Folks, It is not lightly that I've made the decision to write an alternative Actor library and move the Lift code base from the Scala Actors to Lift Actors (working name). I want to spend a little time talking about the steps that led to the decision as well as the impact that it will have on Lift code.
Since November, I've been chasing a series of memory leaks in the Actor library. Philipp Haller from EPFL has been responsive in addressing the individual memory leaks, but the issue seems to be one of whack-a-mole... each time one memory leak is fixed, another one appears. Further, the existing Actor architecture does not lend itself to the kind of Actor usage cycle that we find in Lift apps. Specifically: - Lift creates/destroys an Actor for each Comet request. This rapid creation/destruction of Actors caused memory back-ups, and the existing Actor code seems to be oriented to long running Actors rather than Actors with Object-length lifespans. - The FJ libraries used for Actor scheduling have problems on multi-core machines and are also a source of memory retention issues. - Replacing the FJ libraries with a scheduler based on java.util.concurrent exposes race/deadlock conditions related to the fact that some parts of the Actor processing (e.g., testing mailbox items against partial functions while the Actor itself is synchronized) - The Actors require external threads to function and it's not possible to create external threads in the Google App Engine (making Actor-based functionality including CometActors non-functioning in GAE apps) - Actors are fragile when exceptions are thrown - Actors have running and not running states (as compared with objects which can always respond to message sends). In practice, managing the running and not running states is as hard as managing memory in C. - There are hidden actors associated with each thread which display the above fragility and state management issues - And as a practical matter, I've got a couple of applications that are going into production over the next few weeks and cannot wait for the various fixes to make it into Scala 2.8 and the hacks and work-arounds that I've done to the 2.7.4 Actor libraries became too complex for my comfort. I have written a simple Actor class that is focused on message sending and processing of messages asynchronously. This means there's a single operation that you can perform on Actors, the message send operation. Actors can be specicialized (they only access messages of a certain type). In order to receive a response from an Actor, you can pass in a Future as part of the message and that Future may be satisfied asynchronously. This means that a sender of a message need not be an Actor and that the Actor recipient of a message cannot determine the sender of a message. Actors have two bits of internal state: a mailbox and a flag indicating that the Actor is currently processing messages in its mailbox. The amount of synchronization of Actors is minimal (on inserting messages into the mailbox, on removing messages from the mailbox, and on changing state to/from "processing messages".) An Actor instance must provide a messageHandler method which returns a PartialFunction that is used to pattern match against the messages in the mailbox. The instance may also provide an optional exception handler that is called if an Exception is thrown during the handling of a message. The Actor is guaranteed to only be processing one message at a time and the Actor is guaranteed not to be in a monitor (synchronized) during the processing of messages. An Actor is guaranteed to maintain the order of the messages in its mailbox, however, messages that do not currently match the messageHandler will be retained in the order that they were received in the event that the messageHandler changes and they can be processed. The Lift Actors will, by default, use the java.util.concurrent library for thread pooling, although I have worked out a mechanism for thread-piggy-backing such that if the Actors are running in GAE, they need not use any additional thread (this will enable Lift's comet support in GAE.) There will also be a scheduler (much like the existing ActorPing) which will send a message to an Actor at some time in the future (and on GAE, this scheduler, the Pinger, will not require a separate thread.) The changes that you will have to make to your applications are minimal. Actors will no longer have start(), exit(), or link() methods. Actors will always process messages in their mailbox and will be removed from the system by the JVM's garbage collector. Calls to !? will be replaced by calls to ! with a Future as a parameter to the message. Calls to ActorPing will be replaced by calls to Pinger. You can continue to mix Scala's Actors and Lift's Actors in an application, although Lift's work-arounds to the Scala Actor memory retention issues and scheduling issues will not be turned on by default (they will still be available in the Lift codebase if you're using Scala 2.7.4 and need the work-arounds.) I am happy to share the Lift Actor code with EPFL and if it makes it into the Scala distribution as SimpleActors or something similar, I'm totally cool with that. I'm not interested in owning or maintaining an Actor library. I am however, dedicated to making sure that Lift apps can run in production for months (or even years) without retaining memory or having other problems that can impact the stability of applications. I will have a branch committed up on GitHub tomorrow with Lift ported to the new Actor library. If you have any questions, please let me know. Thanks, David PS -- ESME people, I'll roll these changes into ESME next week -- Lift, the simply functional web framework http://liftweb.net Beginning Scala http://www.apress.com/book/view/1430219890 Follow me: http://twitter.com/dpp Git some: http://github.com/dpp
