Steelhead: Interleaving Partially Synchronous and Asynchronous Commit Rules on a Shared DAG

Steelhead (arXiv:2609.30163) is a dual-mode consensus mechanism that interleaves partially synchronous and asynchronous commit rules on a shared Directed Acyclic Graph (DAG) to achieve both low latency and robust liveness. It determines which rule decides each round slot based solely on the round number, allowing the system to inherit the speed of partial synchrony under healthy conditions while maintaining progress via asynchronous rules during network stalls.

Key features include: Adaptive Period Selection: Validators dynamically adjust the frequency of asynchronous slots by replaying committed data to find the configuration with the lowest expected message delays, without using extra messages or timers. Efficiency: The protocol opens a common coin only every k-th round rather than every round, reducing overhead while ensuring liveness when the partially synchronous path is blocked. Implementation: It is instantiated with Mysticeti and Mahi-Mahi (for n ≥ 3f+1) and BlueBottle variants (for n ≥ 5f+1), with safety and liveness proofs mechanized in Lean 4. Performance: Simulations show Steelhead matches the latency of purely partially synchronous protocols in healthy networks and closely follows asynchronous protocols during adverse conditions, adapting quickly in both directions.

Steelhead is a dual-mode consensus mechanism for DAG-based replicated state machines or blockchains that interleaves two commit rules over a single shared directed acyclic graph: a fast, partially synchronous rule for ordinary operation, and a slower but asynchronous rule that is invoked periodically, specifically every k-th round. The design goal is to capture the low-latency behavior of partially synchronous consensus when the network is well-behaved, while retaining the liveness and safety properties of asynchronous consensus when network delays, leader failures, or other asynchrony assumptions are violated. In effect, the protocol uses the same DAG of votes, proposals, or transaction attestations as the substrate for both modes, rather than maintaining separate ledgers or switching between disconnected protocols.

A central contribution is the way Steelhead composes these two commit rules so that they reinforce each other. The partially synchronous path provides the fast path, allowing the system to make progress with bounded latency under standard partial synchrony assumptions. The asynchronous path, activated on a fixed schedule, acts as a robust checkpoint: every k-th round is decided by an asynchronous rule whose leader is revealed by a common coin after votes are collected. This coin-based leader revelation is intended to provide a mode that remains live even when leader omission or adversarial timing makes partially synchronous progress unreliable. The periodic asynchronous decisions therefore bound how long the system can remain stuck in a bad network regime and give the protocol a principled fallback that does not require abandoning the DAG or the existing vote structure.

The significance of Steelhead lies in addressing a long-standing tension in consensus design: partially synchronous protocols can be fast but are fragile under asynchrony or leader omission, whereas purely asynchronous protocols are robust but often pay a substantial latency or throughput cost. By interleaving the two commit rules on one DAG, Steelhead offers a structured tradeoff parameterized by k: smaller values increase robustness at the cost of more frequent asynchronous rounds, while larger values improve average-case speed but leave the system exposed for longer to adverse conditions. For technically oriented readers, the paper is interesting because it treats dual-mode consensus not as a simple fallback switch, but as a compositional protocol construction in which fast-path performance and asynchronous resilience are woven into the same data structure and decision process.