Data Model Analysis and Verification Tevfik Bultan University

Data Model Analysis and Verification Tevfik Bultan University of California Santa Barbara Joint work with Jaideep Nijjar and Ivan Bocic 1

Acknowledgement: NSF Support 2

Web Application Dependability 3

Web Application Dependability 4

Web Application Dependability President Obama: “I want to go in and fix myself, but I don't write code" 5

Web Application Dependability TRACKS: A todo list application http: //getontracks. org/ Context Recurring Todo Feed the Dog EDIT 6

Web Application Dependability TRACKS • Todos may be interdependent (if a Todo must be completed before another) • Todo cycles. . ? 7

Web Application Architecture Data Model OOP ORM Rel Db RESTful Controller View • Model View Controller (MVC) pattern: Ruby on Rails, Zend for PHP, Cake. PHP, Struts for Java, Django for Python, … • Object Relational Mapping (ORM) Active. Record, Hibernate, … 8

An Example Rails Data Model Static Data Model class User < Active. Record: : Base has_many : todos Data Model Updates: Actions has_many : projects end class Project < Active. Record: : Base class Projects. Controller < Application. Controller def destroy belongs_to : user @project = Project. find(params[: project_id]) has_many : todos @project. notes. each do |note| has_many : notes note. delete end class Todo < Active. Record: : Base @project. delete belongs_to : user belongs_to : project end respond_to(. . . ) end class Note < Active. Record: : Base belongs_to : project end 9

Static Data Model • Active. Record class declarations • sets of objects • Active. Record association declarations • has_one, has_many, belongs_to, has_and_belongs_to_many • Association declarations can be used to declare three basic types of relations between classes • one-to-one • one-to-many • many-to-many 10

Extensions to Static Data Model • : through Option • To express relations which are composition of other relations • : conditions Option • To relate a subset of objects to another class • : polymorphic Option • To express polymorphic relationships • : dependent Option • On delete, this option expresses whether to delete the associated objects or not 11

The : through Option class User < Active. Record: : Base has_one : profile has_many : photos, : through => : profile end class Profile < Active. Record: : Base belongs_to : user has_many : photos end class Photo < Active. Record: : Base belongs_to : profile Profile end 1 1 * 1 User 1 * Photo 12

The : dependent Option class User < Active. Record: : Base has_one : profile, : dependent => : destroy end class Profile < Active. Record: : Base belongs_to : user has_many : photos, : dependent => : destroy end User 1 1 Profile 1 * Photo • : delete directly delete the associated objects without looking at its dependencies • : destroy first checks whether the associated objects themselves have associations with the : dependent option set 13

Data Model Verification • Formalize the static data model as • A set of classes • A set of relations between those classes • A set of constraints on the relations that are imposed by the association declarations • Given a formal data model we can automatically check if a given property holds for the data model • Automated verification determines: Do the constraints of the data model imply the property? 14

i. Da. Ver (Integrated Data Model Verifier) Bound Model Extraction n bound Property formal data model + property BOUNDED VERIFICATION formula Alloy Analyzer instance or unsat Active. Record UNBOUNDED VERIFICATION Alloy Encoder SMT-LIB Encoder Results Interpreter formula SMT Solver Interpreter instance or unsat or unknown 15 Property Verified Property Failed + Counterexample Unknown

How Automated is Automated Verification? • All except one step: Property specification • Example: It is possible to have a User who does not have any Photos. • In Alloy: pred prop { all s: Pre. State | some u: User | all p: Photo | (p not in (s. photo_user). u) } • In SMT-LIB: (assert (exists ((a Polymorphic. Class)) (forall ((p Photo)) (and (is. User a) (not (= p (a user_photo p)))) • Can we make it easier? 16

Property Templates • Property templates for property specification • Language-neutral • Do not require familiarity with SMT-LIB and Alloy • Example property template: • no. Orphans[class. A, class. B] • To check that deleting an object from class. A does not cause related objects in class. B to be orphaned • Easily rerun tool and switch the verification technique, without having to rewrite the property • We developed seven property templates for the most common data model properties 17

Case Study with i. Da. Ver • Lovd. By. Less, a social networking application • LOC: 3787 • Number Active Record files: 13 • i. Daver input: • Path of the directory containing the Active Record files • Name of the file containing the properties to check (Expressed using property templates!) • Verification technique 18

Case Study • Check: some. Unrelated[Forum. Topic, Forum. Post] • Solver: Z 3 19 + Unbounded verification + Sample instance − May report unknown or timeout

Case Study • Check: delete. Propagates[Profile, Photo] • Solver: Alloy 20

Case Study • Check: delete. Propagates[Profile, Photo] • Solver: Alloy + Counterexample data model instance + Always returns a result (for small domains) − Bounded − Slower 21

Can We Do More? 22

Automatic Property Inference • Automatically infer properties based on data model schema • Data model schema: A directed, annotated graph that represents the relations • Look for patterns in the data model schema and infer a property if a pattern that corresponds to a property appears . . . • For example, orphan prevention 0 1 . . . n 23

Automatic Data Model Repair . . . • no. Orphans(X, Y) property failing means deleting an object from class X creates an orphan chain that starts with associated object in class Y • Repair: Set : dependent option to : destroy on association declaration in class X and on remaining relations in the chain that starts with class Y X Y . . . Set : dependent => : destroy on all relations in chain N 24

Summary Formal Data Model + Properties Verification Results Active Records Model Extraction Property Inference Verification for failing properties Data Model Repair 25

Experimental Results Application Lovd. By. Less Substruct Tracks Fat. Free. CRM OSR TOTAL Property Type # Inferred # Timeout # Failed delete. Propagates 13 0 10 no. Orphans 0 0 0 transitive 1 0 1 delete. Propagates 27 0 16 no. Orphans 2 0 1 transitive 4 0 4 delete. Propagates 15 0 6 no. Orphans 1 0 1 transitive 12 0 12 delete. Propagates 32 1 19 no. Orphans 5 0 0 transitive 6 2 6 delete. Propagates 19 0 12 no. Orphans 1 0 1 transitive 7 0 7 145 3 96 26

# Data Model & Application Errors # Data Model Errors # Failures Due to Rails Limitations # False Positives delete. Propagates 1 9 0 0 no. Orphans 0 0 transitive 0 0 0 1 delete. Propagates 1 3 5 7 no. Orphans 0 1 0 0 transitive 0 1 0 3 delete. Propagates 1 1 3 1 no. Orphans 0 0 0 1 transitive 0 7 0 5 delete. Propagates 0 18 1 0 no. Orphans 0 0 transitive 0 0 0 6 delete. Propagates 0 12 0 0 no. Orphans 0 1 0 0 transitive 0 7 0 0 TOTAL 3 60 9 28 Property Type 27

What About Data Model Actions? Static Data Model class User < Active. Record: : Base has_many : todos Data Model Updates: Actions has_many : projects end class Project < Active. Record: : Base class Projects. Controller < Application. Controller def destroy belongs_to : user @project = Project. find(params[: project_id]) has_many : todos @project. notes. each do |note| has_many : notes note. delete end class Todo < Active. Record: : Base @project. delete belongs_to : user belongs_to : project end respond_to(. . . ) end class Note < Active. Record: : Base belongs_to : project end 28

Verification of Data Model Actions 29 29

Abstract Data Stores Rails Abstract Data Store class User has_many : todos has_many : projects end class User { 0+ Todo todos inverseof user 0+ Project projects inverseof user } class Project belongs_to : user has_many : todos has_many : notes end class Project { 0. . 1 User user 0+ Todo todos inverseof project 0+ Note notes inverseof project } class Todo belongs_to : user belongs_to : project end class Todo { 0. . 1 User user 0. . 1 Project project } class Note belongs_to : project end class Note { 0. . 1 Project project } 30

Abstract Data Stores def project_destroy @project = Project. find( params[: project_id]) @project. notes. each do |note| note. delete end @project. delete respond_to(. . . ) end action project_destroy() { at_project = oneof(allof(Project)) foreach note: at_project. notes { delete note } delete at_project invariant(forall{ |project| !project. user. empty? }) invariant(forall{ |user| user. projects. todos. include? (us er) }) forall(Project project: not empty(project. user) ) forall(User user: user in user. projects. todos. users ) } Our library allows developers to specify invariants in native Ruby 31

Extraction is hard for actions • Dynamic type system • Metaprogramming • Eval • Ghost Methods such as: User. find_by_name(‘Rob’) Observations • The schema is static • Action declarations are static • ORM classes and methods do not change their semantic during execution • even if the implementation code is generated dynamically 32

Extraction via Instrumented Execution • Boot-up the Rails runtime in a simulated environment • Without opening sockets or connecting to the database • Prepare action methods for extraction • ORM operations will record their invocation instead of communicating with the database • Method calls propagate instrumentation just before execution • Extraction is path insensitive, executing both branches subsequently • Trigger an HTTP request that triggers an action 33

Verification via Translation to FOL • A predicate is generated for each class and association User(o) means that o is an instance of User Project_user(t) means that t represents an association between a Project object and User object • Type system constraints become axioms ∀u: User(u) → ¬(Project(u) ∨ Todo(u). . . ) • Cardinality of associations is expressed through axioms eg. 0. . 1: ∀t 1, t 2: (Project_user(t 1) ∧ Project_user(t 2) ∧ Project_user_lhs(t 1) = Project_user_lhs(t 2)) → Project_user_rhs(t 1) = Project_user_rhs(t 2) 34

Translation of Statements to FOL • An action is a sequential composition of statements. • Statements • A state is represented with a predicate denoting all entities that exist in a state • A statement is a migration between states e. g. , a create Note statement: ¬pre_state(newly_created()) ¬∃t: post_state(t) ∧ Note_project_lhs(t) = newly_created() ∀o: (post_state(o) ↔ (pre_state(o) ∨ o = newly_created())) 35

Inductive Verification • Inv(s) is a formula denoting that all invariants hold in state s • Action(s, s’) is a formula denoting that the action may transition from state s to state s’ Check if: ∀s, s’: Inv(s) ∧ Action(s, s’) → Inv(s’) 36

Translation of Loops to FOL • We only support For. Each loops (for now) • They correspond to universal quantification • Statements can execute multiple times in loops • Contexts to differentiate iterations • Ordering of iterations • Iteration interdependence 37

Example - Loop Execution pre-state class Posts. Controller def destroy_tags. . . posts = Post. where(id: params[: post_ids]). . . posts. each do |p| p. tags. destroy_all! end. . . end post-state 38

Loop Verification Is Difficult ● ● Iteration interdependency Unknown bound on the number of iterations Standard techniques for loop verification ● Bounded unrolling ● Loop invariants 39

Coexecution ● ● ● Capture the effects (the delta) of each iteration’s execution directly from the pre-state Combine these deltas using the delta union operation Execute the union from the pre-state to reach the post-state 40

Example - Coexecution prestate post-state 41

Coexecutability Condition When are sequential execution and coexecution equivalent? class Posts. Controller def destroy_tags. . . posts = Post. where(id: params[: post_ids]). . . posts. each do |p| p. tags. destroy_all! end. . . end ? = 42

Coexecutability Condition ● Infer the create, delete and read sets of iterations ○ ○ ● Create and delete sets are easily inferred Read sets, not as easy Check if create, delete and read operations may conflict across iterations on the same object ○ ○ read/delete is a conflict read/read is not a conflict delete/delete is not a conflict. . . 43

Experiments Experimented on 3 open source Rails applications • Fat. Free. CRM, Tracks, Kandan • 272 actions, 23 invariants Identified 4 bugs • Reported to original developers • All immediately confirmed and, since, fixed • Missed by previous verification efforts on these applications 44

Experiments 45 45

Experiments on Coexecutability ● All these loops are coexecutable 46

Experiments on Coexecutability Performance: ● At least an order of magnitude 45% of the time ● Four orders of magnitude in 13% of cases 47

Publications • Jaideep Nijjar and Tevfik Bultan. Bounded Verification of Ruby on Rails Data Models. In Proc. International Symposium on Software Testing and Analysis (ISSTA), pages 67 -77, 2011. • Jaideep Nijjar and Tevfik Bultan. Unbounded Data Model Verification Using SMT Solvers. In Proc. 27 th IEEE/ACM Int. Conf. Automated Software Engineering (ASE), pages 210 -219, 2012. • Jaideep Nijjar, Ivan Bocic and Tevfik Bultan. An Integrated Data Model Verifier with Property Templates. In Proc. 1 st FME Workshop on Formal Methods in Software Engineering (Formali. SE 2013). • Jaideep Nijjar and Tevfik Bultan. Data Model Property Inference and Repair. In Proc. International Symposium on Software Testing and Analysis (ISSTA), pages 202 -212, 2013. • Ivan Bocic, and Tevfik Bultan. Inductive Verification of Data Model Invariants for Web Applications. In Proc. International Conference on Software Engineering (ICSE), pages 620 -631, 2014. 48