failSafer.guru
It started with the concept of entropy that I learned in machine learning courses. This is one of the several metric we use to measure similarity. I new as well that physic was using the same word. Entropy, the second
Fail is limitless and omnipresent. We can see example of fails everywhere. Day to day failures, missing something, up to the fail of an airplane that crashes, or fail your life. There is a wide variety of fails.
L’échec nécessite deux éléments : un dessein et un aboutissement qui ne satisfait pas les attentes du point de départ. L’absence du premier élément n’implique pas celle de l’échec.
Pattern: Usually a fail is applied to something. A project, an object (airplaine, ), one's person action, human (a state like greece) we could say that all thoses are systems.
How a system (an airplaine, a project) can collapse? How a state can collapses? (Greece) How a system that human build can collapses?
All those concepts in some way are fails/failures. Is there a stretegy of failure? What are the common pattern of failures? failure theory Tout ces concepts tourne autour du fails/failures
-Risque, -Biais, -Entropie, le bordel , quantifier le bordel. How can one's mesure mess, disorder Mathieu cherche desesperement a mettre de l'ordre dans le chaos qui l'entoure.
-system failure (system Redundancy (engineering)), fail fast -systems designs -second law of thermodynamic -Software entropy -principe de parcimonie -hazard analysis -risk management (decrease the probability of failure or uncertainty)
Aspect exemple monde reel -un repertoire des fails reel: la resistance du reel (-superfail france culture, parce que c'est pas si facile de reussir a echoue) -Danger dans le ciel -blue scren of death -uber -la betise, l'echec de l'intelligence. -Si on a pas rater au moins trois start up dans la silicon valley, on est pas pris au serieux. -Why cutting cost is expensive? Boeing 737 MAX.
Aspect Soft: -Biais -Philosophie de l'echec. Les modalite de l'echec, les facteurs d'echec. La resistance du reel. -Droit. GDPR, Data privacy (NYT data privacy project) -Machine learning desiderata
Aspect pratique coding: -R package failSafeR -R inferno (because I need to master R)
Aspect pratique monde reel: -Value at risk
HOW TO SOLVE THIS? Aspect Statistique pure: -Mouvement brownien. -information de fisher -entropie -perceptron -Self Organizing Map (competitive learning) -Ensemble learning (http://www.scholarpedia.org/article/Ensemble_learning)
A data driven approach of debugging.
Les genes donnent des caracteristiques. Code genetic: lepigenetic, les genes s'expriment non pas seulement parce qu'ils sont dans ton code, mais reagissent aussi a l'environnement/ s'exprime a cause de l'env.
Par exemple si tu vis dans un environenement tres pollue, certains gene s'active et certains gene se desactive. Le code est la, mais les programmes s'allument et s'eteignent.
Chaque code est specifique, pourtant les erreurs leve sont communes a tout les codes. Pourquoi ne pas avoir un debuggeur qui s'adapte a sont environement L'ADN a une forme spaciale, et cette forme change la maniere dont le code est interprete dans le milieu.
Debuggeur qui s'adapte au milieu dans lequel il evolue.
Un debuggeur qui repond a l'environnement de la fonction.
TODO: - Creer une class, object pour recevoir et gerer une debug collection.
- add ML algo clustering Kmeans for categorical data
- detect sign of numeric input
- detect unusual value for factor input or character input
- dim of data frames, could detect empty dataset
- length of object
- give a likelyhood confidence
- track [[ functions : being warn if trying to access an element that doesnot exist. ## has_name
- mastering advanced topic for R
- Possibility to retrieve quickly the input values of the function to try the function and troubleshoot the issue.
- possibility to add a function to track through json file. so the actual code is never touched.
- Create a meaningfull example of a silent error
- Concept of silent errors,
- Concept de contextual debugging.
.onLoad package monitor functions in mc_register
Avoir un comportement different si japelle une fonction de class meta_collection, cad quon doit trigger la sauvegarde du current env.
A class got the attributes: -a function: the function that is monitored -a data frame: where we sotre the data collected for that specific function
-
You add in a json file the function you want to monitor. What is happening in the background? Your functon is slightly modified to be able to track its environment.
-
You run that function a couple of times: what is happening in th bg? It saves all the variables that lives in the function current environment.
-
Let say you run that function, and an error is raised. What is happening in th bg? We look at the stacktrace and filter the monitored functions.
-
It displays the unusual values that showed up among thoses monitored functions.
-
enhance a function
- add patch to the function to be able to track its environment
- wrap the function in a tryCatchish function to be able to subscribe to error event
-
Storage
-
process the stored data to have it in proper way to be consumed in algo
-
extract the information
-
extract meta data
-
extract class (from the actual data.frame to class data.frame, or from 2 to numeric, or from 'Hello World' to character)
-
extract numeric sign (from -2 to negative, 4 to positive)
-
extract numeric length (from c(3,4) to 2, from 4 to 1)
-
extract numeric likely range value with variance, volatility
-
extract numeric type (integer, double)
-
extract numeric distinct values, if we detect that the number of different values is small
-
extract character distinct values
-
extract character length
-
aggregate meta data
- contingency table
-
get the actual best value
-
beta
La vrai intelligence elle reside dans la capacite a voir, quand tu creer tes fonctions meta, voir comment s'articule lensemble des possibilite et etre capable de prendre la plus adapte, la moins couteuse. parce que ta une vue globale de ce qu'il se passe
It is an error up the stack that will not trigger any error/warning by the system because at the moment it was processed it was not an error, because you have to know the constraints down the stack.
how come?
I am implicitly creating a dependency between two objects.
Either I can explicitly tell the system what is the dependance by for example add an assertthat. It is kind of a foolproof/ defensive design.
drawbacks: 1.The problem here is that it is really hard to be aware of all the dependency we are creating between the different pieces of the system. drawback 2. it does add lines of codes drawback 3. when your code change, you have got to change your assertthat.
Because we are just human, that is where come the idee to p**p your mind! with failSafeR.
if(myword == 'this') { do whatever you want }
this bit of code works only for one specific string. It means that it wont work for all the other possibilities. In other words, there is millions of possibilities to mess up and only one way to have it working.
There is clearly an imbalanced. it is the same in internet securities. It is very hard to secure a system because there is thousands way to get in that we didnot though about.
The system is not foolproof.
The idea behind fail safer is to grab the path we know already worked and thus being able to pinpoint when the system try an unusual path or a path that it has never seen before
A function that call another function (and recursively)
Being able to trigger an event when trying to access a list item that doesnot exist.
function [[ we have got you covered!
Sequential pattern mining.
Example:
How is it possible to handle this?
Knowing within a function the space of acceptable values.
Knowing between functions the expected flow of values.
TO UNDERSTAND: the callstack what are the different items, what does the numbers reffers to
Fit a beta distribution.
l'interdependance entre parametres. Mis en avant par la clusterisation.
- univariate variable to be tested (x, y, ...)
- properties of variables (class, length, dim, )
- common properties (class, mode)
- inner/specific properties of variables (most often based on the class) (dim dataframes, sign numerical, distinct characters)
- Be able to give a likelihood/probablity
- Plot the success and failing space. There is a space where the function actually works. Cluster analysis.
- Quand une variable est reutilise dans la fonction, il faut pouvoir reinitialise la colonne.
Ce package peut s'appliquer a nimporte quel system qui genere un flux de donnee. Poser une API dessus permet de le monetiser. (du code, des logs machines, ...)
Peut s'appliquer a nimporte quel langage, puisque traite de la data, bak end is R.
On faisant touner ma function. je donne de linformation. et lapprentissage il se fait la.
having only NA in a column of integer make that bit of code crashing. because only NA in column make the column as character, where the system was expecting numeric
where m is a list of calls $orders sum(orders, na.rm = T)
$revenue sum(revenue, na.rm = T)
data %>% group_by_(.dots = dots) %>% summarise(!!!m[1:7])
injection de dependances le code nest pas responsable dinstancier les nouveau object
SSL certificate https://www.digitalocean.com/community/tutorials/how-to-secure-nginx-with-let-s-encrypt-on-ubuntu-16-04
Docker https://docker-curriculum.com/ https://docs.docker.com/storage/volumes/
S3 Classes https://www.cyclismo.org/tutorial/R/s3Classes.html
Logisitic regression https://www.math.univ-toulouse.fr/~besse/Wikistat/pdf/st-m-app-rlogit.pdf
Fidel Castro https://fr.wikipedia.org/wiki/Fidel_Castro
Platon le banquet. https://papyrus.bib.umontreal.ca/xmlui/bitstream/handle/1866/3551/Fortin_Jerome_2009_memoire.pdf?sequence=4&isAllowed=y
New York https://www.youtube.com/watch?v=igG6xDyCZFA
Deep learning with R https://manning-content.s3.amazonaws.com/download/9/9a3b0d8-e651-4239-8c4f-94267be64fee/SampleCh03.pdf
Logique https://www.editions-ellipses.fr/PDF/9782340014787_tdm.pdf
Une remise en question de quelaue chose qui parait evident au yeux de tout le monde.
https://medium.com/civis-analytics/experimenting-with-pyspark-to-match-large-data-sources-abf7ec44b8c2 https://medium.com/@suci/running-pyspark-on-jupyter-notebook-with-docker-602b18ac4494 https://spark.apache.org/docs/latest/quick-start.html
https://www.youtube.com/watch?v=DV44pB9fpf8&list=LLxCFwDyjOTXOfOIOTaVv1tw&index=2