18 KiB
lez/chain_state — Two-Tier Chain State
Design doc for the shared block-apply engine and two-tier chain state that
backs decentralized sequencing. Status: interface freeze — the
apply_block signature and the ChainState tip/state shape below are the
contract the produce-on-turn and follow-blocks tracks build against. Changing
them after the tracks split forces rework in both.
Branch: erhant/lez-two-tip-chain-state (off erhant/indexer-recoverable-invalid-blocks).
1. Motivation
Decentralized sequencing requires every honest node — sequencer or indexer — to
converge on the same chain and the same state by running one deterministic
validate-then-apply path over blocks pulled from the channel. That path today
lives only inside the indexer (lez/indexer/core/src/block_store.rs), where the
recoverability work built a park-and-skip ingest: accept_block validates a
block against the current tip, applies it to a scratch clone of state atomically,
and on any failure records a StallReason, freezes the tip, and marks the bad
block processed without applying it. The sequencer has no equivalent — it only
produces blocks and reads peer inscriptions for finalization; it never executes
peer blocks into its own state.
This crate lifts that logic into a shared home and generalizes it into a two-tier state machine so the sequencer can produce on the head while both sequencer and indexer expose their exact current state.
2. Crate placement & layering
lee_core ← lee (owns V03State) ← common (owns Block, BedrockStatus, clock_invocation, recompute_hash) ← lez/chain_state ← { indexer/core, sequencer/core }
- Not
lee: the apply logic needsBlock/BedrockStatus/clock_invocationfromcommon, andcommondepends onlee— putting it inleeinverts the layer. lez/chain_statesits abovecommon, depends oncommon+lee, and is consumed by bothindexer/coreandsequencer/core.
Persistence boundary. chain_state holds the in-memory state machine and the
pure logic; it performs no I/O. Each consumer keeps its own RocksDBIO and
drives the scratch → put_block → commit ordering, exactly as accept_block does
today. This keeps the crate fully unit-testable without a DB.
3. The apply_block entry point
A single pure function, called identically whether the block was produced by us, adopted from a peer, or read finalized from the channel:
/// Validate `block` against `tip`, then apply it to `state`. Pure: no I/O.
/// Mutates `state` only on success; on failure `state` is untouched and the
/// caller parks.
fn apply_block(
tip: Option<&Tip>,
block: &Block,
state: &mut V03State,
) -> Result<(), BlockIngestError>;
Validation order (unchanged from the indexer): hash integrity
(recompute_hash) → block-id continuity → prev_block_hash linkage, with a
None tip expecting the genesis block. Application splits off the mandatory
trailing clock tx, executes user txs (genesis = public-only), applies the clock
last.
Shared types moved into the crate: AcceptOutcome, BlockIngestError,
StallReason, and Tip.
struct Tip { block_id: u64, hash: HashType, l1_slot: Slot }
enum AcceptOutcome { Applied, AlreadyApplied, Parked(BlockIngestError) }
Tip carries l1_slot (recorded atomically with the tip) because the anchor /
chain-consistency logic keys on the inscription slot, not just (id, hash).
4. The two-tier ChainState
struct ChainState {
final_state: V03State, // driven by finalized channel ops
final_tip: Option<Tip>,
head_state: V03State, // final_state + applied head blocks
head_blocks: Vec<HeadEntry>, // ordered, above final_tip
final_stall: Option<StallReason>, // the one stall — persisted to RocksDB. See §4a
}
struct HeadEntry { this_msg: MsgId, block: Block }
The head tier is a MsgId-keyed chain (adopted/orphaned reference
this_msg/parent_msg); the final tier is block-id-keyed. apply_block
validation stays LEZ-level (block_id + prev_block_hash) — the two chains run
in parallel and must agree, so we validate via apply_block and track MsgId
for revert correlation.
Operations:
apply_adopted(inscription) -> AcceptOutcome— dedup bythis_msgagainst our outbox, elseapply_blockon the head tip; on success pushHeadEntry. On failure, do nothing durable: the head tip simply stays at the last valid block. No stall is recorded (see §4a) — the head self-heals from the stream.apply_channel_update(orphaned, adopted)— revert everyorphanedbythis_msg, re-derivehead_state(clonefinal_state, replay survivors), then apply everyadoptedin order. Atomic per event.finalize_up_to(block_id)— movehead_blocksup toblock_idintofinal_state(already validated; a move, not a re-apply).apply_finalized(inscription)— steady state: if present in head bythis_msg,finalize_up_to; cold-start backfill (not in head):apply_blockdirectly tofinal_state, mirror into head. If a finalized block fails to apply, setfinal_stalland persist it — this is the only stall (see §4a).rollback_orphan(this_msg)— drop from that entry forward, re-derive head.status() -> { final_height, head_height, final_stall }— for RPC/UI. A derived "head blocked" indicator can be computed on demand (see §4a) without persisting.
For the indexer (finalized-only next_messages stream), head_blocks stays
empty and head == final; it exercises only apply_finalized. The sequencer
uses both tiers from day one.
4a. One stall — final_stall, persisted
There is a single stall, final_stall, on the final tier. The head tier does
not carry its own stall, and this is deliberate.
The head and the final tier never represent two independent problems: a block
always reaches the head first (as adopted) and only later the final tier (as
finalized), so a would-be "head stall" is just the earlier, provisional sighting
of the exact block that final_stall records durably if it finalizes — the same
event modeled twice.
And the head does not need a recorded reason to freeze. The tip-freeze is intrinsic: not applying a bad block is what freezes the tip; no marker is required. The head's freeze is also transient and self-healing — the bad block is either orphaned (a competing valid block at the same height wins and applies on its own) or it finalizes. Subsequent adopted blocks that chain on the bad one fail validation by themselves, and the producer builds on the head tip regardless of any marker. So a persisted head stall would be redundant (re-derived from the stream on restart) and buys no behavior.
final_stall is the stall that does real work:
- The indexer already requires it and ships it today. The indexer has only a
final tier (finalized-only stream, no head); its startup chain-consistency /
anchor check reads the persisted stall to know where it is parked. The shared
final_stallserves that unchanged. - It survives restart and is what we surface as
Stalled. - "A bad block finalized" is the only irreversible, actionable condition — the signal the committee acts on to evict a bad sequencer. A provisional head block that may vanish on the next reorg is not something to evict over.
The one thing we forgo is an early warning that a sequencer is posting garbage
before it finalizes. That condition frequently self-heals via reorg, so alarming on
it is mostly noise; if wanted, it is a derived, non-persisted indicator (e.g.
"k adopted inscriptions above the head tip remain unapplied"), computed on demand —
not a second StallReason in the struct.
So the sequencer and the indexer share exactly one stall concept, keeping the two consumers uniform.
4b. Producer contract — write on turn, build on last valid
The sequencer publishes only on its own turn (the SDK queues out-of-turn publishes). When it is our turn we build the next block on the current head tip, which is by construction the last validly-applied block. So if the head is frozen on a peer's bad block, we build on that frozen valid tip — the same parent every honest sequencer chooses — and thereby skip the bad block rather than extend it. A parked node keeps following peers' valid blocks as they arrive; the moment it also gets a turn, it produces the next valid block on its last valid tip. Net: parking never stops us from producing correctly on our turn.
5. Event → tier mapping
Event::BlocksProcessed { checkpoint, channel_update: { orphaned, adopted }, finalized }:
| Input | Source | Effect |
|---|---|---|
| adopted inscription | channel_update.adopted |
validate + apply to head |
| orphaned inscription | channel_update.orphaned |
revert from head, return txs to mempool |
| finalized inscription | finalized[].ops (FinalizedOp::Inscription) |
move head→final, or apply directly on backfill |
| own publish | publish-return | optimistically apply to head, record this_msg |
Golden rules: (1) validation is deterministic, so every honest node makes the same accept/park decision. (2) An invalid block is processed but discarded — never applied, never halts the node. (3) Finalized is never reverted. (4) We rebuild orphaned blocks ourselves; we do not trust the SDK's republish (it keeps stale LEZ contents — prev-hash, tx selection, and resulting state were all computed against the old parent).
6. Scenarios
Processing one BlocksProcessed event
flowchart TD
EV["Event::BlocksProcessed"] --> ORPH{"orphaned<br/>non-empty?"}
ORPH -->|yes| REV["For each orphaned by this_msg:<br/>drop from head_blocks,<br/>return its txs to mempool"]
REV --> RED["Re-derive head_state:<br/>clone final_state, replay survivors"]
ORPH -->|no| ADO
RED --> ADO{"adopted<br/>non-empty?"}
ADO -->|"yes, in order"| DEDUP{"this_msg in<br/>our outbox?"}
ADO -->|no| FIN
DEDUP -->|"yes (our own)"| SKIP["skip: already applied optimistically"]
DEDUP -->|no| VAL["apply_block on head tip"]
VAL --> OUT{"AcceptOutcome"}
OUT -->|Applied| APP["append this_msg+block to head,<br/>advance head tip, clear stall"]
OUT -->|AlreadyApplied| SKIP
OUT -->|"Parked(err)"| PARK["freeze head tip — do NOT apply.<br/>No stall recorded (self-heals<br/>via reorg/finalization)"]
SKIP --> FIN
APP --> FIN
PARK --> FIN
FIN{"finalized<br/>inscriptions?"}
FIN -->|"already in head (steady state)"| MOVE["finalize_up_to:<br/>move head→final, trim head_blocks"]
FIN -->|"not in head (cold-start backfill)"| DIRECT["apply_block directly to final"]
DIRECT --> DOK{"applied?"}
DOK -->|yes| MIRROR["mirror into head"]
DOK -->|"no (invalid finalized)"| FSTALL["record StallReason on FINAL,<br/>freeze final tip"]
FIN -->|none| CP
MOVE --> CP
MIRROR --> CP
FSTALL --> CP
CP["persist checkpoint atomically"]
Park / recovery status
stateDiagram-v2
[*] --> Syncing
Syncing --> CaughtUp: stream drained, no stall
CaughtUp --> Syncing: new adopted / finalized arrives
Syncing --> Parked: invalid block FINALIZED (apply_finalized fails)
Parked --> Parked: further non-chaining finalized blocks (orphans_since++)
Parked --> Syncing: valid successor finalizes on frozen final tip → stall cleared
note right of Parked
final_stall — the one stall. Persisted, survives restart.
Head-tier bad blocks do NOT enter this state:
the head tip freezes intrinsically (no stall) and
self-heals via reorg/finalization. Producer (on our
turn) builds on the last valid tip either way.
end note
Scenario table
Normal flow
| # | Scenario | Handling | Expected |
|---|---|---|---|
| 1 | Adopted block chains cleanly on head tip | apply_block → Applied; append {this_msg, block} to head |
head advances; converges with peers |
| 2 | Our own block comes back in adopted |
dedup by this_msg against outbox → skip |
no double-apply |
| 3 | Adopted block later finalizes | finalize_up_to moves head→final, trims head_blocks |
final advances; no re-apply |
| 4 | Re-delivery of an already-applied block | id ≤ tip & stored hash matches → AlreadyApplied |
idempotent, no state change |
Reorg / orphan
| # | Scenario | Handling | Expected |
|---|---|---|---|
| 5 | Our block orphaned at turn handoff (stale-parent race) | revert by this_msg, return txs to mempool, rebuild on new head tip (not SDK republish) |
our txs re-queued; next block on correct parent |
| 6 | Batch reorg: some orphaned + some adopted in one event |
revert all orphaned, re-derive head, then apply all adopted in order | deterministic convergence |
| 7 | Orphan chain (parent transitively off canonical) | SDK surfaces all affected as orphaned; revert each, replay survivors |
head_state matches new canonical branch |
Invalid / bad block — "stall" below means the one persisted final_stall
(§4a). A bad block seen only in adopted records no stall — the head tip just
freezes and self-heals; it becomes a final_stall only if it finalizes.
| # | Scenario | Handling | Expected |
|---|---|---|---|
| 8 | Authorized sequencer posts a block with an invalid state transition | head: apply_block → Parked, freeze head tip, no stall recorded. If it finalizes: persisted final_stall |
park-and-skip; no apply, no halt |
| 9 | Broken chain link / hash mismatch / unexpected id in adopted | Parked(BrokenChainLink / HashMismatch / UnexpectedBlockId); same park |
frozen tip; peers park identically |
| 10 | Undeserializable inscription payload | park with Deserialize (no header); processing advances |
recover when a valid block chains on frozen tip |
| 11 | Valid successor after a park (recovery) | block chaining on frozen tip → Applied → clear stall |
head resumes automatically; no divergence |
| 12 | Further non-chaining blocks while parked | keep first StallReason, bump orphans_since |
original cause preserved; still parked |
Producing while parked
| # | Scenario | Handling | Expected |
|---|---|---|---|
| 13 | It's our turn but head is parked on a bad block | producer builds on the frozen valid tip (head tip = last valid), skipping the invalid block | we emit the next valid block on the same parent honest peers use — chain moves on our turn |
Startup / backfill
| # | Scenario | Handling | Expected |
|---|---|---|---|
| 14 | Cold start / reconnect backfill | history via finalized, empty channel_update; apply directly to final + mirror head |
head == final until live deltas start |
| 15 | Local store belongs to a different chain (L1 reset) | anchor-based chain_consistency check at startup: wipe+reindex if allow_chain_reset, else error |
no silent divergence |
7. Invariants
Should-never-happen conditions — assert/log, don't silently absorb:
- An
orphanedentry never references a block at or below the final tip — finalized is irreversible. If seen, it is a bug. headtip ≥finaltip at all times;head_blocksholds exactly the blocks between them.- After processing any event,
head_state == final_statereplayed throughhead_blocks(the re-derivation is the source of truth). - A parked node's frozen tip is identical across all honest nodes for the same invalid block (deterministic validation).