The goal of this tutorial is to highlight how to perform complex data transformation and validation operations, and output the results to either to disk or a database. This is commonly referred to as ‘ETL’ for Extract-Transform-Load.
Again we’ll use the tutorials/woodworking directory from the git repository.
Since we are not using any worker nodes in this demo, we configure Essentia to be in ‘local’ mode using:
Essentia$ ess instance local
Essentia treats data as a ‘stream’, similar to Unix pipes. As an example, let’s simply count the lines in one week of the log files that we classified in the previous tutorial:
$ for i in {1..7}; do funzip ./diy_woodworking/browse_2014090${i}.gz | wc -l ; done
On some systems, you may use zcat instead of funzip.
The Essentia equivalent is:
$ ess task stream browse 2014-09-01 2014-09-07 'wc -l'
Some notes here. The bash version is fairly straightforward in this case, but gets much more complicated if you
want to traverse dates that span weeks, months, or years. Essentia handles the decompression of the data and
piping it to the command you specify. All you need to do is specify the category and date range to process.
Since we are in local mode, each file is processed sequentially. If we had worker nodes (i.e. the cloud version), the processing would be done in parallel, with each node responsible for a subset of the files.
aq_pp commandThe above is a rather trivial ETL example. More typical cases involve validating the data, filtering data, and deriving entirely new columns based on some math operations or string processing. For that, we developed a set of command line programs called the “AQ tools”, which are part of the Essentia distribution.
Written in C to achieve a high level of performance, the AQ tools are able to manipulate and transform raw input
data into a format more easily handled by other AQ or third party tools. In particular,
the aq_pp program does the heavy lifting for all ETL operations.
The command structure of aq_pp consists of an input specification specifying which file(s) to take the data
from,
various processing specifications to determine how data is processed, and output specifications describing how
and where to put the results of your command.
There are also a variety of global options that modify the environment and default variables used in aq_pp.
The following provides some working examples of aq_pp commands. Data and scripts are found under
tutorials/etl-engine in the git repository.
First let’s create a simple command that imports our example file chemistry.csv and defines its columns.
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin
-f specifies the file to operate on (chemistry.csv). It accepts an optional ATTRIBUTE in ,+1, which
means to skip the first line (header in this case)-d defines the column names and types. The format is t,attribute:name with ‘t’ being the type. An ‘X’
means to ignore a column. In this example, we load the names and final grades as strings (forcing the last name to
be upper case), the student id as an integer, and the midterm grade as a float.Since there are no processing or output specifications given, the the output is simply:
"id","lastname","firstname","chem_mid","chem_fin"
1,"DAWSON","Leona",76.5,"B-"
2,"JORDAN","Colin",25.899999999999999,"D"
3,"MALONE","Peter",97.200000000000003,"A+"
By default, aq_pp will validate the input against the type you defined it as. For instance if a letter grade
was accidentally placed in lieu of the midterm percentage, the program will exit with an error. By specifying the
optional eok attribute along with -f,+1, the program will simply ignore the input row.
This feature makes it easy to produce validated output.
The process specs define transformation operations on your data. They fall into three groups: * Conversion operations (string to numeric and vice versa) * Numerical operations (math etc) * String operations (merge strings, extract substrings, etc)
For a simple example, let’s say that the midterm grades for the chemistry final need to be revised downward so that the distribution falls within acceptable limits (i.e. grading on a curve):
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin \
-evlc 'chem_mid' 'chem_mid*0.8'
"id","lastname","firstname","chem_mid","chem_fin"
1,"DAWSON","Leona",61.200000000000003,"B-"
2,"JORDAN","Colin",20.719999999999999,"D"
3,"MALONE","Peter",77.760000000000005,"A+"
Here we use the math switch evlc to adjust the chem_mid column down 20%.
By default, all known columns are output to stdout. The -o switch allows users to specify an output file, and
the -c switch allows one to designate explicitly what columns to output.
For example:
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin \
-c id chem_fin
"id","chem_fin"
1,"B-"
2,"D"
3,"A+"
This simply restricts the output to the two designate columns:
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin \
-o newtable.csv -c id chem_fin
Similar, but the output is to a file named newtable.csv instead of the stdout.
Instead of the output being routed into the stdout or a file, it can also be directly imported into the UDB, which is an extremely powerful part of the Essentia toolkit. We expand on this more in the In-memory Database tutorial.
There are a number of scenarios (particularly with log data) where merging two different types of files is useful. Lets consider the case where we want to merge our chemistry and physics grades into a single table:
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin \
-cat,+1 physics.csv i:id s,up:lastname s:firstname f:phys_mid s:phys_fin
"id","lastname","firstname","chem_mid","chem_fin","phys_mid","phys_fin"
1,"DAWSON","Leona",76.5,"B-",0,
2,"JORDAN","Colin",25.899999999999999,"D",0,
3,"MALONE","Peter",97.200000000000003,"A+",0,
1,"DAWSON","Leona",0,,88.5,"A"
3,"MALONE","Peter",0,,77.200000000000003,"B"
4,"CANNON","Roman",0,,55.799999999999997,"C+"
The -cat option is used for such a merge, and it is easiest to think of it as the aq_pp specific version of
the unix command of the same name. The difference here is that aq_pp will create new columns in the output,
while simply concatenating the two files will result in just the same 5 columns as before.
However most users will want to JOIN datasets based on common values between two files. In this case, the first and
last name, as well as the country, are the common columns between the two files. The -cmb option is similar to
-f and -d since it defines the number of lines to skip and the column specification for the second file.
Records will be matched based on all the columns that share the same names between the two files. For example:
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin \
-cmb,+1 physics.csv i:id X X f:phys_mid s:phys_fin
"id","lastname","firstname","chem_mid","chem_fin","phys_mid","phys_fin"
1,"DAWSON","Leona",76.5,"B-",88.5,"A"
2,"JORDAN","Colin",25.899999999999999,"D",0,
3,"MALONE","Peter",97.200000000000003,"A+",77.200000000000003,"B"
Users familiar with SQL will recognize this as a LEFT OUTER JOIN. All the data from the first file is preserved,
while data from the second file is included when there is a match. Where there is no match,
the value is 0 for numeric columns, or the empty string for string columns. In this case,
since the label i:id is common between both file specifications, that is the join key.
We could also have joined based off multiple keys as well: For example matching first AND last
names will achieve the same result:
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin \
-cmb,+1 physics.csv X s,up:lastname s:firstname f:phys_mid s:phys_fin
An important type of dataset joining is replacing some value in a file with a matching entry in a lookup table. In the following example, we wish to convert a students letter grade from ‘A,B,C...’ etc into a simple PASS/FAIL:
aq_pp -f,+1 chemistry.csv -d i:id s,up:lastname s:firstname f:chem_mid s:chem_fin \
-sub,+1,pat chem_fin grades.csv
"id","lastname","firstname","chem_mid","chem_fin"
1,"DAWSON","Leona",76.5,"PASS"
2,"JORDAN","Colin",25.899999999999999,"FAIL"
3,"MALONE","Peter",97.200000000000003,"PASS"
Note the use of the pat attribute when we designate the lookup table. This means that column 1 of the lookup
table can have a pattern instead of a static value. In our case, we can cover grades ‘A+,A,
and A-‘ by the pattern ‘A*’.
The -cmb can be used substituting data, but for situations similar to the one above, -sub is preferred because:
-cmb does. Values are modified in place.-sub can match regular expressions and patterns, while -cmb is limited to exact matches.-sub is faster.The input specification defines all the input columns we have to work with. The goal of the process spec is to modify these data according to various rules.
The -evlc switch allows users to overwrite or create entirely new columns based on some operation with existing
columns or built-in variables. The types of operations are broad, covering both string and numerical data.
For example, if we want to merge our id, ‘first’ and ‘last’ name columns from the chemistry file to create a new column, we can do:
aq_pp -f,+1 chemistry.csv -d i:id s:lastname s:firstname f:chem_mid s:chem_fin \
-evlc s:fullname 'ToS(id)+"-"+firstname+" "+lastname'
"id","lastname","firstname","chem_mid","chem_fin","fullname"
1,"Dawson","Leona",76.5,"B-","1-Leona Dawson"
2,"Jordan","Colin",25.899999999999999,"D","2-Colin Jordan"
3,"Malone","Peter",97.200000000000003,"A+","3-Peter Malone"
Note the use of a built in function ToS which converts a numeric to a string. There are many such built in
functions, and users are free to write their own to plug into the AQ tools. Note also that since we created a new
column, we had to provide the ‘column spec’, which in this case is s:fullname to designate a string labeled
“fullname”.
It may be useful to note the the record number or file ID in the output table. The aq_pp handles this via
built-in variables. In the example below, we augment the output with a row number. We add 1 to it to compensate for
skipping the header via the -f,+1 flag
aq_pp -f,+1 chemistry.csv -d i:id s:lastname s:firstname f:chem_mid s:chem_fin \
-evlc i:row '$RowNum+1'
"id","lastname","firstname","chem_mid","chem_fin","row"
1,"Dawson","Leona",76.5,"B-",2
2,"Jordan","Colin",25.899999999999999,"D",3
3,"Malone","Peter",97.200000000000003,"A+",4
Other built-ins are $Random for random number generation, and $FileId to reference the file being processed.
With raw string data, it is often necessary to extract information based on a a pattern or regular expression. Consider the simple case of extracting a 5 digit zip code from data which looks like this
91101
91101-1234
zipcode: 91101 1234
A unix regular expression of ([0-9]{5}) would easily capture the 5 digit zip code. In this 1 column example the
command would be:
aq_pp -f zip.csv -d s:zip -rx_syntax extended -maprx zip "([0-9]{5})" 'zip=%%1%%'
"zip"
"zip=91101"
"zip=91101"
"zip=91101"
aq_pp has a number of options related to pattern matching. First and formost, it supports regular expressions
and a format developed for another product called RT metrics. Regex is more widespread, but the RT format has
certain advantages for parsing log based data. Full details can be found in the aq_pp
manual.
Back to the example above, we use a global option to specify the type of regex we want to use, and then the
-maprx switch to identify the column to work with and the regex. Finally, the captured value (in this case the
first group, or ‘1’, is mapped to a string using %%1%%. The output string can contain other text.
This example highlights extraction and overwriting a single column. We can also merge regex matching from multiple columns to overwrite or create a new column. For example, we can take our chemistry students and create nicknames for them based on the first three letters of their first name, and last 3 letters of their last name:
aq_pp -f,+1 chemistry.csv -d i:id s:lastname s:firstname f:chem_mid s:chem_fin \
-rx_syntax extended -mapfrx firstname "^(.{3})" -mapfrx lastname "(.{3})$" -mapc s:nickname "%%1%%%%2%%"
"id","lastname","firstname","chem_mid","chem_fin","nickname"
1,"Dawson","Leona",76.5,"B-","Leoson"
2,"Jordan","Colin",25.899999999999999,"D","Coldan"
3,"Malone","Peter",97.200000000000003,"A+","Petone"
Instead of -maprx, we use multiple -mapfrx statements and then -mapc to map the matches to a new nickname
column.
Often it is necessary to use a global variable that is not output as a column but rather acts as an aid to calculation.
Consider the following where we wish to sum a column:
echo "1\n2\n3" | aq_pp -f - -d i:x -var 'i:sum' 0 -evlc 'sum' 'sum+x' -ovar -
"sum"
6
We defined a ‘sum’ global variable and for each validated record we added a value to it. Finally, we use -ovar -
to output our variables to the stdout (instead of the columns).
Filters and if/else statements are used by aq_pp to help clean and process raw data.
For example, if we want to select only those Chemistry students who had a midterm score greater than 50%, we can do:
aq_pp -f,+1 chemistry.csv -d i:id s:lastname s:firstname f:chem_mid s:chem_fin \
-filt 'chem_mid > 50.0'
"id","lastname","firstname","chem_mid","chem_fin"
1,"Dawson","Leona",76.5,"B-"
3,"Malone","Peter",97.200000000000003,"A+"
Another useful option is the -grep flag, which has utility similar to the Unix command of the same name. Given a
file containing a ‘whitelist’ of students, we are asked to select only the matching students from our Chemistry class:
aq_pp -f,+1 chemistry.csv -d i:id s:lastname s:firstname f:chem_mid s:chem_fin \
-grep lastname whitelist.csv X FROM
"id","lastname","firstname","chem_mid","chem_fin"
2,"Jordan","Colin",25.899999999999999,"D"
The format of the grep switch allows the whitelist to contain multiple columns. We select the column to use via
the ‘FROM’ designator. grep also accepts attributes. For instance with grep,ncas, we would have matched
Peter Malone as well in the example above.
A final yet incredibly useful technique for processing your data is to use conditional statements ‘if, else, elif, and endif’
Let’s extend the previous example by boosting the midterm scores of anyone in the whitelist by a factor of 2, and leaving the others untouched:
aq_pp -f,+1 chemistry.csv -d i:id s:lastname s:firstname f:chem_mid s:chem_fin \
-if -grep lastname whitelist.csv X FROM -evlc chem_mid 'chem_mid*0.8' -endif
"id","lastname","firstname","chem_mid","chem_fin"
1,"Dawson","Leona",76.5,"B-"
2,"Jordan","Colin",51.799999999999997,"D"
3,"Malone","Peter",97.200000000000003,"A+"
At the start of this tutorial, we demonstrated how we can use Essentia to select a set of log files and pipe the
contents to the unix wc command. In a similar manner, we can do with with aq_pp,
enabling us to apply more complex ETL operations on a large set of files. In this tutorial we will focus on
‘extract and transform’, and detail how to load the data onto other platforms in other sections.
First, lets switch back to the tutorials/woodworking directory.
For our first example, we are tasked with generating a cleaned version of each file,
and saving it as a comma separated file with bz2 compression:
$ mkdir bz2
$ ess task stream browse 2014-09-01 2014-09-30 "aq_pp -f,+1,eok - -d %cols -notitle | bzip2 - -c > ./bz2/%file.bz2"
We can break down the command (everything within the double quotes) as follows:
aq_pp that the first line should be skipped (+1), that errors are OK (eok)
and that the input is being piped in via stdin.
With eok set, whenever aq_pp sees
an articleID (which we defined as an integer) with a string value, it will reject it. This takes care of the ‘TBD’
entries. Normally aq_pp would halt upon seeing an error. This allows users to use aq_pp as both a data
validator and a data cleaner.aq_pp what the column specification is. We determined this in the previous tutorial where we setup our
datastore and categorized our files. The %cols is a substitution string. Instead of having to enter the
columns each time by hand, Essentia will lookup the column spec from your datastore settings and place it here.
There are several substitution strings that can be used, and they are listed in the section:
Reference Materialbzip utility. We use the substitution string %file to
generate the same filename as the input, except with a bz2 extension.You will probably see a lot of output in the form of:
<stdin>: Bad field value: byte=164727+30 line=5734+1 field=articleID
This is natural, and tells you about problems in your data. The above indicates that many records were rejected
because the ‘articleID’ could not be created as an integer. If you recall from our Tutorial Data documentation,
this was expected and desired behavior, since some articles were dead links that returned a string “TBD”. To turn
off error reporting, add the attribute ,qui right after the eok in the command.
If instead you wish import the “TBD” record anyways, you would need to cast the entire column as a string.
The purchase data needs the articleID corrected for all dates on and after the 15th of September. There are a few ways to achieve this, but the most robust is the following:
1 2 3 4 5 6 7 8 9 | $ ess task stream purchase 2014-09-01 2014-09-30 \
"aq_pp -f,+1,eok,qui - -d %cols \
-evlc is:t 'DateToTime(%date_col,\"%date_fmt\") - DateToTime(\"2014-09-15\",\"Y.m.d\")' \
-if -filt 't>0' \
-evlc articleID 'articleID+1' \
-endif \
-c purchaseDate userID articleID price refID \
-notitle \
| bzip2 - -c > .bz2/%file.bz2"
|
Note
The use of quotations in Unix commands invariably leads to a need to escape characters in order
for them to be recognized.
Line 3 creates a new column ‘t’, which is a signed integer, and it is assigned a value equal to the difference between the time of the current record and the cutoff time of September 15. Positive values of ‘t’ indicate that the record was collected after the 15th.
Line 4 creates a filter condition, which is triggered for all records on or after the 15th.
Line 5 adjusts the articleID to correct for the website error.
Line 6 ends the block
Line 7 specifies the output columns. If not provided, it would also output our new ‘t’ column which we used only for temporary purposes.
We could have just issued 2 Essentia commands, one with dates selected before the 15th and another for dates after. In this case it would have been easy, but there are other scenarios where it becomes more problematic.
This tutorial was designed to teach users how to use aq_pp, but did not compare it against other possible solutions.
To demonstrate the utility of aq_pp, let’s look at the following problem:
We have sales data from a fictional store that caters to international clients. We record the amount spent for each purchase and the currency it was purchased with. We wish to compute the total sales in US Dollars. We have 2 files to process. The first contains the time, currency type, and amount spent, and the second is a lookup table that has the country code and USD exchange rate.
sales data:
transaction_date,currency,amount
2013-08-01T07:50:00,USD,81.39
2013-08-01T08:22:00,USD,47.96
2013-08-01T08:36:00,CAD,62.59
exchange data:
currency,rate
EUR,1.34392
CAD,0.91606
USD,1.00000
Let’s compare 2 solutions against aq_pp. If you wish to execute the commands to see for yourself,
the data are in the tutorial/etl-engine directory.
SQL:
select ROUND(sum(sales.amount*exchange.rate),2) AS total from sales INNER JOIN exchange ON sales.currency = exchange.currency;
SQL is straightforward and generally easy to understand. It will execute this query very quickly, but this overlooks the hassle of actually importing it into the database.
AWK:
awk 'BEGIN {FS=","} NR==1 { next } FNR==NR { a[$1]=$2; next } $2 in a { $2=a[$2]; sum += $2*$3} END {print sum}' exchange.csv sales.csv
AWK is an extremely powerful text processing language, and has been a part of Unix for about 40 years. This legacy means that it is stress tested and has a large user base. But it is also not very user friendly in some circumstances. The language complexity scales with the difficulty of the problem you are trying to solve. Also, referencing the columns by positional identifiers ($1, $2 etc) makes AWK code more challenging to develop and maintain.
AQ_PP:
aq_pp -f,+1 sales.csv -d s:date s:currency f:amount -cmb,+1 exchange.csv s:currency f:rate -var f:sum 0.0 -evlc 'sum' 'sum+(amount*rate)' -ovar -
The AuriQ preprocessor is similar in spirit to AWK, but it simplifies many issues.
We’ll detail the specifics in the rest of the documentation, but even without knowing all of the syntax, the
intent of the command is fairly easy to discern. Instead of positional arguments, columns
are named, therefore making an aq_pp command more human readable.
Additionally, it is very fast, in fact an order of magnitude faster in this example.