Difference between revisions of "Transactions"

m (Tabs.)
m (added ECMA category)
Line 1: Line 1:
 +
[[Category:ECMA]]
 +
 
This is something I've been looking at for the last day or so.  I was really looking for a concurrency solution that solves issues of concurrency from a language standpoint and it seems like SCOOP is currently at a standstill research-wise.
 
This is something I've been looking at for the last day or so.  I was really looking for a concurrency solution that solves issues of concurrency from a language standpoint and it seems like SCOOP is currently at a standstill research-wise.
  

Revision as of 11:18, 6 April 2007


This is something I've been looking at for the last day or so. I was really looking for a concurrency solution that solves issues of concurrency from a language standpoint and it seems like SCOOP is currently at a standstill research-wise.

I'm not an Eiffel compiler writer so I don't know exactly how every piece can be implemented but it seems doable. The proposal relies on transactional memory, probably implemented as software transactional memory right now.

If this proposal is feasible and desirable, it would need to be formalized by someone (not me) in to some language rules.

I'm crossing my fingers for a language solution to concurrency!


A proposal for transaction based concurrency in Eiffel by Colin LeMahieu,

Rationale: The un-computability of determining if a program based on locks will deadlock in general necessitates using alternative forms of concurrency.

Abstract: The Eiffel language already supports the basic idea of transactions in features. A feature either completes in its entirety or an exception is raised. The only shortfall of the current implementation is that any side effects made before the feature completes are not aborted in the transactional sense. If we adapt the semantics of a feature to include the idea of a transactional feature, we will be able to implement concurrency through transactions in the Eiffel language cleanly.

Behavior:

  • A feature can be a transaction
  • All feature calls are asynchronous
  • Transactional features guarantee atomicity and isolation across the execution of the feature
  • When a transactional feature calls another transactional feature, the transactions are composed.
  • The compiler and runtime determine the scheduling either through compiler analysis, runtime statistics, or runtime transactions.
  • When a transactional feature is aborted it is automatically retried until it is not aborted.
  • External features without abort and commit semantics are executed when all transactions have been committed or aborted and executed serially with respect to the system, this is ensured by the runtime.

Syntax: Use the 'transaction' keyword to define a feature as a transaction

feature
 obligatory_transfer_example(account, amount) is
  transaction
   withdraw(account, amount)
   deposit(account, amount)
  end

 withdraw(account, amount: INTEGER) is
  transaction
   account.balance.withdraw(amount)
  end

 deposit(account, amount: INTEGER) is
  transaction
   account.balance.deposit(amount)
  end

Use the 'external', 'abort', and 'commit' keywords to define a transactional external feature. 'External' defines the external transaction functionality. 'Abort' defines functionality to roll back any state changed in 'external'. 'commit' defines functionality to commit the transaction of 'external'.

feature external_feat is
 external
  ""
 abort
  ""
 commit
  ""
 end

Design by contract:

  • Preconditions and postconditions retain ECMA semantics.
  • Features with preconditions and postconditions can only be called from a feature that is a transaction as a validation constraint.
  • If a transactional feature 'f' wants to call feature 'g' with a precondition 'a', 'f' can check 'a' before calling 'g' and is assured that 'a' will be the same once 'g' is being executed by the transactional nature of 'f'; 'a' will either remain valid or 'f' will be aborted and retried.
  • Postcondition wait semantics as introduced with SCOOP do not need to be used as the compiler or runtime will determine the execution schedule of a program.
  • Class invariants determine the consistency of transactions and are checked every time a transaction is committed and after it is aborted

Deferred features:

  • Deferred features can be redefined to be either a transaction or not. If the feature must be transactional, the designer should encapsulate the call to the deferred feature within a transaction feature.

Agents:

  • Agent semantics will not change.

Exceptions:

  • The feature {ANY}.last_exception is obsolete
  • The ANY class contains a new feature last_exceptions and is a collection of exception objects that have been thrown within the context of the current feature.
  • All features within a feature must be executed and either succeed or generate an exception before the enclosing feature can move to the retry clause or propagate an exception.
  • If a feature doesn't handle all exceptions, its transaction is failed(aborted and not retried) with an exception of type ROUTINE_FAILURE
  • When a lower priority feature has a transactional dependency on a higher priority feature that fails, it will fail with an exception of type TRANSACTION_FAILURE which is a subtype of ROUTINE_FAILURE

Priority:

  • If two calls, 'g' and 'h' within a feature 'f' are in conflict, priority is assigned by order of program text and the feature appearing later in the program text will be aborted and retried.
  • If a looping construct is in conflict with earlier iterations of its loop, priority is assigned to earlier iterations of the loop and later iterations will be aborted.
  • If a recursive call is in conflict with itself, priority is assigned to the top level nesting and the deeper nesting calls will be aborted.

Thread library:

  • The thread library is obsolete. All classes should be marked obsolete.

Threads:

  • Explicitly creating threads is redundant.
  • Legacy external features that create threads must block until the threads are complete. Code inheriting from THREAD in the EiffelBase threading library and effecting execute will continue to work simply by using new feature call semantics; the library will need to be changed to allow the new runtime to schedule the feature's execution.

Mutexes:

  • Code using mutexes will continue to work, however they should be replaced with transactional code as soon as possible to avoid the possibility of deadlocks associated with locking.

Legacy:

  • This language change is fully backward compatible with serial code, however, performance benefits would only be seen when using code that does not reference external features that don't have transaction semantics.
  • Building transactions over code that references legacy externals can cause runtime routine failures; the overlying solution would be to eliminate legacy external features.
  • If a legacy external feature is executed within the context of a transaction, the routine is not run and generates an exception of type EXTERNAL_ROUTINE_FAILURE which is a subtype of ROUTINE_FAILURE.
  • External clauses without abort and commit clauses are called 'legacy externals' and are marked as obsolete by compilers supporting legacy externals.
  • Systems that allow legacy external features, without abort and commit clauses, cannot be executed within transactions. The compiler could detect many of these through system analysis however the runtime will need to make explicit checks against legacy external calls to make sure they do not exist within the context of a transaction for things such as agents or inheritance substitution where a feature is redefined to be implemented by a legacy external.
  • Legacy external features are executed by stalling transaction execution, making sure all transactions have committed or aborted, making sure nothing else is executing in the system, and running the legacy external.
  • We're essentially forcing legacy external features to be transactions by making sure no other transactions are currently in progress and there is nothing executing in the system.

Notes:

  • Things that are seemingly serial, such as writing text to the screen in a certain order, are actually dependency relationships; specifically appending to the screen. If there are 2 calls writing to the screen, the second call depends on the result of the first call and the compiler or runtime will either optimize this and have the second call wait, or the second call will abort at runtime until the first call completes. These two could actually run in parallel in the case where the first call writes nothing to the screen.
  • Unlike database transactions, durability is not guaranteed with feature transactions, memory is inherently volatile in the face of system errors.
  • It's impossible for a transactional feature to reliably act on a legacy external feature. During the execution of a transactional feature, any of the actions may be aborted and retried and since legacy external features cannot be aborted, there is no reliable way to undo that action.
  • The asynchronous semantics of feature calls gives the interesting effect of the programmer not being aware of the order in which feature calls are executed. This allows the compiler or runtime to determine the serialization schedule of the program and allows for a level of parallelism not easily thought of by a human.
  • Previous work focused on an idea called 'compensation' which has no real foundation in transactions. Things that are not transactional as currently implemented such as writing files, sending network data, etc, will need classes encapsulating a transaction around that write or send. This may be a type of journaling at the file access or network layer.
  • A feature that encounters exceptions tries to execute all code paths in order to gain a consistent view of the exceptions caused. If this wasn't done, the exception reported as causing the routine to fail might vary depending on the execution plan.
  • Executing all code paths to collect all exceptions might be too restrictive for things such as the main loop of a program. It might not be beneficial to make this change.
  • It might be better to use a keyword to define a feature as non-transactional instead of using a keyword to define a feature as transactional.