Diagnose Metric Anomalies with Single-Click Machine-Learning Analytics Magic

App Insights Analytics just released Smart Diagnostics, and it is by far the best application of Machine Learning analytics in the service to date.

I’ve posted before about some ML features such as autocluster and smart alerting, but this one really takes the cake as the most powerful and useful yet:

  • It’s super-duper easy to use! Despite the huge complexity of the Machine Learning algo behind the scenes.
  • It’s fast!
  • It can give you awesome answers that save you lots of investigation time and agony.

It works by analyzing spikes in charts, and giving you a pattern that explains the sudden change in the data.

So let’s give it a go!

Analyze spike in dependency duration

I run a service that has all kinds of remote dependencies – calls to Azure blobs, queues, http requests, etc.

In my devops hat, I run this simple query almost daily just to gauge the health of my service – a look at the 95th percentile for call duration by dependency type:

| where timestamp > ago(1d)
| where operation_SyntheticSource == ""
| summarize percentile(duration, 95) by bin(timestamp,30m), type
| render timechart

The results look like this:


Right off the bat I can see something very funky going on in my http calls. I wanna know exactly what’s going on, but drilling in to the raw data can be a messy business.

If only there was a way to analyze that spike with just one click…. !!!

Fortunately, there’s a small purple dot on that spike. It signifies that this spike is available for analysis with Machine Learning (aka Smart Diagnostics).

Once I click on it, the magic happens.


Smart Diagnostics just told me that the cause for the spike in call duration was:

  • name: https://../observation
  • operation_Name: HealthyFlow
  • resultCode: 409
  • success: False


Furthermore, looking at the chart I see that calls without this pattern are totally flat in this time period.

This is definitive proof that the spike was caused by failed calls to this dependency. My work here is done in seconds.


Creating Beautiful Sankey Diagrams From App Insights Custom Events

I came across this tweet the other day:

It sounded a lot like a challenge to me, so I just couldn’t resist!!

Sankey diagrams for those who don’t know are an amazing tool for describing user flows AND are the basis for one of the most famous data visualizations of all-time. But really, I had no idea how to create one. So, I googled Sankey + PowerBI and came across this fantastic Amir Netz video:

So all you need to create the diagram is a table with 3 columns:

  • Source
  • Destination
  • Count

And PowerBI takes care of the rest.

We already know you can take any App Insights Analytics query and “export” the data to PowerBI.

So the only problem now is how do I transform my AppInsights custom events table into a Sankey events table?

Let’s go to Analytics!

First it’s best to have an idea of what we’re trying to do. Decide on 5-10 events that make sense for a user session flow. In my case, I want to see the flow of users through a new feature called “CTD”. So the events I chose are:

  • CTD_Available (feature is available for use)
  • CTD_Result (user used the feature and got a result)
  • CTD_DrillIn (user chose to further drill-in the results)
  • CTD_Feedback (user chose to give feedback)

In every step, I’m interested in seeing how many users I’m “losing”, and what they’re doing next.

Ok, let’s get to work!

First query we’ll

  • Filter out only relevant events
  • Sort by timestamp asc (don’t forget that in this chart, order is important!)
  • Summarize by session_id using makelist, to put all events that happened in that session in an ordered list. If you’re unfamiliar with makelist, all it does is take all the values of the column, and stuffs them into a list. The resulting lists are the ordered events that users triggered in each session.
| where timestamp > ago(7d)
| where name=="CTD_Available" or name=="CTD_Result" or 
        name=="CTD_Drillin" or name== "CTD_Feedback"
| sort by timestamp asc
| summarize l=makelist(name) by session_Id

Next step I’ll do is add an “EndSession” event to each list, just to make sure my final diagram is symmetric. You might already have this event as part of your telemetry, I don’t. This is optional, and you can choose to remove this line.

| extend l=todynamic(replace(@"\]", ',"EndSession" ]', tostring(l))) 

Next step, I’d like to create “tuples” for source and destination from each list. I want to turn:

Available -> Result -> Feedback -> EndSession


[Available, Result], [Result, Feedback], [Feedback, EndSession]

To do this, I need to chop off the first item in the list and “zip” it (like a zipper) with the original list. In c# this is very easy – list.Zip(list.Skip(1))..

Amazingly, App Analytics has a zip command! Tragically, it doesn’t have a skip… :(. Which means we need to do some more ugly regex work in order to chop off the first element.

| extend l_chopped=todynamic(replace(@"\[""(\w+)"",", @"[", tostring(l)))

Then I zip, and use mvexpand, to create one row per tuple created

| extend z=zip(l, l_chopped) 
| mvexpand z

And I remove the lonely “EndSession”s which are useless artifacts.

| where tostring(z[0]) != "EndSession"

Last thing left to do is summarize to get the counts for each unique tuple, and remove the tuples with identical source and destinations.

The final query:

| where timestamp > ago(7d)
| where name=="CTD_Available" or name=="CTD_Result" or 
        name=="CTD_Drillin" or name== "CTD_Feedback"
| sort by timestamp asc
| summarize l=makelist(name) by session_Id
| extend l=todynamic(replace(@"\]", ',"EndSession" ]', tostring(l))) 
| extend l_chopped=todynamic(replace(@"\[""(\w+)"",", @"[", tostring(l)))
| extend z=zip(l, l_chopped) 
| mvexpand z
| where tostring(z[0]) != "EndSession"
| summarize cnt() by source=tostring(z[0]),dest=tostring(z[1])
| where source!=dest

We’re in business!

Now I want to get this data to Power BI…

  • Download Power BI desktop.
  • Goto the Power BI visuals gallery, search “sankey” and download the visuals you want (my preference is “sankey with labels”)
  • Open Power BI, and goto GetData -> BlankQuery
  • In Query editor, goto View -> Advanced Editor
  • Now go back to the prepared AppAnalytics query, and hit export to Power BI


  • Take the query in the downloaded text file, and paste it into the query editor
  • Press done

Now the data should be there! Just one more step!

  • Back in PowerBI, import the visuals you previously downloaded


  • Select Source -> source, Destination -> dest, Weight -> count_



Cool uses for the top-nested operator

There’s a pretty nice operator in Kusto (or App Insights Analytics) called top-nested.

It basically allows you to do a hierarchical drill-down by dimensions. Sounds a bit much, but it’s much clearer when looking at an example!

So a simple use for it could be something like getting the top 5 result-codes, and then a drill down for each result code of top 3 request names for each RC.

| where timestamp > ago(1h)
| top-nested 5 of resultCode by count(),
  top-nested 3 of name by count()

So I can easily see which operation names are generating the most 404’s for instance.

This is pretty cute, and can be handy for faceting.

But I actually find it more helpful in a couple of other scenarios.

First one is getting a chart of only the top N values. For instance, if I chart my app usage by country, I get a gazillion series of all different countries. How can I easily filter the chart to show just my top 10 countries? Well one way is to do the queries separately, and add a bunch of where filters to the chart…

But top nested can save me all that work:

let top_countries = view()
  | where timestamp > ago(3d)
  | top-nested 5 of client_CountryOrRegion by count()
| join kind= inner
    | where timestamp >= ago(3d)
   ) on client_CountryOrRegion
| summarize count() by bin(timestamp, 1h), client_CountryOrRegion
| render timechart


A beautiful view of just my top 5 countries…

I’ve actually used the same technique for a host of different dimensions (top countries, top pages, top errors etc.), and it can also be useful to filter OUT top values (such as top users skewing the numbers), by changing the join to anti-join.

The second neat scenario is calculating percentages of a whole. For instance – how do you calculate the percentage of traffic per RC daily?

Yeah, you can do this using a summarize and the (newly-added) areachart stacked100 chart kind:

| where timestamp >= ago(3d)
| where isnotempty(resultCode)
| summarize count() by bin(timestamp, 1h), resultCode
| render areachart kind=stacked100


But this only partially solves my problem.

Because ideally, I don’t want to look at all these 200’s crowding my chart. I would like to look at only the 40X’s and 500’s, but still as a percentage of ALL my traffic.

I could do this by adding a bunch of countif(rc=403)/count(), countif(rc=404)/count()… ad nauseum, but this is tiresome + you don’t always know all possible values when creating a query.

Here’s where top-nested comes in. Because it shows the aggregated value for each level, creating the percentages becomes super-easy. The trick is simply doing the first top-nested by timestamp:

| where timestamp > ago(14d)
| top-nested 14 of bin(timestamp, 1d) by count() ,
  top-nested 20 of resultCode by count()
| where resultCode !startswith("20")
| where resultCode !startswith("30")
| project pct=aggregated_resultCode * 1.0 / aggregated_timestamp, 
          timestamp, resultCode 
| render timechart


Pretty nice, no?

Calculating Stickiness Using AppInsights Analytics


There is a new, simpler, better way to calculate usage metrics such as stickiness, churn and return rate.

In previous posts I  demonstrated some simple yet nifty tricks to get stuff done in app insights analytics – like extracting data from traces, or joining tables.

Those were mostly pretty simple queries, showing some basic Kusto techniques.

In this post I’m going to show something much more complex, with some advanced concepts.

We’re gonna take it slow, but be warned!

What I wanna do is calculate “Stickiness“. This is a measure of user engagement, or addiction to your app. It’s computed by dividing DAU (daily active users) by MAU (monthly active users) in a rolling 28 day window. It basically shows what percentage of your total user base is using your app daily.

Computing your DAU is pretty simple in analytics, and can be done using a simple dcount aggregation:

| where timestamp > ago(60d)
| summarize dcount(user_Id) by bin(timestamp, 1d)

But how do you compute a rolling 28-day window unique count of users? For this we’re gonna need to get familiar with some new Kusto operators:

hll() – hyperloglog – calculates the intermediate results of a dcount.

hll_merge() – used to merge together several hll intermediate results.

dcount_hll() – used to calculate the final dcount from an hll intermediate result.

range() – generates a dynamic array with equal spacing

mvexpand() – expands a list into rows

let – binds names to expressions. I’ve already shown a use for let in a past post.

It’s kind of a lot, but let’s get going and see how we’re gonna use each of these along the way.

Let’s do this in steps. Our goal is to calculate a moving 28 day window MAU. First thing, instead of dcount we’ll use hll, to get the intermediate results:

| where timestamp > ago(60d)
| summarize hll(user_Id) by bin(timestamp, 1d)

With the intermediate results in place, the next phase is to think about which dates will use each intermediate result. If we take 20/1/2017 as an example, well, we know that each subsequent day, 28 days forward, will want to use this hll for it’s moving window result. So we build a list of [21/1/2017, 22/1/2017 … 18/2/2017].

So what we do here, and this is a little dirty, is create a list of all the future dates that will need this result. We do this using the range operator:

| where timestamp > ago(60d)
| summarize hll(user_Id) by bin(timestamp, 1d)
| extend periodKey = range(bin(timestamp, 1d), timestamp+28d, 1d)

Now let’s turn every item in the periodKey column list, into a row in the table. We’ll do this with mvexpand:

| where timestamp > ago(60d)
| summarize hll(user_Id) by bin(timestamp, 1d)
| extend periodKey = range(bin(timestamp, 1d), timestamp+28d, 1d)
| mvexpand periodKey

So now, when sorting by periodKey, each date in that column has exactly 28 rows, each with an hll from a different date it needs to calculate the total dcount. We’re almost done! Let’s calculate the dcount:

| where timestamp > ago(60d)
| summarize hll(user_Id) by bin(timestamp, 1d)
| extend periodKey = range(bin(timestamp, 1d), timestamp+28d, 1d)
| mvexpand periodKey
| summarize rollingUsers = dcount_hll(hll_merge(hll_user_Id)) by todatetime(periodKey)

That’s the 28 day rolling MAU right there!

Now let’s make this entire query modular, so we can calculate any length rolling dcount we’d like – including a zero day rolling (DAU actually) – and calculate our metric:

let start=ago(60d);
let period=1d;
let RollingDcount = (rolling:timespan)
| where timestamp > start
| summarize hll(user_Id) by bin(timestamp, period)
| extend periodKey = range(bin(timestamp, period), timestamp+rolling, period)
| mvexpand periodKey
| summarize rollingUsers = dcount_hll(hll_merge(hll_user_Id)) by todatetime(periodKey)
| join RollingDcount(0d) on periodKey
| where periodKey < now() and periodKey > start + 28d
| project Stickiness = rollingUsers1 *1.0/rollingUsers, periodKey
| render timechart



App Analytics: Using “Let”, and a really useful investigation query

So here’s just a small tidbit that can be useful.

First the “let” keyword – it basically allows you to bind a name to an expression or to a scalar. This of course is really useful if you plan to re-use the expression.

I’ll give an example that I use in real-life – a basic investigative query into failed requests. I’m joining exceptions and failed dependencies (similar to NRT proactive detection). I’m using the let keyword to easily modify the time range of my query.

Here it is, enjoy!


let investigationStartTime = datetime("2016-09-07");
let investigationEndTime = investigationStartTime + 1d;
| where timestamp > investigationStartTime
| where timestamp < investigationEndTime
| where success == "False"
| join kind=leftouter(exceptions
   | where timestamp > investigationStartTime
   | where timestamp < investigationEndTime
   | project exception=type , operation_Id ) on operation_Id
| join kind=leftouter (dependencies
   | where timestamp > investigationStartTime
   | where timestamp < investigationEndTime
   | where success == "False"
   | project failed_dependency=name, operation_Id ) on operation_Id
| project timestamp, operation_Id , resultCode, exception, failed_dependency

Tweaking Proactive Alerts

I’ve already talked about how cool proactive alerts are, but one thing I missed is that you can actually tweak these alerts through the Azure portal.

Go to the “Alerts” blade, and there you should find a single “Proactive Diagnostics” alert.


If you click it and go to the alert configuration, you can set email recipients, setup a webhook,  and enable/disable.

One thing that is really useful is you can set “Received detailed analysis” checkbox:


This will make sure you get the entire deep-dive analysis of the incident straight to your email inbox, without needing to go to the portal. This can save some very valuable minutes during a live-site incident!