Skip to content

validator: attest and send sync committee messages when the slot's block arrives, and PTC votes when its payload does #829

Open

Description

@MegaRedHand

The validator client signs attestations, sync committee messages and PTC votes only at their clock offsets. It never signs earlier when the slot's block, or at gloas its payload, arrives. The specification asks for whichever comes first. All seven other validator clients read below attest on the block's arrival by default.

Code references are at main @ 5d8628e9.

What happens today

After the proposal, the loop sleeps to the slot's first duty (crates/validator/src/lib.rs:515-520). Then serve_slot_duties sleeps each duty to its own offset:

  • Attestation: clock.until_attestation (lib.rs:659). That is ATTESTATION_DUE_BPS, 3333 (4 s into a 12 s slot), or from gloas ATTESTATION_DUE_BPS_GLOAS, 2500 (3 s) (slot_clock.rs:129-136).
  • Sync committee message: clock.until_sync_message (lib.rs:690), at SYNC_MESSAGE_DUE_BPS or its gloas value (slot_clock.rs:164-171).
  • PTC vote: clock.until_payload_attestation (lib.rs:664), at PAYLOAD_ATTESTATION_DUE_BPS, 7500 (9 s) (slot_clock.rs:142-144).

The client reads nothing from the beacon node's event stream. Its clock is independent of the stream on purpose: "that stream is best-effort by contract, dropping events for a slow subscriber, and a dropped event must not mean a missed duty" (slot_clock.rs:1-6).

What the specification says

At consensus-specs master @ 2ea9d96a:

  • Attestation: "when either (a) the validator has received a valid block from the expected block proposer for the assigned slot or (b) get_attestation_due_ms() milliseconds has transpired since the start of the slot -- whichever comes first" (specs/phase0/validator.md:610-614). Gloas changes only the deadline (specs/gloas/validator.md:99).
  • Sync committee message: the same two conditions, with get_sync_message_due_ms() (specs/altair/validator.md:346-352).
  • PTC vote: "as soon as it has seen the execution payload envelope and blob data for the block, and no later than get_payload_attestation_due_ms()" (specs/gloas/validator.md:402-404). blob_data_available may be set to False only once that deadline has passed (gloas/validator.md:425-427). So a PTC vote is triggered early by the payload and its blob data. The block alone is not enough, and an early vote is always one that saw both.

Why it matters

  • Less time to reach the aggregators. A subnet's aggregators publish at AGGREGATE_DUE_BPS: 8 s into the slot before gloas, 6 s at gloas. An attestation sent at the deadline gets 4 s, or 3 s at gloas, to reach them. Sent when the block arrives, usually well before the deadline, it gets that time plus the wait. A vote that misses every aggregate depends on a proposer's node having seen it on its own.
  • The same for sync committee messages, ahead of the contributions at CONTRIBUTION_DUE_BPS.
  • PTC votes have to reach the next proposer. A vote sent at 9 s has 3 s to propagate before the next slot's block is built. Sent when the payload and blobs are available, it has most of the slot.
  • The beacon node's latency comes on top. The attestation data request starts at the deadline, so a node that takes a second to answer makes the vote a second late. Started on the block's arrival, the same wait mostly falls before the deadline.

None of this has been measured for this client.

What other clients do

Client Attestation Sync committee message PTC vote Read at
Lighthouse A head event for the slot or the deadline, whichever comes first, on by default (attestation_service.rs:205-216). It checks the produced data's block root against the event's, and falls back to the deadline on a mismatch (:285-325). Optimistic heads do not trigger it (beacon_head_monitor.rs:194-200). --disable-beacon-head-monitor turns it off The same trigger (sync_committee_service.rs:145-152) On execution_payload_available, only when the data says payload_present, and again at the deadline if that attempt fails (payload_attestation_service.rs:170-178, :308-316). --disable-payload-available-monitor unstable @ 11813cfb
Teku On head events, on by default, --validators-early-attestations-enabled (EventSourceHandler.java:174-180, ValidatorConfig.java:65) At the deadline only (SyncCommitteeScheduler.java:125) On execution_payload_available for the last reported head (EventSourceHandler.java:203-209) master @ 62a54b96
Prysm A head event at the slot or the due time, on by default, --disable-attest-timely (attest.go:305-312) The same wait (sync_committee.go:41) execution_payload_available or the deadline (wait_helpers.go:86-94) develop @ 0c1e3d24
Nimbus A block event or the deadline (attestation_service.nim:540-544), then a further half of the attestation offset after the block (validator_duties.nim:68-74). On with the default --block-monitor-type=event (conf.nim:1141-1145) The same wait (sync_committee_service.nim:362) execution_payload_available or PAYLOAD_DUE_BPS, and once more at the deadline when the blob data was missing (payload_attestation_service.nim:85) unstable @ 9bed158a
Lodestar Races the due time against a head event at the slot (attestation.ts:80-85). A hidden --afterBlockDelaySlotFraction, 0 by default, delays the publish after an early block (attestation.ts:29, :176-178) The same race (syncCommittee.ts:89-93) execution_payload_available for the head at the slot, or the deadline (ptc.ts:70-76) unstable @ 77088eb7
Vouch "Fast track": on a head event for the current slot it waits 500 ms, then runs the slot's attestation job (events.go:200-212, main.go:262-264) The same, controller.fast-track.sync-committees (events.go:214-219) The first execution_payload_available, with the deadline as the fallback (events.go:42-60) master @ 691f3652
Vero The first head_v2 from any node, always on (attestation.py:172-177) The same trigger (sync_committee.py:154-160) The first execution_payload_available (ptc.py:91-105) master @ 9d110c1d

File names are relative to each client's validator client directory, except Vouch's main.go, at its repository root.

All seven attest on the block's arrival by default, and all seven vote PTC on execution_payload_available. Six of the seven also send sync committee messages early; Teku does not.

What it takes

  • An event stream in the beacon node client. A server-sent events reader for GET /eth/v1/events?topics=head, plus execution_payload_available at gloas, behind a new BeaconNodeApi method. It has to reconnect after a drop.
  • Triggers next to the clock. The slot clock stays the deadline, and whichever fires first runs the duty, once.
    • A head event whose slot is the current slot and whose execution_optimistic is false wakes the slot's attestation and sync committee message.
    • An execution_payload_available event for the current slot's block wakes the PTC vote.
  • One attempt per duty. attest must not be called twice in a slot (attestation.rs:83-94): a second call signs new data that slashing protection refuses. The early trigger and the deadline have to share one attempt.
    • Falling back to the deadline is safe only when the early attempt signed nothing. Lighthouse falls back on a block-root mismatch, which it detects before signing.
  • A check of what came back. The attestation data's beacon_block_root should be the event's block, or the duty waits for the deadline (Lighthouse, attestation_service.rs:285-292).
  • An early PTC vote only when it says the payload is there. If the data does not say payload_present, wait for the deadline, as Lighthouse does (payload_attestation_service.rs:308-316). blob_data_available may be False only once the deadline has passed, so an early vote must also have it True.
  • Node side, for the PTC trigger. ethlambda's node accepts but never emits execution_payload_available or head_v2 (crates/blockchain/src/events.rs:296-311; docs/rpc.md:1031-1036 lists them as never emitted). It emits head (docs/rpc.md:1023), so attestations and sync committee messages can be triggered against it today, while the PTC trigger needs the node to emit the event first.

To decide

  • A delay after the block. Vouch waits 500 ms, Nimbus half the attestation offset, Lodestar nothing by default. A short wait lets the block propagate before votes for it arrive at other nodes.
  • Which node's events, with several configured. Vero takes the first event from any node, then cross-checks the data with a majority of nodes. Lighthouse opens a head stream on each of its candidate nodes (beacon_head_monitor.rs:160-164).
  • A flag to turn it off. Lighthouse, Teku and Prysm have one; Vero and the PTC paths of most clients do not.
  • The aggregation. It stays at its own offset. The early attestation's data is the data it asks the aggregate for, as today (attestation.rs:21-50).

Tests

  • A mock event stream that delivers a head for the current slot one second in: assert the attestation and the sync committee message are published before the attestation offset.
  • No event: assert both still go out at their offsets.
  • A head event for the wrong block root, or an optimistic one: assert the duty waits for the deadline.
  • An event arriving after the deadline attempt has started: assert nothing is signed twice.
  • An execution_payload_available whose data says payload_present = false: assert the PTC vote waits for its deadline.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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