Scaling and Optimization
Remove bottlenecks only after identifying whether execution, proving, publication, or settlement is limiting throughput.
Scaling a rollup is an end-to-end problem. Increasing sequencer throughput can simply move the backlog to proving, batch submission, data publication, indexing, or the bridge.
Measure the pipeline before changing it: transaction admission, execution, state persistence, batch construction, data publication, proof generation, settlement inclusion, indexing, and user finality. Track throughput, latency percentiles, backlog age, cost per batch, and failure rate at each stage.
Prefer reversible changes. Increase capacity with load tests, set explicit queue limits and backpressure, and keep enough headroom for settlement congestion or a failed worker. A faster local chain that cannot publish or settle its work is not a successful optimization.
Capacity review cadence
Review capacity after material traffic changes, stack upgrades, settlement-fee changes, and every incident that caused a backlog. Keep a forecast for the next peak period that includes a failed worker or unavailable provider, not only the healthy average case.
| Bottleneck | Useful mitigation | Guardrail |
|---|---|---|
| Sequencer CPU or database writes | Add capacity, tune batching, partition safe read workloads. | Preserve ordering and durability guarantees. |
| Proving backlog | Add compatible workers, aggregate proofs, reduce batch size. | Verify proof latency and cost do not regress. |
| Settlement congestion | Adapt batch policy and fee strategy. | Bound the unsubmitted backlog and reserve funds. |
| DA publication delay | Increase publisher throughput or use a tested fallback. | Confirm independent retrieval before resuming normal intake. |
| Indexing lag | Scale consumers and replay from durable checkpoints. | Keep user APIs clear about their latest indexed block. |
Capacity changes should include a before-and-after load test, rollback condition, and dashboard comparison. Optimizing one stage without verifying downstream queues only moves the incident.