Sharding

Overview

Sharding divides a system into partitions and distributes responsibility across validator committees or nodes. Instead of every participant executing and storing every operation, each shard handles a defined portion of the workload. Designs vary between execution sharding, state sharding, and data-availability sharding. They require rules for assigning validators, routing transactions, exchanging authenticated messages, and finalizing results that touch more than one shard.

Parallelism can increase throughput and reduce per-node requirements, but it complicates composability and security. Cross-shard latency, data withholding, uneven load, and committee capture must be addressed. Users should not compare sharded capacity from shard count alone; effective performance depends on workload distribution, coordination overhead, and the security budget protecting each partition. A commitment can authenticate data without ensuring that users or validators can obtain enough of it to verify state or produce an exit. Sharding should be analyzed by separating data publication, commitment, sampling or retrieval, reconstruction, and final settlement. Risk exposure depends on encoding rules, sampling assumptions, participant distribution, withholding resistance, and the party that declares data available. Committee-based, peer-to-peer, and base-layer publication models expose different trust and liveness boundaries.

A rollup or shard can continue producing commitments while users remain unable to reconstruct the information required for independent verification. Applications relying on Sharding should identify how missing data is detected, how long data is retained, and what recovery path exists when normal providers fail.

Operational monitoring should track publication delay, sampling success, reconstruction failures, provider diversity, and the relationship between availability evidence and canonical settlement. Capacity claims should be based on deployed parameters rather than projected roadmap throughput.