Range sync: step past a span with no blocks instead of failing the peer that serves it

Open
#812 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 6 days

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
35/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
rust

Research direction

Start with RangeSyncState::complete_batch in crates/net/p2p/src/lib.rs and the Lean and Beacon BlocksByRange handlers named in the issue. Read the stall-window note in docs/spec_deviations.md and the linked discussion in lambdaclass/ethlambda#791 to understand the prior attempt. A fix should advance across empty spans, verify skips using the next batch's parent link, and retry or count a failure when that check does not match.

Written by the indexing model from the issue text.

Description

Problem. Range sync cannot pass a stretch of slots with no block on the chain, such as a long run of missed proposals during a non-finality incident.

  • A BlocksByRange answer with no block in the requested range counts as a failed request. Lean: crates/net/p2p/src/req_resp/handlers.rs:943-953. Beacon: handlers.rs:2905-2913.
  • The range session's cursor only moves when a batch delivers blocks (RangeSyncState::complete_batch, crates/net/p2p/src/lib.rs:865-867). Across a span a whole batch long with no block on the chain, range sync therefore never advances.
  • The honest peers asked for that span cool down. After three such answers in a row they become suspects (SUSPECT_AFTER_FAILURES, crates/net/p2p/src/peer_heads.rs), and their claims stop counting towards the network head.
  • The node can then read as synced while it is still behind, and only the parent lookups of newer gossip blocks fill the gap.

The trade-off is stated in docs/spec_deviations.md (the stall-window bullet of the sync-gate section). It was discussed in rounds seven, eight, eleven and thirteen of lambdaclass/ethlambda#791.

Why it is out of that PR's scope.

  • main already stops at such a span. Its lean handler fails an empty answer and returns before complete_batch (crates/net/p2p/src/req_resp/handlers.rs:817-821 on main), so its cursor does not move either.
  • Excusing empty answers in place was tried and reverted in lambdaclass/ethlambda#791. A peer claiming a high head was then excused on every batch below it, and took a batch on every Status it sent.
  • The fix is a change to range sync itself, not to the sync gate.

What a fix could look like.

  • Advance the cursor: when a peer whose Status head is past the span answers it empty, move the cursor past it.
  • Remember the skip: keep the skipped span and who answered it.
  • Check it with the next batch: use the parent link HeadWatch::note_delivery already uses. If the next batch's first block names, as its parent, the last block delivered below the span, the span was empty and the peer is credited.
  • Otherwise: re-ask the span from another peer, and count a failure against the peer that answered it empty.
Dominant language
Rust
Stars
86
Forks
34
Avg merge
9d 20h
Merged PRs (30d)
77

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from lambdaclass/ethlambda

All issues in lambdaclass/ethlambda

Similar issues

More Rust issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.