Repository navigation
Description
Problem. A beacon gossip message whose stateful checks answer Outcome::Queue is reported to gossipsub as a final IGNORE straight away (disposition in crates/net/p2p/src/beacon/verdict.rs:566, Outcome::Queue(_) => (MessageAcceptance::Ignore, true) at :569, reported by report at :577) and handed to the chain actor. The actor often imports it moments later, but by then gossipsub has its verdict, so this node never forwards it.
#802 makes this visible for gloas blocks. A block built on a FULL parent whose envelope is not verified yet is now queued (QueueReason::ParentPayloadUnverified, gossip::block::stateful_checks) and held by the actor until the envelope verifies (hold_block_for_parent_payload and release_blocks_awaiting_parent_payload, crates/blockchain/src/beacon_envelope.rs:959 and :1067). In the devnet trace in that PR, the envelope landed 40 ms after the block: the block now imports, but this node, and every other node that hits the same race, is a dead end for it. The block reaches the rest of the mesh only through nodes that did not race.
This applies to every topic that answers Queue, not only that rule: beacon_block (parent_unknown, parent_not_ready, shuffling_unavailable, parent_payload_unverified), data_column_sidecar (parent_unknown, parent_not_ready, shuffling_unavailable, and from gloas block_unknown and block_not_ready) and execution_payload (block_unknown, block_not_ready, state_not_cached).
Why it was out of scope for #802. That PR changes one rule's verdict from IGNORE to queue. This change is about how every queued verdict is reported. It also needs a signal that does not exist yet: nothing tells the p2p actor that a queued object has settled. And an Accept must not rest on the actor's import alone. Whatever gossip rules the queue returned before still have to pass, along with the seen cache. For parent_payload_unverified those are now only the finalized-ancestry walk's queue and the lookahead gate, since #802 moved every REJECT that does not read the parent's payload ahead of it. For parent_unknown they are nearly all of them.
The window. validate_messages() holds every beacon message until it has a verdict, and the message cache keeps it for history_length(6) heartbeats of 700 ms (gossipsub_config, crates/net/p2p/src/lib.rs:917, history at :933), about 4.2 s. A verdict reported after that propagates nothing (report_message_validation_result returns false, counted by lean_beacon_gossip_verdict_expired_total).
What a fix could look like.
- On
Queue, do not report yet. Park theGossipId(message id and propagation source) with the validated object, keyed by the root it waits on, under a deadline comfortably inside the window (say 3 s). Bound the parked set per root and in total. - Add a message from the chain actor to p2p (
crates/net/api) for the events that settle a queued object: a block imported, an envelope verified. On it, re-run the object's stateful checks against the store as it now is. Report the new verdict: Accept records the seen entry throughsettleand propagates; anything else reports as it would today. - At the deadline, report IGNORE, as today.
- Count holds, settles and expiries in metrics, so it is visible whether queued messages settle inside the window.
A test should queue a gloas block on an unverified parent payload, verify the envelope inside the window, and assert that the block's verdict is reported as Accept exactly once.
Activity