Description

See ETCM-313 and ethereum/go-ethereum#2315

Proposed Solution

1. From one peer, called the master peer: Download a skeleton of N headers¹, spaced by N-1 headers in between, starting at bestBlockHeaderNumber + N (i.e. block headers bestBlockHeaderNumber + kN where 1 <= k <= N). To achieve this, use the skip parameter in the GetBlockHeaders request.
2. When the skeleton is received, validate that it corresponds to the requested headers and also check the PoW of the headers.
3. If something is wrong, blacklist the peer and restart.
4. Enqueue the numbers of the headers starting each batch in the skeleton. This means, enqueue for downloading block numbers: bestBlockHeaderNumber + 1 + kN where 0 <= k <= N-1.
5. Download the batches of N sequential headers, in parallel from multiple peers (potentially the master peer as well) by using the enqueued block numbers. Notice there are N batches of size N to download in parallel, so we are requesting N*N block headers.
6. For each downloaded batch, validate if:
   1. It is a valid sequence of headers, with valid parent-child hashes
   2. It fits the skeleton
7. To "fit the skeleton" means:
   1. The first header of the batch must have number equal to any batch starting number (those requested in 4.).
   2. The last header of the batch must have number equal to any skeleton header (those requested in 1.).
   3. The last header's hash must be equal to the hash of the skeleton header found in ii.
   4. The penultimate header's hash must be equal to the parent hash of the skeleton header found in ii.
8. If all validations pass, save this batch (for now, in memory) and continue downloading the rest.
9. If validations 6.i, 7.i, or 7.ii fail, blacklist the peer and enqueue again the batch for downloading.
10. If validations 7.iii or 7.iv fail, increment the counter batchFailuresCount. This is because these validation failures might mean that the master peer is sending us wrong data and that we couldn't find a matching batch from another peer.
11. If on previous step fastSyncMaxBatchRetries is reached, start branch resolution.
12. When all batches are downloaded, process the whole chain of headers as always (chain numbers & hashes with parent hashes).
13. Enqueue block bodies and receipts as always.
14. Start with a new skeleton once all header + body + receipts are downloaded and validated.

¹ N is determined by blockHeadersPerRequest. Currently set exactly to its value but this is open for discussion.

A class called HeaderSkeleton handles the logic of the skeleton size, numbers, batches, gap size, and all that is needed from the FastSync actor. It is kept in memory so if the node goes down everything is recalculated.

Pending work

• Before calling for the branch resolution, the FastSync actor has to rewind a configured number of blocks so the resolution actor starts from there.
• Once we choose a master peer, it's better to stick to it and continue getting the skeletons from that same master peer unless it starts sending us wrong data. Currently on each round of assignBlockchainWork we just take the peer with the highest block number.
• If a peer sends us an incomplete batch (one with less than N headers but valid nonetheless): We could check against that peer's best block so we know if it's malicious, and therefore should be blacklisted, or it may be just the highest block that the peer hast to offer, in which case just retry same batch from another peer.
• We need to integrate with the new actors BlacklistSupport and PeerListSupport mentioned in the task ETCM-531.