Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MuFASA -- Asynchronous Checkpoint for Weakly Consistent Fully Replicated Databases

Raaghav Ravishankar, Sandeep Kulkarni, Nitin H Vaidya

TL;DR

This work addresses checkpointing in fully replicated, weakly consistent databases by introducing DTCS, a checkpoint notion that combines transactional and distributed snapshot consistency. It proposes MuFASA, an asynchronous, size-minimal snapshot algorithm that uses a vPoGC construct and a color-based protocol to collect exactly one copy of each object across $n$ replicas with only $O(n)$ control messages, without blocking ongoing transactions. The authors prove that MuFASA yields a DTCS snapshot and supports crash recovery by rolling back to the stored checkpoint and reapplying subsequent committed transactions. The approach enables strong, invariant-preserving views in otherwise eventual systems, with practical implications for main-memory, highly available databases and avenues for extension to partitioned or quorum-based replication.

Abstract

We focus on the problem of checkpointing in fully replicated weakly consistent distributed databases, which we refer to as Distributed Transaction Consistent Snapshot (DTCS). A typical example of such a system is a main-memory database that provides strong eventual consistency. This problem is important and challenging for several reasons: (1) eventual consistency often creates anomalies that the users do not anticipate. Hence, frequent checkpoints to ascertain desired invariants is highly beneficial in their use, and (2) traditional checkpoints lead to significant overhead and/or inconsistencies. By showing that the traditional checkpoint leads to inconsistencies or excessive overhead, we define the notion of size-minimal checkpointing for fully replicated databases. We present an algorithm for checkpointing with minimal checkpointing overhead (only O(n) new messages and addition of a single counter for existing messages). It also provides a significant benefit over existing checkpointing algorithms for distributed systems and main-memory databases. A key benefit of DTCS is that it summarizes the computation by a sequence of snapshots that are strongly consistent even though the underlying computation is weakly consistent. In essence, when anomalies arise in an eventually consistent system, DTCS enables one to concentrate solely on the snapshots surrounding the time point of the anomaly.

MuFASA -- Asynchronous Checkpoint for Weakly Consistent Fully Replicated Databases

TL;DR

This work addresses checkpointing in fully replicated, weakly consistent databases by introducing DTCS, a checkpoint notion that combines transactional and distributed snapshot consistency. It proposes MuFASA, an asynchronous, size-minimal snapshot algorithm that uses a vPoGC construct and a color-based protocol to collect exactly one copy of each object across replicas with only control messages, without blocking ongoing transactions. The authors prove that MuFASA yields a DTCS snapshot and supports crash recovery by rolling back to the stored checkpoint and reapplying subsequent committed transactions. The approach enables strong, invariant-preserving views in otherwise eventual systems, with practical implications for main-memory, highly available databases and avenues for extension to partitioned or quorum-based replication.

Abstract

We focus on the problem of checkpointing in fully replicated weakly consistent distributed databases, which we refer to as Distributed Transaction Consistent Snapshot (DTCS). A typical example of such a system is a main-memory database that provides strong eventual consistency. This problem is important and challenging for several reasons: (1) eventual consistency often creates anomalies that the users do not anticipate. Hence, frequent checkpoints to ascertain desired invariants is highly beneficial in their use, and (2) traditional checkpoints lead to significant overhead and/or inconsistencies. By showing that the traditional checkpoint leads to inconsistencies or excessive overhead, we define the notion of size-minimal checkpointing for fully replicated databases. We present an algorithm for checkpointing with minimal checkpointing overhead (only O(n) new messages and addition of a single counter for existing messages). It also provides a significant benefit over existing checkpointing algorithms for distributed systems and main-memory databases. A key benefit of DTCS is that it summarizes the computation by a sequence of snapshots that are strongly consistent even though the underlying computation is weakly consistent. In essence, when anomalies arise in an eventually consistent system, DTCS enables one to concentrate solely on the snapshots surrounding the time point of the anomaly.

Paper Structure

This paper contains 20 sections, 2 theorems, 1 equation, 3 figures, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Checkpointing in Distributed Systems
  4. Checkpointing in Database Systems
  5. Defining Checkpoint Consistency for a Replicated Database
  6. Limitations of existing checkpoint approaches
  7. Defining Checkpoint Consistency in Replicated Database Systems
  8. Challenges in Checkpointing Asynchronously Replicated Distributed Databases
  9. System Model : Model of the Database
  10. Asynchronous DTCS Algorithm Requirements
  11. MuFASA : Multi-replica Fully Asynchronous Snapshot Algorithm
  12. Overview
  13. Elements of the Algorithm
  14. The Checkpointing Algorithm
  15. Proof
  16. ...and 5 more sections

Key Result

Theorem 3

Assume that at most one DTCS operation takes place at a given time. Let $\sigma_1$ and $\sigma_2$ be any two sequential histories that include any DTCS $CP$ and all committed transactions. Then the index of $CP$ in $\sigma_1$ is the same as the index of $CP$ in $\sigma_2$.

Figures (3)

  • Figure 1: Examples for Transaction Convergence and Distributed Snapshot Consistency.
  • Figure 2: Transaction flow and Commit Log Coloring.
  • Figure 3: Execution of MuFASA in a 3 replica system. Arrows (messages) that reached a replica are messages that are considered delivered, the rest are in transit.

Theorems & Definitions (4)

  • Definition 1
  • Definition 2
  • Theorem 3
  • Theorem 12