Hi All, I have implemented a flatmap function and I want to collect metrics for average time for this function which I plan to monitor via prometheus. What would be good approach for it? I have added a gauge to the method(extending Gauge interface from flink API). Would it work for my needs? |
Sure, that might work. Be aware though that time measurements are,
compared to the logic within a function, usually rather expensive and may impact performance. On 1/12/2021 10:57 AM, Manish G wrote: > Hi All, > > I have implemented a flatmap function and I want to collect metrics > for average time for this function which I plan to monitor via prometheus. > > What would be good approach for it? I have added a gauge to the > method(extending Gauge interface from flink API). Would it work for my > needs? > > |
My code is: And in flatmap function:
On Tue, Jan 12, 2021 at 3:53 PM Chesnay Schepler <[hidden email]> wrote: Sure, that might work. Be aware though that time measurements are, |
startTime is set at start of function: long startTime = System.currentTimeMillis(); On Tue, Jan 12, 2021 at 3:59 PM Manish G <[hidden email]> wrote:
|
A gauge just returns a value, and Flink
exposes it as is. As such you need to calculate the average over
time yourself, taking 2 time measurements (before and after the
processing of each).
On 1/12/2021 11:31 AM, Manish G wrote:
|
OK, got it. So I would need to accumulate the time value over the calls as well as number of times it is called...and then calculate average(accumulated time/ number of times called) and then set calculated value into gauge as above. On Tue, Jan 12, 2021 at 4:12 PM Chesnay Schepler <[hidden email]> wrote:
|
That approach will generally not work
for jobs that run for a long time, because it will be nigh
impossible for anomalies to affect the average. You want to look
into exponential moving averages.
Alternatively, just expose the diff as
an absolute value and calculate the average in prometheus.
On 1/12/2021 11:50 AM, Manish G wrote:
|
Can you elaborate the second approach more? Currently I am exposing the difference itself. OR do you mean the cumulative difference?ie I maintain a member variable, say timeSoFar, and update it with time consumed by each method call and then expose it. Something like this: timeSoFar += timeConsumedByCurrentInvocation this.simpleGaug.setValue(
timeSoFar
); On Tue, Jan 12, 2021 at 4:24 PM Chesnay Schepler <[hidden email]> wrote:
|
I mean the difference itself, not
cumulative.
On 1/12/2021 12:08 PM, Manish G wrote:
|
The cumulative time probably isn't that
useful to detect changes in the behavior of the application.
On 1/12/2021 12:30 PM, Chesnay Schepler
wrote:
|
Prometheus provides avg_over_time for a range vector. That seems to be better suited for this usecase. On Tue, Jan 12, 2021 at 6:53 PM Chesnay Schepler <[hidden email]> wrote:
|
This approach has an issue. Even for those periods when there is no activity, still the latest gauge value is used for calculations and this generates graphs which are not correct representation of the situation. On Tue, Jan 12, 2021 at 7:01 PM Manish G <[hidden email]> wrote:
|
If you want the gauge to only represent
recent activity then you will need to use a timer of sorts to
reset the gauge after N time (something larger than the reporter
interval) unless it was changed in the meantime (e.g., by also
recording a timestamp within SimpleGauge)
On 1/13/2021 9:33 AM, Manish G wrote:
|
Free forum by Nabble | Edit this page |