Skip to content

kona-engine: keep SafeDB history after an outage longer than the sequencing window #24131

Open

Description

@einar-oplabs

Summary

After an outage longer than one sequencing window, kona-node's engine reset picks a head above the SafeDB tip, and derivation clears the whole history. op-node keeps the history in the same situation.

Found by the safedb-reviewer pass for #23473 on develop at 7f7abd37fd.

Path

  • The reset walk-back (rust/kona/crates/node/engine/src/sync/mod.rs:78-101) goes back from the unsafe head and stops at the sequencing window, the finalized head or genesis. It never caps the result at the EL's safe head. op-node does cap it (op-node/rollup/sync/start.go:278).
  • When the reset head lands above the tip, truncate_safe_db clears the database (rust/kona/crates/node/service/src/actors/derivation/actor/core.rs:432-438).
  • op-node keeps a tip that is still canonical and resumes from it (op-node/rollup/engine/engine_controller.go:959-967, the fix(op-node): preserve safedb across EL-sync restarts #21111 fix).

Examples: a restart with EL sync after more than about 12 hours of downtime on mainnet, or a batcher outage longer than the sequencing window.

Impact

All earlier history becomes unavailable. Queries return not found or l1 at safe head history unavailable, so games over that range can no longer be checked against the SafeDB. Answers are missing, not wrong. Separately, derivation adopts an unsafe head it never derived. That head is not recorded.

Status: the walk-back was checked in code. The scenario is inferred.

Fix

Cap the reset walk-back at the EL's safe head, as op-node does, so the reset head stays at or below the SafeDB tip whenever that tip is still canonical.

The same PR also carries the low-severity findings from the same review (see #23473):

  • F4: update docs that are out of date. docs/ai/safedb-review.md:30 still describes kona's SafeDB as "crate only; not yet wired", and rust/kona/crates/node/safedb/src/traits.rs:28-31 states a resume rule kona no longer follows.
  • F5: make RollupNodeBuilder itself refuse a SafeDB together with delegation. Today only the CLI guard (bin/node/src/commands/node/command.rs:333-339) prevents serving a database that nothing writes to.

Acceptance criteria

  • After an outage longer than the sequencing window, a restart keeps the SafeDB history that is still canonical, matching op-node.
  • A test covers that scenario and fails without the fix.
  • The two docs above are updated.
  • RollupNodeBuilder rejects a SafeDB combined with delegation, with a test.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions