Static vs dynamic SAGAs

1
Static vs dynamic SAGAs
Ivan Lanese Computer Science Department University
of Bologna/INRIA Italy
2
Roadmap

• Long running transactions
• A journey in static SAGAs
• A journey in dynamic SAGAs
• Static vs dynamic
• Conclusions

3
Roadmap

• Long running transactions
• A journey in static SAGAs
• A journey in dynamic SAGAs
• Static vs dynamic
• Conclusions

4
Today computing systems

• Computing systems of increasing complexity
• Many components
• Interactions
• Distribution
• Many sources of unreliability
• Other components
• Communication middleware (wireless, )
• Computing systems have to provide reliable
  services to the users
• Important to handle unexpected events

5
Long running transactions

• Unexpected events make activities to abort
• One has to manage the abort to allow the whole
  application to reach a consistent state
• Impossible to have perfect rollback (as for ACID
  transactions)
• Irreversible actions sending of an email
• Penalties booking of an airplane ticket
• Approaches based on long running transactions and
  compensations
• A compensation is executed to take the system to
  a consistent state

6
Formal models for long running transactions

• Different formal models for long running
  transactions
• To clearly specify the expected behavior
• To prove properties of systems
• Interaction based compensations
• Extending name passing calculi such as
  pi-calculus with operators for error handling
• c-join, webp, dcp,
• Compensable flow composition approaches
• Analyzing how compensations of simple activities
  are composed
• cCSP, StAC, SAGAs calculi,
• Need for some order

7
Comparing models

• A difficult task
• Models differ in many aspects
• Atomic activities, communication, state
• Different levels of abstraction
• Different primitives
• We will concentrate on one kind of model, SAGAs,
  and one particular aspect, static vs dynamic
  compensations

8
Static vs dynamic

• Static the possible orders of execution of
  compensations depend only on the structure of the
  term
• Example 1 to compensate PQ execute the
  compensation of Q then the compensation of P
• Example 2 to compensate PQ execute concurrently
  the compensations of P and of Q
• Dynamic the possible orders of execution of
  compensations may depend on runtime information
• Example to compensate PQ execute the
  compensations of P and Q in reverse order of
  completion of P and Q

9
Static vs dynamic in the literature

• Static vs dynamic for interaction based
  compensations at ESOP 2010 (with Vaz, Ferreira)
• Classic SAGAs calculi (Bruni, Melgratti,
  Montanari, POPL 2005) are static
• Dynamic SAGAs calculus at SEFM 2009 (with
  Zavattaro)
• Which are the relations between static and
  dynamic SAGAs calculi?

10
Something more on SAGAs calculi

• The basic building blocks are compensable actions
  AB
• Execute activity A, if the SAGA aborts execute
  activity B as compensation for A
• SAGAs can be composed
• sequence PQ
• parallel PQ
• nested P

11
Which semantics for SAGAs?

• Many possible choices
• Static vs dynamic
• Interruption vs no interruption
• Centralized vs distributed compensations
• We consider the semantics with interruption and
  centralized compensation
• Interruption for avoiding unnecessary
  computations
• Semantics of distributed compensations
  unrealistic for real systems

12
Nesting

• Useful for modeling complex systems andfor
  refinement
• Allows to see a SAGA as an activity
• Present in the original proposal (POPL 2005)
• Never defined for static SAGAs with interruption
  and centralized compensation nor for dynamic
  SAGAs
• Second aim of the paper extending the two
  approaches with nesting
• Not a trivial issue

13
Roadmap

• Long running transactions
• A journey in static SAGAs
• A journey in dynamic SAGAs
• Static vs dynamic
• Conclusions

14
Static SAGA semantics

• Big-step semantics
• G is an environment describing basic activities
• The final outcome of a SAGA can be
• success
• abort
• failure, if a compensation aborts
• Observation a the activities that have been
  successfully executed
• Compensation ß starting compensation
• Compensation ß1 final compensation

15
Sample static rule

• A rule for sequential composition
• Observations can also include parallel
  composition
• We refer to the paper for the whole semantics

16
Static SAGA semantics

• Activities are executed and compensations stored
  for later retrieval
• Executed in case of failure by the innermost SAGA
  (centralized compensation)
• If a branch aborts and/or fails, other branches
  should be notified
• Abort makes the other branches to compensate
• Fail is catastrophic and blocks all the
  activities
• Can only occur with nesting

17
Branch outcome notification

• Notification blocks execution of other branches
• We need to model incomplete executions
• abort because of external abort (successful
  compensation)
• failure because of external failure (no
  compensation)
• failure because of external abort and failure
  of the compensation
• Parallel composition rules describe the allowed
  combinations
• Notifications should be also propagated to
  subSAGAs

18
Roadmap

• Long running transactions
• A journey in static SAGAs
• A journey in dynamic SAGAs
• Static vs dynamic
• Conclusions

19
Dynamic SAGA semantics

• Small-step semantics
• The final result of a dynamic SAGA can be
  either an intermidiate process P or a final
  outcome , or
• Observation a can be an activity or empty
• A SAGA gives rise to computations to reach a
  final outcome

20
Sample dynamic rules

• A few rules for sequential composition
• Allow to compute inside P and to complete P
  execution successfully
• Again, we refer to the paper for the whole
  semantics

21
Dynamic SAGA execution

• When a subSAGA executes it produces some items of
  compensation
• Compensations of subSAGAs and of the main SAGA
  should not mix
• Auxiliary syntax is needed
• Running SAGA P,ß
• SAGA executing P with stored compensation ß

22
Compensation execution

• Running compensations should not be interrupted
  by external aborts
• Should execute in a protected way
• Auxiliary syntax is needed again
• ß is a running compensation
• Two possible causes for compensation execution
• Internal if compensation is successful then the
  SAGA is successful
• External the SAGA can not succeed anyway, abort
  is at the upper level
• Two forms of running compensations ß and ß

23
Roadmap

• Long running transactions
• A journey in static SAGAs
• A journey in dynamic SAGAs
• Static vs dynamic
• Conclusions

24
Static vs dynamic SAGAs

• Two different intuitions about compensation order
• Big-step vs small-step semantics
• Which is the relation between them?
• Are the two definitions coherent?
• Good hint of correctness

25
Towards the correspondance

• Static big-steps correspond to sets of possible
  dynamic computations
• Big-step with label (AB)C corresponds to the
  set of computations with sequences of labels
  A,B,C or B,A,C
• Steps with empty labels are deleted

26
From dynamic to static

• For each complete dynamic computation there is a
  static big-step with a compatible label and with
  the same outcome
• Proof by induction on the SAGA structure
• Auxiliary results to relate partial computations
  to big-steps with outcomes , or

27
From static to dynamic

• For each static big-step there is a dynamic
  computation with a compatible label and with the
  same outcome
• There is not one such computation for each
  possible interleaving of the parallel
  observations
• Dynamic SAGAs have more constraints on order of
  execution of actions
• In ABCD compensations B and D can be executed
  in any order in static SAGAs
• In dynamic SAGAs if A is executed before C then D
  has to be executed before B
• ACBD is valid for static SAGAs but not for
  dynamic SAGAs

28
Comparison outcomes

• Static and dynamic SAGAs are strongly related
• Static SAGAs allow for more nondeterminism in the
  order of compensation of parallel actions
• The strong relation is a good hint about the
  correctness of the formalization

29
Roadmap

• Long running transactions
• A journey in static SAGAs
• A journey in dynamic SAGAs
• Static vs dynamic
• Conclusions

30
Results

• Formalization of
• Nested static SAGAs with interruption and
  centralized compensations
• Nested dynamic SAGAs
• Proved a strong relationship between the two
  models

31
Future work

• Fully analyze dynamic SAGAs
• Define a realistic semantics for SAGAs with
  interruption and distributed compensations (WADT
  2010)
• Continue to study the relationships between
  different approaches to long running transactions
• Still lot of work to do
• Hierarchical vs flat

32
The end

• Thanks!

• Questions?