search

spcl / faaskeeper

A fully serverless implementation of the ZooKeeper coordination protocol.
BSD 3-Clause "New" or "Revised" License
18 stars 13 forks source link
aws-lambda faas serverless zookeeper

readme

FaaSKeeper

The ZooKeeper-like serverless coordination service FaaSKeeper.

GH Actions

The main implementation of the serverless FaaSKeeper service. At the moment we support an AWS deployment only. Adding support for other commercial clouds (Azure, GCP) is planned in the future.

To use a deployed FaaSKeeper instance, check our Python client library: spcl/faaskeeper-python.

Paper

When using FaaSKeeper, please cite our upcoming HPDC'24 paper. You can find more details about research work in this paper summary and paper preprint on arXiv.

@inproceedings{copik2024faaskeeper,
  author = {Copik, Marcin and Calotoiu, Alexandru and Zhou, Pengyu and Taranov, Konstantin and Hoefler, Torsten},
  title = {FaaSKeeper: Learning from Building Serverless Services with ZooKeeper as an Example},
  year = {2024},
  series = {HPDC '24},
  booktitle = {Proceedings of the 33rd International Symposium on High-Performance Parallel and Distributed Computing}
}

Dependencies

We use the serverless framework to manage the deployment of serverless functions and cloud resources. The framework with all dependencies and plugins is installed automatically. Configure the provider credentials according to instructions provided in the serverless framework, e.g., instructions for AWS.

Currently, serverless has a bug causing it to generate empty deployment packages (bug 1, bug 2), and the node 15.4.0 is confirmed to work - any newer version is NOT guaranteed to work.

Configuration.

The FaaSKeeper architecture can be customized with the following settings:

Installation & Deployment

To install the local development environment with all necessary packages, please use the install.py script. The script takes one optional argument --venv with a path to the Python and Node.js virtual environment, and the default path is python-venv. Use source {venv-path}/bin/activate to use it.

The deployment with serverless framework is wrapped with a helper executable fk.py. Use the JSON config example in config/user_config.json to change the deployment name and parameters. Use this to deploy the service with all functions, storage, and queue services:

./fk.py deploy service config/user_config_final.json --provider aws --config config/user_config.json

The script will generate a new version of the config in config/user_config_final.json, which includes data needed for clients, such as the location of S3 data bucket - its name is partially randomized to generate unique S3 bucket names. To update functions, they can be redeployed separately:

./fk.py deploy functions --provider aws --config config/user_config.json

The existing deployment can be cleared by removing entire service before redeployment:

./fk.py deploy service --provider aws --clean --config config/user_config.json

To enable verbose debugging of functions, set the flag verbose in the config.

Using CLI

A CLI for FaaSKeeper is available in bin/fkCli.py. It allows to run interactive FaaSKeeper session, and it includes history and command suggestions.

bin/fkCli.py <config-file>
[fk: aws:faaskeeper-dev(CONNECTED) session:f3c1ba70 0] create /root/test1 "test_data" false false

Design

Storage

FaaSKeeper uses two types of storage - system and user. For system, we require strong consistency, as different writer functions must be able to safely modify data in parallel. For user data storage, we require strong consistency as well to prevent stale reads that would violate ZooKeeper consistency principles.

Components

All resources are allocated with the help of Serverless framework. See config/aws.yml for an example on the AWS platform.

Storage

We use key-value tables to store system data, user store, and to function as queus for incoming requests. Furthermore, we allocate persistent object storage buckets to store user data as well.

Functions

We create four basic functions: heartbeat, watch, writer and distributor. Users invoke writer indirectly via a queue or a table to proces a write request, and this function in turn invokes distributor and watch to process updated data nad new watch events. Heartbeat is invoked periodically by the system.

Communication

We use queues to process incoming requests and invoke functions. Queues must uphold FIFO ordering in FaaSKeeper.

Functionalities

Watches

To register a watch, the client performs the following sequence of operations: read node, add watch in storage, read node again. If the node has not been update, then we have inserted the node correctly. This is guaranteed by the fact that system first updates the data, and then it sends watch notifications.

If there's a concurrent update happening in the background, then interleaving between update and watch setting can happen. If the client manages to add notification before the update, it will be notified. However, if the data is updated, then client has no guarantee that it managed to create watch before watch function started delivering notifications - thus, watch creation failed. The watch might have been created correctly, but this case is managed by adding timestamps to let watch function detect when the watch is new enough that it shouldn't be trigerred.

Bad interleaving - entire watch process happens between system writing data and starting notifications. The system should not trigger the watch and retain them for future usage.

GetData

The watch is triggered by set_data and delete operations on the node. To detect if the watch is not set on an older version of the node, the timestamp is compared.

Exists

The watch is triggered by create, set_data, and delete call on the node. To detect if the watch is not set on an older version of the node, the timestamp is compared. When the node does not exist, a "none" timestamp is set - such watch is retained when the update was delete and triggered when the update was create.

GetChildren

Triggered by delete on the node and create and delete on its children. To detect if the watch is not set on an older version of the node, we use the children timestamp.

Development

We use black and flake8 for code linting. Before commiting and pushing changes, please run tools/linting.py functions to verify that there are no issues with your code.

We use Python type hints (PEP, docs) to enhance the readability of our code. When adding new interfaces and types, please use type hints. The linting helper tools/linting.py includes a call to mypy to check for static typing errors.

Authors