseTime is cumulative since the server started, so it's
> > not useful for monitoring purposes, but only benchmarking purposes
(it
> > won't tell you if your queries are getting slower all of a sudden).
> > (it will also count slower warming queries, not just live queries).
> >
> > requestsPerSecond is likewise flawed... it won't let you detect a
> > flood of traffic or a dropoff. Also, if you turned off traffic to
the
> > server yesterday, that will continue to be reflected in the
> > requestsPerSecond today.
> >
> > Since it seems like these parameters are only useful for
benchmarking
> > (which can easily be done from log files), perhaps we should defer
> > adding them until we can come up with versions that are useful for
> > monitoring?
> >
> > -Yonik
> >
>
>
Re: requestsPerSecond, averageResponseTime
Posted by Ian Holsman <li...@holsman.net>.
Clay Webster wrote:
> Hey Ian, these version with all the parameter options only shows the
> table headers.. no data. (No requests?)
it's a bug with the collector
try http://pyro.holsman.net:9081/top/?hours=72
and I'll work on getting more load pushed through it. (It comes with a
log replay tool as well)
>
> PS: I think there's interest. ;-)
>
> --cw
>
>
> On 6/19/07, Ian Holsman <li...@holsman.net> wrote:
>>
>> I've been working on a tool to parse log files to get some of this kind
>> of information as well
>>
>> it's really alpha, but if your curious the dummy system is here:
>>
>> http://pyro.holsman.net:9081/top/ -- slightly obfuscated queries (to
>> roll them up)
>> http://pyro.holsman.net:9081/overall/?period=5m&hours=12 -- # of
>> requests, response time, and deviation in that
>>
>> http://pyro.holsman.net:9081/overall/?period=5m&hours=12&format=csv&cols=1,2,5,6,7,8
>>
>> - same thing as a CSV file and showing selected columns
>>
>>
>> The aim is to use this as a data source for something like cacti and
>> sticking a flash graph on top of it.
>>
>> If there is enough interest I can contribute this to solr
>>
>> Yonik Seeley wrote:
>> > requestsPerSecond and averageResponseTime were added to statistics for
>> > each response handler. Are these statistics really useful enough to
>> > keep as-is?
>> >
>> > averageResponseTime is cumulative since the server started, so it's
>> > not useful for monitoring purposes, but only benchmarking purposes (it
>> > won't tell you if your queries are getting slower all of a sudden).
>> > (it will also count slower warming queries, not just live queries).
>> >
>> > requestsPerSecond is likewise flawed... it won't let you detect a
>> > flood of traffic or a dropoff. Also, if you turned off traffic to the
>> > server yesterday, that will continue to be reflected in the
>> > requestsPerSecond today.
>> >
>> > Since it seems like these parameters are only useful for benchmarking
>> > (which can easily be done from log files), perhaps we should defer
>> > adding them until we can come up with versions that are useful for
>> > monitoring?
>> >
>> > -Yonik
>> >
>>
>>
>
Re: requestsPerSecond, averageResponseTime
Posted by Clay Webster <we...@gmail.com>.
Hey Ian, these version with all the parameter options only shows the
table headers.. no data. (No requests?)
PS: I think there's interest. ;-)
--cw
On 6/19/07, Ian Holsman <li...@holsman.net> wrote:
>
> I've been working on a tool to parse log files to get some of this kind
> of information as well
>
> it's really alpha, but if your curious the dummy system is here:
>
> http://pyro.holsman.net:9081/top/ -- slightly obfuscated queries (to
> roll them up)
> http://pyro.holsman.net:9081/overall/?period=5m&hours=12 -- # of
> requests, response time, and deviation in that
>
> http://pyro.holsman.net:9081/overall/?period=5m&hours=12&format=csv&cols=1,2,5,6,7,8
> - same thing as a CSV file and showing selected columns
>
>
> The aim is to use this as a data source for something like cacti and
> sticking a flash graph on top of it.
>
> If there is enough interest I can contribute this to solr
>
> Yonik Seeley wrote:
> > requestsPerSecond and averageResponseTime were added to statistics for
> > each response handler. Are these statistics really useful enough to
> > keep as-is?
> >
> > averageResponseTime is cumulative since the server started, so it's
> > not useful for monitoring purposes, but only benchmarking purposes (it
> > won't tell you if your queries are getting slower all of a sudden).
> > (it will also count slower warming queries, not just live queries).
> >
> > requestsPerSecond is likewise flawed... it won't let you detect a
> > flood of traffic or a dropoff. Also, if you turned off traffic to the
> > server yesterday, that will continue to be reflected in the
> > requestsPerSecond today.
> >
> > Since it seems like these parameters are only useful for benchmarking
> > (which can easily be done from log files), perhaps we should defer
> > adding them until we can come up with versions that are useful for
> > monitoring?
> >
> > -Yonik
> >
>
>
Re: requestsPerSecond, averageResponseTime
Posted by Ian Holsman <li...@holsman.net>.
I've been working on a tool to parse log files to get some of this kind
of information as well
it's really alpha, but if your curious the dummy system is here:
http://pyro.holsman.net:9081/top/ -- slightly obfuscated queries (to
roll them up)
http://pyro.holsman.net:9081/overall/?period=5m&hours=12 -- # of
requests, response time, and deviation in that
http://pyro.holsman.net:9081/overall/?period=5m&hours=12&format=csv&cols=1,2,5,6,7,8
- same thing as a CSV file and showing selected columns
The aim is to use this as a data source for something like cacti and
sticking a flash graph on top of it.
If there is enough interest I can contribute this to solr
Yonik Seeley wrote:
> requestsPerSecond and averageResponseTime were added to statistics for
> each response handler. Are these statistics really useful enough to
> keep as-is?
>
> averageResponseTime is cumulative since the server started, so it's
> not useful for monitoring purposes, but only benchmarking purposes (it
> won't tell you if your queries are getting slower all of a sudden).
> (it will also count slower warming queries, not just live queries).
>
> requestsPerSecond is likewise flawed... it won't let you detect a
> flood of traffic or a dropoff. Also, if you turned off traffic to the
> server yesterday, that will continue to be reflected in the
> requestsPerSecond today.
>
> Since it seems like these parameters are only useful for benchmarking
> (which can easily be done from log files), perhaps we should defer
> adding them until we can come up with versions that are useful for
> monitoring?
>
> -Yonik
>