esr_linux_uncounted.Rmd

---
title: Have we missed profiles on ESR and Linux?
author: Saptarshi Guha <joy@mozilla.com>
date: "`r format(Sys.time(), '%H:%M %B %d, %Y',tz='America/Los_Angeles',usetz=TRUE)`"
output:
 html_document:
    mathjax: default
    self_contained: false
    theme: readable
    highlight: haddock
    number_sections: true
    toc_depth: 5
    toc: true
---



```{r child = '/home/sguha/mz/template2.Rmd'}
```


```{r options, echo=FALSE}
knitr::knit_engines$set(python = reticulate::eng_python)
```

# Background

For much more informed background speak with nanj on slack.Briefly, Activity
Stream sent a daily API call that didn't respect the privacy settings of the
browser. This call was sent daily and once to our servers for versions 40 ... 59
inclusive. See
[this](https://sql.telemetry.mozilla.org/queries/52620/source#140092) for an
example query.

On ESR, companies often turn off telemetry (which is our way of getting profile
counts) and Linux distros customize to their taste which often includes turning
of telemetry.

With that said, i don't see any reason why the blocklist ping be turned of (since
this has no client information, cannot be used to track and protects the
browser). Hence the AS ping ought to be similar to the blocklist ping . However,
given this is an assumption, I'm inclined to believe the AS ping as the true ESR
and Linux number assuming it roughly matches the block list/submission_date based
daily counts on other OS/Channels.



# Getting Data

## AS Data
I used [this](https://sql.telemetry.mozilla.org/queries/53315/source)  query to
get the AS counts by version, os and date.

```{r as1, cache=TRUE}
as <- fread("New_Query_2018_05_14.csv",head=TRUE)
setnames(as,c("date","version","os","asi"))
as <- as[, isesr:=grepl("^(52|45)", version)][,]
as <- as[, os2:=sapply(os,function(s){
    if(grepl("Windows",s)) "Windows"
    else if(s %in% c("Arch Linux","BSD","CentOS","Debian","Fedora",
                      "FreeBSD","Gentoo","Kubuntu","Linux","Linux Mint","Mandriva",
                      "NetBSD","OpenBSD","openSUSE","PCLinuxOS","Red Hat","Slackware",
                      "Solaris","SUSE","Ubuntu")) "Linux"
    else if(s %in% c("Mac OS","Mac OS X")) "OSX"
    else "Others"
})][,]
as2 <- as[grepl("^(40|41|42|43|44|45|46|47|48|49|50|51|52|53|54|55|56|57|58|59)\\.",version)==TRUE,]
as2 <- as2[,ver2:=substring(version,1,2)][,]
as2 <- as2[, asi:=1*asi][,]
d1 <- strftime(as.Date(min(as2$date)), "%Y%m%d")
d2 <- strftime(as.Date(max(as2$date)),"%Y%m%d")
```

## Blocklist Count

```{r as2, cache=TRUE,eval=TRUE}
d <- odbc()
adi <- data.table(d$q(sprintf("select
 bl_date as date,
 prod_os as os,
 v_prod_major as version,
 sum(tot_requests_on_date) as adi
 from
 copy_adi_dimensional_by_date
 where bl_date >='%s' and bl_date <='%s' and product='Firefox'
 group by 1,2,3
 order by 1,2,3
",min(as2$date),max(as2$date))))

adi <- adi[, os2:=sapply(os,function(s){
    if(grepl("Windows",s)) "Windows"
    else if(s %in% c("Arch Linux","BSD","CentOS","Debian","Fedora",
                      "FreeBSD","Gentoo","Kubuntu","Linux","Linux Mint","Mandriva",
                      "NetBSD","OpenBSD","openSUSE","PCLinuxOS","Red Hat","Slackware","SunOS",
                      "Solaris","SUSE","Ubuntu")) "Linux"
    else if(s %in% c("Darwin","Mac OS","Mac OS X")) "OSX"
    else "Others"
})][,]
adi2 <- adi[grepl("^(40|41|42|43|44|45|46|47|48|49|50|51|52|53|54|55|56|57|58|59)\\.",version)==TRUE,]
adi2 <- adi2[,ver2:=substring(version,1,2)][,] 

```

## DAU count based of `submission_date_s3`

```{pydbx as3, cache=TRUE, dependson='as1',storein='dau'}
dau=spark.sql("""
select 
submission_date_s3,
app_version as version,
os as os,
count(distinct(client_id))*100 as dau
from main_summary
where app_name='Firefox'
and sample_id='42'
and submission_date_s3>='(__REPLACE__d1)' and submission_date_s3<='(__REPLACE__d2)'
group by 1,2,3 order by 1,2,3
""").toPandas()
```

## Munging
```{r as4, cache=TRUE, dependson=c('as1','as2','as3')}
dau <- dau[, os2:=sapply(os,function(s){
    if(grepl("Windows",s)) "Windows"
    else if(s %in% c("Arch Linux","BSD","CentOS","Debian","Fedora",
                      "FreeBSD","Gentoo","Kubuntu","Linux","Linux Mint","Mandriva",
                      "NetBSD","OpenBSD","openSUSE","PCLinuxOS","Red Hat","Slackware","SunOS",
                      "Solaris","SUSE","Ubuntu")) "Linux"
    else if(s %in% c("Darwin","Mac OS","Mac OS X")) "OSX"
    else "Others"
})][,]
dau2 <- dau[grepl("^(40|41|42|43|44|45|46|47|48|49|50|51|52|53|54|55|56|57|58|59)\\.",version)==TRUE,]
dau2 <- dau2[,ver2:=substring(version,1,2)][,] 
dau2 <- dau2[, date:=as.character(as.Date(as.character(submission_date_s3),format="%Y%m%d"))][, submission_date_s3:=NULL][,]

adi3 <- adi2[,list(adi=sum(adi)), by=list(date, os2,ver2)]
as3 <- as2[, list(asi=sum(asi)), by=list(date,os2,ver2)]
dau3 <- dau2[,list(dau=sum(dau)), by=list(date,os2,ver2)]
ma <- merge(merge(as3,adi3, by=c("date","os2","ver2")),dau3,by=c("date","os2",'ver2'))
ma <- ma[, isesr:=grepl("^(52|45)", ver2)][,]
```

    
# Analyses

## Definitions

- **asi**  unique installations active on a day based off the AS ping
- **adi**  unique installations active on a day based off the Blocklist ping
- **dau**  unique installations active on a day based off Unified Telemetry

## (BASELINE) Compare Known Values (Windows Non ESR)

Windows,non ESR ADI and ASI track each other fairly well. ADI is larger. That
they track each other implies ASI is a consistent signal of volume, though the
number may not be entirely accurate.

```{r fig.width=10,cache=TRUE,fig.cap='Plot of ASI, ADI and DAU (7 day smoothed) (WINDOWS ONLY, NON-ESR)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[!isesr & os2=='Windows', list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]

xyplot( sdau/1e6+sasi/1e6+sadi/1e6 ~ date, data=ma2,type=c('g','l'),auto.key=list(columns=3)
       ,scales=list(x=list(format='%Y-%m-%d',rot=45)))
```

The differences are in the order of a few millions and increasing over time
(i.e. ADI is greater than ASI by a few million). Also  the relative difference  is
increasing time.


<div class = "row">

<div class = "col-md-6">
```{r cache=TRUE,fig.cap='Plot of ASI,ADI  difference (WINDOWS, non ESR)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[!isesr & os2=='Windows', list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]
ma2 <- ma2[, reldiff:= (sadi - sasi)]
xyplot( reldiff/1e6~ date, data=ma2,type=c('g','l','r'),ylab='Difference\n(ADI-ASI) mm',
       scales=list(x=list(format='%Y-%m-%d',rot=45)))
```
</div>


<div class = "col-md-6">
```{r cache=TRUE,fig.cap='Plot of ASI,ADI  relative difference %(WINDOWS, non ESR)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[!isesr & os2=='Windows', list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]
ma2 <- ma2[, reldiff:= (sadi - sasi)/sasi * 100]
xyplot( reldiff~ date, data=ma2,type=c('g','l','r'),ylab='Relative Difference\n(ADI-ASI)/ASI %',
       scales=list(x=list(format='%Y-%m-%d',rot=45)))
```
</div>
</div>


## Compare ESR (all OS)

Plot of ESR volumes. Based on the veracity of the above figures, ESR is not
being completely measured by current methods(ADI/Telemetry).


```{r fig.width=10,cache=TRUE,fig.cap='Plot of ASI, ADI and DAU (7 day smoothed) (ESR)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[isesr==TRUE, list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]

xyplot( sdau/1e6+sasi/1e6+sadi/1e6 ~ date, data=ma2,type=c('g','l'),auto.key=list(columns=3)
       ,scales=list(x=list(format='%Y-%m-%d',rot=45)))
```



The difference between ADI and ASI (ASI-ADI) is increasing and is in the order
of several millions. The discrepancy is large: for Windows, Non ESR there is a
6-8% relative difference between ADI and ASI. For ESR, there is a 45-55%
relative difference!!

Looking at the trend, the discrepancy  might even be getting larger over time.

**As opposed to the graphs above, we look at $\frac{ASI-ADI}{ASI}$ (not
$(ADI-ASI)$) since ASI for ESR/ Linux is higher than ADI**

<div class = "row">
<div class = "col-md-6">

```{r cache=TRUE,fig.cap='Plot of ASI,ADI  difference (ESR)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[isesr==TRUE, list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]
ma2 <- ma2[, reldiff:= (sasi - sadi)]
xyplot( reldiff/1e6~ date, data=ma2,type=c('g','l','r'),ylab='Difference\n(ASI-ADI) mm',
       scales=list(x=list(format='%Y-%m-%d',rot=45)))
```
</div>
<div class = "col-md-6">

```{r cache=TRUE,fig.cap='Plot of ASI,ADI relative  difference %(ESR)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[isesr==TRUE, list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]
ma2 <- ma2[, reldiff:= (sasi - sadi)/sasi*100]
xyplot( reldiff~ date, data=ma2,type=c('g','l','r'),ylab='Difference\n(ASI-ADI)/ASI %',
       scales=list(x=list(format='%Y-%m-%d',rot=45)))
```
</div>
</div>



## Compare Linux 

Plot of Linux (irrespective of ESR)  volumes. Based on the veracity of the above figures, ESR is not
being completely measured by current methods(ADI/Telemetry).

**Cannot explain odd bump in March 2018**

```{r fig.width=10,cache=TRUE,fig.cap='Plot of ASI, ADI and DAU (7 day smoothed) (Linux)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[os2=='Linux', list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]

xyplot( sdau/1e6+sasi/1e6+sadi/1e6 ~ date, data=ma2,type=c('g','l'),auto.key=list(columns=3)
       ,scales=list(x=list(format='%Y-%m-%d',rot=45)))
```



The difference between ADI and ASI (ASI-ADI) is increasing and is in the order
of several millions. The discrepancy is large: for Windows, Non ESR there is a
6-8% relative difference between ADI and ASI. For Linux, there is a  70-90%
relative difference!!

Looking at the trend, the discrepancy  might even be getting larger over time.


<div class = "row">
<div class = "col-md-6">

```{r cache=TRUE,fig.cap='Plot of ASI,ADI  difference (Linux)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[os2=='Linux', list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]
ma2 <- ma2[, reldiff:= (sasi - sadi)]
xyplot( reldiff/1e6~ date, data=ma2,type=c('g','l','r'),ylab='Difference\n(ASI-ADI) mm',
       scales=list(x=list(format='%Y-%m-%d',rot=45)))
```
</div>
<div class = "col-md-6">

```{r cache=TRUE,fig.cap='Plot of ASI,ADI relative  difference %(Linux)',echo=FALSE}
ma <- ma[, date:=as.Date(date)][,]
ma2 <- ma[os2=='Linux', list(adi=sum(adi),asi=sum(asi),dau=sum(dau)),by=list(date)]
ma2 <- ma2[,":="(sasi=filter(asi, rep(1,7)/7, sides=1),
                 sadi=filter(adi, rep(1,7)/7, side=1),
                 sdau=filter(dau, rep(1,7)/7, side=1))][,]
ma2 <- ma2[, reldiff:= (sasi - sadi)/sasi*100]
xyplot( reldiff~ date, data=ma2,type=c('g','l','r'),ylab='Difference\n(ASI-ADI)/ASI %',
       scales=list(x=list(format='%Y-%m-%d',rot=45)))
```
</div>
</div>