I'm having a tough time with Google Analytics, trying to understand why the value of metrics changes when segments are applied.
There is a standard audience overview report, which is based on 100% of sessions (no sampling) and the view is not filtered. The period is March of 2017.
Standard "All visitors" segment looks like this:
Then, there is another built-in segment called "Bounced Sessions". When I apply this segment, the "All visitors" values changes:
Amount of users increases, but the count of pageviews decreases.
Any ideas how to explain this?.. Thank you in advance!
Oki, there can be, multiple reasons. Let me explain first how these numbers are calculated, then we move on to your query.
There two types of data gathering and manipulation from google.
Pre-calculated data -- pre-aggregated tables
These are the precalculated data that Google uses to speed up the UI. Google does not specify when this is done but it can be at any point of the time. These are known as pre-aggregated tables
Data calculated on the fly
Some that you do which result in computation or manipulation falls under this category. Like using segments, creating custom reports etc.
Coming to your problem. When you apply segment, every metric that it effects will be calculated again. Thus it may result in numbers greater than you see in normal view.
Standard audience overview report is pre-aggregated at some point of the day. When you apply segment, the results will be calculated with the fresh data. Since latter is the latest, it will automatically give you increased number of the metrics. Even you can see a decrease as well, all depends on your data and user behavior.
Resolution: If you are a premium user, use Big Query. You must rely on big query for every metric as they are fresh and calculated on the fly
Related
I'm logging some custom metrics in Application insights using the TelemetryClient.TrackMetric method in .NET, and I've noticed that occasionally some of the events are duplicated when I view them in the Azure portal.
I've drilled into the data, and the duplicate events have the same itemId and timestamp, but if I show the ingestion time by adding | extend ingestionTime = ingestion_time() to the query then I can see that the ingestion times are different.
This GitHub issue indicates that this behavior is expected, as AI uses at-least-once delivery.
I plot these metrics in charts in the Azure portal using a sum aggregation, however these duplicates are creating trust issues with the charts as the duplicates are simply treated as two separate events.
Is there a way to de-dupe the events based on itemId before plotting the data in the Azure portal?
Update
A more specific example:
I'm running an algorithm, triggered by an event, which results in a reward. The algorithm may be triggered several dozen times a day, and the reward is a positive or negative floating point value. It logs the reward each time to Application Insights as a custom metric (called say custom-reward), along with some additional properties for data splitting.
In the Azure portal I'm creating a simple chart by going to Application Insights -> Metrics and customising the chart. I select my custom-reward metric in the Metric dropdown, and select Sum as the aggregation. I may or may not apply splitting. I save the chart to my dashboard.
This simple chart gives me a nice way of monitoring the system to make sure nothing unexpected is happening, and the Sum value in the bottom left of the chart allows me to quickly see whether the sum of the rewards is positive or negative over the chart's range, and by how much.
However, on occasion I've been surprised by the result (say over the last 12 hours the sum of the rewards was surprisingly negative), and on closer inspection I discovered that a few large negative results have been duplicated. Further investigation shows this has been happening with other events, but with smaller results I tend not to notice.
I'm not that familiar with the advanced querying bit of Application Insights, I actually just used it for the first time today to dig into the events. But it does sound like there might be something I can do there to create a query that I can then plot, with the results deduped?
Update 2
I've managed to make progress with this thanks to the tips by #JohnGardner, so I'll mark that as the answer. I've deduped and plotted the results by adding the following line to the query:
| summarize timestamp=any(timestamp), value=any(value), name=any(name), customDimensions=any(customDimensions) by itemId
Update 3
Adding the following line to the query allowed me to split on custom data (in this case splitting by algorithm ID):
| extend algorithmId = tostring(customDimensions.["algorithm-id"])
With that line added, when you select "Chart" in the query results, algorithmId now shows up as an option in the split dropdown. After that you can click "Pin to dashboard". You lose the handy "sum over the time period" indicator in the bottom left of the chart which you get via the simple "Metrics" chart, however I'm sure I'll be able to recreate that in other ways.
if you are doing your own queries, you would generally be using something like summarize or makeseries to do this deduping for a chart. you wouldn't generally plot individual items unless you are looking at a very small time range?
so instead of something like
summarize count() ...
you could do
summarize dcount(itemId) ...
or you might add a "fake" summarize to a query that didn't need it before with by itemId to coalesce multiple rows into just one, using any(x) to grab any individual row's value for each column for each itemId.
but it really depends on what you are doing in your specific query. if you were using something like sum(itemCount) to also deal with sampling, you have other odd cases now, where the at-least-once delivery might have duplicated sampled items? (updating your question to add a specific query and hypothetical result would possibly lead to a more specific answer).
I would love to understand what I'm looking at - why are the numbers different in this report when I add a segment?
This is the report without any segmentation:
This is the same report with the Mobile Traffic segment:
There two methods that Google uses to identify the number of users.
Calculation 1: Pre-calculated data
This calculation relies only on the number of sessions in the given date range and the time of each session. (This is determined by technology managed on the device, like a web browser, and is often referred to as the client-side time.) Because the result of this calculation can be added to the pre-aggregated data tables, Analytics can reference the table to quickly retrieve and serve this data in a report, including when you change the date range.
Calculation 2: Data calculated on the fly
Calculation 2 is based on the way you assign, collect, and store persistent data about your traffic. There are many solutions you can implement to customize this, but the most common way this data is going to be assigned and stored is through cookies managed via a web browser.
Adding a segment will force GA to calculate the data on the fly and that's why you are seeing a difference in the numbers.
Are you using GA free or 360? and the time range you are using is same in both reports?
You can also have a look into the Google article https://support.google.com/analytics/answer/2992042?hl=en
You are victim of sampling:
https://support.google.com/analytics/answer/2637192?hl=en
Sampling applies when:
you customize the reports
the number of sessions for the report time range exceeds 500K (GA) or 100M (GA 360)
The consequence is that:
the report will be based on a subset of the data (the % depends on the total number of sessions)
therefore your report data won't be as accurate as usual
What you can do to reduce sampling:
increase sample size in UI (will only decrease sampling to a certain extend, but in most cases won't completely remove sampling)
reduce time range
create filtered views so your reports contain the data you need and you don't have to customize them
The goal flow report on my google analytics account shows some strange sampling behavior. While I can usually select up to a month of data before sampling starts it seems to be different for the goal flow report.
As soon as I select more than one day of data the used data set is getting smaller very fast. At three days the report ist based on only 50% of the sessions, which, according to analytics, comes to only 35 sessions.
Has anyone experienced a similar behavior of sampling although only very small data-sets are used?
Sampling is induced when your request is calculation-intensive; there's no 'garunteed point at which it trips.
Goal flow complexity will increase exponentially as you add goals, so even a low number of goals might make this report demand a lot of processing.
Meanwhile you'll find that moast of the standard reports can cover large periods of time without sampling; they are preaggreated, so it's very cheap to load them.
If you want to know more about sampling, see here:
https://stackoverflow.com/a/37386181/5815149
Is there a way to simply show the change of a value over the selected time period? All I'm interested in is the offset of the last value compared to the initial one. The values can vary above and below these over the time period, it's not really relevant (and would be exceptions in my case).
For an initial value of 100 and an final value of 105, I'd expect a single stat box displaying 5%.
I have the feeling I'm missing something obvious obvious, but can't find a method to display this deceptively simple task.
Edit:
I'm trying to create a scripted Grafana dashboard that will automatically populate disk consumption growth for all our various volumes. The data is already in Graphite, but for purposes of capacity management and finance planning (which projects/departments gets billed) it would be helpful for managers to have a simple and coarse overview of which volumes grow outside expected parameters.
The idea was to create a list of single-stat values with color coding that could easily be scrolled through to find abnormalities. Disk usage would obviously never be negative, but volatility in usage between the start and end of the time period would be lost in this view. That's not a big concern for us as this is all shared storage and such usage is expected to a certain degree.
The perfect solution would be to have the calculations change dynamically based on the selected time period.
I'm thinking that this is not really possible (at least not easily) to do with just Graphite and Grafana and have started looking for alternative methods. We might have to implement a different reporting system for this purpose.
Edit 2
I've tried implementing the suggested solution from Leonid, and it works after a fashion. The calculations seems somewhat off from what I expected though.
My test dashboard looks like follows:
If I were to calculate the change manually, I'd end up with roughly 24% change between the start (7,23) and end (8.96) value. Graphite calculates this to 19%. It's probably a reason for the discrepancy, probably something to do with it being a time-series and not discreet values?
As a sidenote: The example is only 30 days, even though the most interesting number would be a year. We don't have quite a year of data in Graphite yet and having a 30 day view is also interesting. It seems I have to implement several dashboards with static times.
You certainly can do that for some fixed period. For example following query take absolute difference betweent current metric value and value that metric has one minute ago (i.e. initial value) and then calculate it's percentage of inital value.
asPercent(absolute(diffSeries(my_metric, timeShift(my_metric, '1m'))), timeShift(my_metric, '1m'))
I believe you can't do that for time period selected in Grafana picker.
But is that really what you need? It's seems strange because as you said value can change in both directions. Maybe standard deviation would be more suitable for you? It's available in Graphite as stdev function.
I try to pull out the (unique) visitor count for a certain directory using three different methods:
* with a profile
* using an dynamic advanced segment
* using custom report filter
On a smaller site the three methods give the same result. But on the large site (> 5M visits/month) I get a big discrepancy between the profile on one hand and the advanced segment and filter on the other. This might be because of sampling - but the difference is smaller when it comes to pageviews. Is the estimation of visitors worse and the discrepancy bigger when using sampled data? Also when extracting data from the API (using filters or profiles) I still get DIFFERENT data even if GA doesn't indicate that the data is sampled - ie I'm looking at unsampled data.
Another strange thing is that the pageviews are higher in the profile than the filter, while the visitor count is higher for the filter vs the profile. I also applied a filter at the profile to force it to use sample data - and I again get quite similar results to the filter and segment-data.
profile filter segment filter#profile
unique 25550 37778 36433 37971
pageviews 202761 184130 n/a 202761
What I am trying to achieve is to find a way to get somewhat accurat data on unique visitors when I've run out of profiles to use.
More data with discrepancies can be found in this google docs: https://docs.google.com/spreadsheet/ccc?key=0Aqzq0UJQNY0XdG1DRFpaeWJveWhhdXZRemRlZ3pFb0E
Google Analytics (free version) tracks only 10 mio page interactions [0] (pageviews and events, any tracker method that start with "track" is an interaction) per month [1], so presumably the data for your larger site is already heavily sampled (I guess each of you 5 Million visitors has more than two interactions) [2]. Ad hoc reports use only 1 mio datapoints at max, so you have a sample of a sample. Naturally aggregated values suffer more from smaller sample sizes.
And I'm pretty sure the data limits apply to api access too (Google says that there is "no assurance that the excess hits will be processed"), so for the large site the api returns sampled (or incomplete) data, too - so you cannot really be looking at unsampled data.
As for the differences, I'd say that different ad hoc report use different samples so you end up with different results. With GA you shouldn't rely too much an absolute numbers anyway and look more for general trends.
[1] Analytics Premium tracks 50 mio interactions per month (and has support from Google) but comes at 150 000 USD per year
[2] Google suggests to use "_setSampleRate()" on large sites to make sure you have actually sampled data for each day of the month instead of random hit or miss after you exceed the data limits.
Data limits:
http://support.google.com/analytics/bin/answer.py?hl=en&answer=1070983).
setSampleRate:
https://developers.google.com/analytics/devguides/collection/gajs/methods/gaJSApiBasicConfiguration#_gat.GA_Tracker_._setSampleRate
Yes, the sampled data is less accurate, especially with visitor counts.
I've also seen them miss 500k pageviews over two days, only to see them appear in their reporting a few days later. It also doesn't surprise me to see different results from different interfaces. The quality of Google Analytics has diminished, even as they have tried to become more real-time. It appears that their codebase is inconsistent across API's, and their algorithms are all over the map.
I usually stick with the same metrics and reporting methods, so that my results remain comparable to one another. I also run GA in tandem with Gaug.es, as a validation and sanity check. With that extra data, I choose the reporting method in GA that I am most confident with and I rely on that exclusively.