Stateful Distributed Applications over Serverless Architectures
Function-as-a-Service (FaaS) enables programmers to develop cloud applications as individual functions that can run and scale independently. However, storage and compute resources must be separated in FaaS. Applications that require support for mutable state and synchronization, such as machine learning and scientific computing, are hard to build.
Crucial is a system to program highly-concurrent stateful applications on serverless architectures. It keeps a simple programming model and allows to port effortlessly multi-threaded algorithms to FaaS.
Crucial associates each thread in a multi-threaded application with a FaaS function invocation. We call this new form of thread cloud thread.
With our abstraction here, a regular Java thread creation such as
Thread t = new Thread(runnable);
can be run as a cloud function with a simple change:
Thread t = new CloudThread(runnable);
To remotely share objects, Crucial maintains them in a separated in-memory layer. Shared objects are made so by annotating their declaration with
@Shared(key="foo")
Foo bar = new Foo();
These objects are wait-free and linearizable, meaning that concurrent method invocations behave as if they were executed by a single thread. Shared objects are simple POJOs with the restriction that they must implement Externalizable and have a no-arg constructor.1
You can find application examples in our examples repository.
Crucial encompasses three main components:
- the FaaS computing layer that runs the cloud threads
- the DSO layer (distributed shared object) that stores the shared objects
- the client application
The client application creates and manages the execution of cloud threads, and all of them access the the shared data using the DSO layer.
This project is built on top of a certain version of Infinispan that requires Java 8. Additionally, AWS Lambda only supports (out-of-the-box) Java 8 runtimes. Therefore, this projects needs to be compiled and executed with Java 8. Newer versions will result in compilation/execution errors.
Applications are built like simple Java code that uses Crucial's abstractions. See this repository for examples.
Crucial's DSO layer (this repository) relies on a client-server model.
It is built atop Infinispan.
To build the project, use maven package
at the root of the project.
Since the applications will need the client, you may want to install it
(maven install
) in your local Maven repository.
The server will be contained in an archive file at server/target/server-*.tar.gz
.
Then, to launch the server, uncompress and run the script server.sh
where
desired. This script binds by default to 0.0.0.0
(when launched with no arguments).
If you want to allow remote access, you will have to modify the script.
Every class used by a client should be known at the server.
This requires to add the appropriates .class
or jar
files to the classpath
of the server. The server can dynamically load new jars located at a specified
directory (by default, it loads from /tmp
).
When running a Crucial DSO cluster in Amazon EC2, you will need to create an
AWS S3 bucket (e.g. crucial-s3ping
) that will be used by the Crucial DSO
server as cluster information store.
Edit the jgroups config file (jgroups-ec2.xml
) and configure the previously
created S3 bucket under the S3_PING
tag. Example:
<S3_PING
location="crucial-s3ping" <!-- bucket name -->
access_key="AKIAAAAAAAAAAAAAAAAA"
secret_access_key="xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
/>
All the servers set up with the previous configuration will automatically form a cluster.
To run in EC2, launch the server with this argument on each VM:
./server.sh -ec2
With this argument, the script automatically binds the DSO server to the public IP address.
If the Crucial DSO cluster is inside a VPC, run the following command instead:
./server.sh -vpc
"On the FaaS Track: Building Stateful Distributed Applications with Serverless Architectures", Daniel Barcelona-Pons, Marc Sánchez-Artigas, Gerard París, Pierre Sutra, Pedro García-López. ACM/IFIP International Middleware Conference, Davis, CA, USA, December 9-13, 2019.
- Objects must be Externalizable in order to support cluster topology changes.