Recap of ICDCS 2023 and a workshop where we presented our early-stage work on ScalarDL

From July 18 to 21, we attended the ICDCS 2023 conference and presented our early-stage work regarding ScalarDL in a workshop held before the conference. In this blog post, I will briefly introduce the paper that we presented and share an overview of the conference, research trends, and a few interesting papers.

About the ICDCS

The International Conference on Distributed Computing Systems (ICDCS), which has a long history of 43 years, is one of the top-tier international conferences for distributed systems and is sponsored by IEEE. In recent conferences, researchers and practitioners have discussed cutting-edge ideas on various distributed-systems topics, including cloud and edge computing, distributed systems for AI and machine learning, blockchain, and security. Attending this conference was a first for me, as I’m active mainly in the database systems community. This year’s conference was held in Hong Kong and consisted of several workshops on the first day, followed by three days for the main conference.

Presentation about lazy Byzantine fault detection (BFD)

At the workshop, I presented the paper titled “The Case for Lazy Byzantine Fault Detection for Transactional Database Systems.” We are currently working on a new approach for Byzantine fault detection (BFD) as an evolved form of ScalarDL, one of our core products, which we presented at the VLDB 2022 conference last year. This paper discusses the initial study, prototype, and evaluation.

In a bit more detail, Byzantine fault detection in ScalarDL currently requires synchronous agreement (i.e., verifies if no fault, including data tampering, occurs in a real-time manner) between the two database replicas during transaction execution. This synchronous protocol has an impact on transaction performance and computational resources. So, we proposed lazy BFD, which only logs some necessary information at the execution time and then verifies transactions lazily. Although lazy BFD will not be able to detect tampering or other faults in real time, we expect it to provide better transaction performance and resource management, particularly if users need to verify data only periodically and don’t need real-time detection (e.g., before a quarterly audit).

Research trends in distributed systems

I will share an overview of the trends in this year’s ICDCS and briefly introduce a few interesting papers related to my research field.

This year, 439 papers were submitted, and only 83 papers were accepted (an 18.9% acceptance ratio). By category, AI for Distributed Systems and Edge Computing had the most acceptances, with 12 papers each, followed by Blockchain, with 11 papers accepted.

AI and machine-learning research actually existed in more than the 12 papers accepted. For example, there were also several machine-learning papers in the Edge Computing and Security categories (e.g., those related to federated learning and adversary attacks). So, I felt that about one-third of the papers were AI and machine-learning related. Although I’m not an expert in machine learning, this situation made me think about what we can and should do to ensure data authenticity and privacy in machine learning as a company working to help ensure absolute data reliability.

Blockchain was also actively discussed. What I found unique about the blockchain field was that almost all papers were for public or permissionless blockchains. This was a bit surprising, as I have felt that private or permissioned blockchains are more common in the database community, where I often participate. However, the papers were exciting because there were many similarities in terms of challenges and opportunities.

Let me pick up an interesting paper related to the parallel execution of smart contracts[1]. Basically, blockchain systems decide a total order of transactions, and then each node in the system needs to deterministically execute the transactions (i.e., smart contracts) in that order without any discrepancies among nodes. So, executing transactions in parallel in the latter part is challenging. Qi et al. proposed an approach to execute transactions in parallel by using a data dependency graph constructed based on not only the static source code of the smart contract but also runtime values. Note that ScalarDL executes transactions in parallel without deciding orders for non-conflicting transactions (i.e., with partial order), thus allowing it to detect Byzantine faults while benefiting from the improved performance through the entire transaction processing.

A paper regarding data reduction in blockchain systems is very motivating. Recently, blockchain systems that use a directed acyclic graph (DAG) structure have been deeply explored to address the performance issue caused by the total ordering of blockchains, as mentioned above. Specifically, a transaction-based DAG, which manages a graph with a fine-grained transaction basis, is a promising approach to improve system performance. However, its high storage overhead is a challenging problem depending on the workload because maintaining metadata (e.g., hash values) for each transaction is necessary. So, Dai et al. proposed GeckoDAG that merges and reduces the metadata for past transactions while keeping order information (i.e., data dependencies).

In addition to such metadata reduction, I thought that pruning the data itself (e.g., transaction payloads and records) would be an effective approach if we could relax the assumption that all histories must be kept forever. Achieving both data traceability and reducibility at the same time would be an ongoing challenge in distributed ledgers.

[1] Xiaodong Qi et al., “Smart Contract Parallel Execution with Fine-Grained State Accesses”, ICDCS 2023, 2023, pp. 841–852.

[2] Xiaohai Dai et al., “GeckoDAG: Towards A Lightweight DAG-based Blockchain via Reducing Data Redundancy”, ICDCS 2023, 2023, pp. 451–462.

Wrap-up

In this blog post, I briefly explained the paper we presented in the ICDCS 2023 workshop and shared an overview of the conference, research trends, and some interesting papers. We will continue to further enhance lazy BFD for productization and submit the full paper to top-tier conferences in the future.