-
Notifications
You must be signed in to change notification settings - Fork 7
Blockpool
technical notes to the go-ethereum blockchain synchronisation module
An ethereum node needs to acquire the set of blocks from which the current consensus can be proved. This includes scenarios where the node starts up from scratch (i.e., with an empty database) in which case the block pool acts as a download manager as well as normal operation where the node needs to synchronize recent blocks by finding the best candidate new blocks. Since network latency and disruption can cause any amount of blocks to be missed the two scenarios can be considered the same task: finding the best chain to give to the blockchain manager.
Goal: The Blockpool needs to find the best chain to be inserted in the blockchain.
In order the achieve this, the blockpool can only rely on p2p communication using the ethereum wire protocol.
The blockchain database is a tree structure with exactly one root, the genesis block and contains at least one leaf node. The consensus protocol specifies that the block with the highest total difficulty is the consensus state. The total difficulty of a block is always strictly higher than that of its parent, so a consensus state is always a leaf node. The current known best block is also called the head block of the node. When the node is starting up the first time with an empty database, the head block is the genesis block itself.
A chain is any sequence of blocks connecting a root with a leaf node. The leaf node with the highest total difficulty is the head of the block chain which specifies the consensus state of the network. The chain connecting the genesis block with the head is the canonical chain.
A sequence of blocks or block hashes that form a contiguous subsequence of a chain is called a section. The youngest section of blocks headed by the peer's current head is called the peer's head section.
If there are two different blocks with the same parent, the chain has a fork. Those blocks which do not end up on the canonical chain are called stale blocks, a sequence of such blocks is a stale section.
The Blockpool stores a pool of blocks which have not been processed yet, thus it acts as a queue or processing buffer for the blockchain manager. Therefore the notions of chain, fork, section, etc. carry over to sequences of blocks and blockhashes in the blockpool.
As a proper decentralised consensus system, Ethereum relies only on other peers to synchronise. The eth wire protocol defines the messages that the blockpool can use to communicate. For a naive new node, the only entry point to the blockchain is the head block advertised by a peer as part of their status message (StatusMsg, i.e., protocol handshake) or when a NewBlockMsg is received from a miner. A sequence of block hashes can be requested and received with getBlockHashesMsg/BlockHashesMsg message pair. A set of blocks can be requested and received with getBlocksMsg/BlocksMsg message pair.
Although the exact way synchronisation is done is not specified by either the wire protocol or the block chain consensus protocol, the messages provided restrict the synchronisation process:
Synchronisation: proceeds from head to known block by requesting and fetching block hashes iteratively from young to old (head to root). Based on block hashes, blocks can be requested. Based on the parenthash of a block, independent sections can be linked and a chain established. By checking if a block is found in the block chain the root of the blockpool can be established and the chain can be inserted in the blockchain.
We will elaborate on how exactly this is done below.
Our implementation currently provide 4 entry points to the blockpool for the eth protocol instances running on each peer. These get called when a message relevant to blockchain synchronisation is received from the peer.
AddPeer is called by the eth protocol right after receiving the status message (or protocol handshake) from a peer. The blockpool registers the peer with information about its total difficulty and current head block. It also registers peer callbacks so that the blockpool can send requests directly to a peer (requestBlockHashes and requestBlocks) and report a peer error (peerError).
Once a peer is removed (disconnect or eviction), the protocol calls back to the BlockPool to unregister the peer with RemovePeer.
The blockpool synchronisation relies on choosing the best peer out of the connected peers. The best peer is defined as the peer with the highest advertised total difficulty that is ahead of us. The blockpool works by following and replicating the best peer's canonical chain.
Strategy: follow the best peer and replicate its canonical chain.
In our architecture the Blockpool controls peer selection only indirectly.
If a peer violates a blockpool policy, the blockpool reports it as a peer error. If the error is fatal, the peer is disconnected and suspended for PeerSuspensionInterval. Synchronisation therefore crucially relies on a healthy network (i.e., connectivity as well as efficient message relaying).
Every time a peer is registered, its total difficulty is checked against the best peer. If it is higher, the peer is promoted as best peer. Every time the best peer disconnects, a new "second best" peer is promoted. Surely this is limited to peers that are ahead of us, i.e., ones with advertised total difficulty strictly greater than ours.
Since our own total difficulty can change when mining, the blockpool needs to receive updates of the last total difficulty. This is achieved by subscribing to ChainHeadEvent posted in the blockchain manager. A blockpool update routine is listening to these total difficulty updates and if it finds that our own total difficulty goes above that of the best peer, the peer is demoted (our own node becomes the best peer as it were). Note that because of forks it is not sufficient to check if we already have the head block of a peer.
The total difficulty determining best peer status is communicated in one of the following 4 ways:
- when a peer is connected, the
ethprotocol receives its status message that contains the total difficulty of the peer as well as its current block hash. If this total difficulty is higher than that of self and other peers, a peer switch happens: the new peer is promoted best peer, the current best peer (if there was one) is demoted. - when a new block message arrives, it contains the new total difficulty, as well as the head block (including its header hash). To allow for the update of peers head info and a potential peer switch, the protocol calls
AddPeer. In the special case when a new block message arrives from the best peer, its head information is updated and a new process is launched to obtain the peer's new head section. - when a peer sends a block, it can optionally include the total difficulty td. The head information on the peer gets updated in this case as well. Since we requested this block, it cannot be the head of the pool. However if the current best peer as well as all others with difficulty above td get disconnected, this peer can be choosen as best peer.
The third entry poing, AddBlockHashes is called by the eth protocol when a blockHashesMsg (blockhashes message) arrives. Since the blockpool needs to follow the canonical chain of the best peer at all times, only the best peer can add block hashes. If this is not the case, AddBlockHashes returns without effect.
AddBlocks is called by the eth protocol when a blocksMsg (blocks message) arrives. The various blocks are requested from multiple peers therefore they are accepted from any peer. The peer is recorded on the pool node as the source of the block, this makes it possible to assign an error to the peer in case the block is invalid.
Once a peer is promoted as best peer (total difficulty, and current chain head block registered) a head section process is started, which first requests from the peer the head block itself. Once the head block is received, blockhashes starting from the peer's head block are requested from the best peer.
Once a response is received (and the protocol calls AddBlockHashes), the sequence of block hashes in the response from the best peer are used to build up a sequence nodes replicating the head section of the peer's canonical chain. If the peer fails to respond to requests, after a period of blockTimeout, an ErrInsufficientChainInfo error is raised. As a consequence, the peer is disconnected and suspended for PeerSuspensionInterval during which it is not allowed to reconnect.
Once the head section nodes are set up, the blockpool starts requesting blocks for that section. The requests are distributed among multiple peers so that fetching is optimised.
If the root block of a section is received, then we can connect a section to its parent section (the root block's parent is the parent section's head block). We can repeat requesting blockhashes for the parent section since now we have a way to tell if they arrived.
Once a batch of hashes is received, the node skeleton for the new section is built, and a process similar to the head section process is repeated in a somewhat simplified form. To recap, each section runs its own parallel process with 2 main objective:
- get hashes for its parent section, and
- obtain the blocks in the section. This is achieved by
- requesting hashes starting from the root (bottom) node of the section
- requesting blocks based on the hashes in the section
Block requests are distributed among connected peers to optimise bandwidth utilisation.
With this recursive strategy, a chain is getting built from young to old blocks section by section. This process is repeated until a known block is reached. If the known block is found in the blockchain, the descendent blocks in the chain can be inserted to the blockchain. If the block is known to the blockpool, i.e., it is found in a known section, then the peer is registered with the section's process. This means that the section is requesting and receiving blocks because the section is part of the best peer's canonical chain. If block insertion reveals an invalid block, its source (may not be online any more)
is given a ErrInvalidBlock error resulting in disconnect and suspension.
If a block process does not complete within a set period of time blockTimeout, the chain is killed and the synchronisation is reattempted with (potentially) new peers. Note that these timeouts are needed to protect against attacks where a rogue peer is sending random blockhashes indefinitely.
The interface of the Blockpool with the core is defined with the help of 4 entry points. These are specified as parameters to the blockpool constructor.
Initial block validation that does not require the block to have a known (already processed and valid) parent block. Soft proof of work validation is such a step.
This is used as a first line of defence: when received, blocks are verified, which by putting a cost to make a peer accept your block, protects against simple spamming. If PoW verification fails the sender peer receives an ErrInvalidPoW error. As a consequence, it is disconnected and not allowed to reconnect for PeerSuspensionInterval
Add one or more blocks on top of a known block. See ChainManager.InsertChain The chain manager runs the vm and does proper block validation as well as establishes which block has the highest total difficulty defining the head of the node. If the chain manager finds a block invalid, the peer that supplied the block receives an ErrInvalidBlock error and as a consequence, gets disconnected and suspended for PeerSuspensionInterval. After successfully inserting blocks, if the block is or was the head block of a peer, the blockpool also checks if the block's actual total difficulty is identical to one advertised by the peer. If it is not, the peer received an ErrIncorrectTD and as a consequence gets disconnected and suspended for PeerSuspensionInterval. This protects against rogue peers advertising a high total difficulty and forcing us to follow their (potentially non-canonical) chain.
When receiving block hashes and blocks, we need to check whether the block is already in the blockchain. See ChainManager.HasBlock.
Subscription to new block event, to help set and reset current total difficulty of our head block. This is needed to filter out candidate best peers that are behind and therefore useless.
- maximise bandwidth utilisation by running parallel section processes fetching blocks
- peer switch: section processes that are part of a stale fork (not on the canonical chain of the new best peer) are put to idle mode, i.e., parent section hash requests and block requests are not sent, missing blocks are not checked. Only the absolute deadline timer is active on this process.
- resilience to quick peers switches when several competing miners are connected
- section process caching: if a registered active peer is promoted best peer not for the first time during its connection, all non-contiguous series of sections are activated (they may or may not have been part of the canonical chain of the previous best peer)
- In order to help optimise mining, blocks are inserted into the blockchain the earliest possible time (not waiting for the entire section to be processed for instance).
The parameters are not currently optimised or even tried against other settings. This includes batch size, section length, timeouts, and block request distribution strategy.
When distributing block requests, we could in principle be smarter and do not request blocks from peers with a known lower difficulty. It is unclear how certain we can be of peers state, since they could have caught up synchronising since we received any information about their head.
In case of a network outage, the timeouts are still ticking, if the connection is intermittent, this could cause a lot of repeated work rebuilding the same sections over and over.
The Blockpool uses these errors that are meant to be 'assigned' to peers.
-
ErrInvalidBlock: Invalid block. Block insertion fails. (Fatal) -
ErrInvalidPoW: Invalid PoW. Light block verification fails. (Fatal) -
ErrUnrequestedBlock: Unrequested block. Error when there is an attempt to add a block to the pool, but there is no skeleton node for the blockhash. -
ErrInsufficientChainInfo: Insufficient chain info. This error is raised if the best peer fails to provide the block for their advertised current block hash or fail to provide a sequence of ancestor hashes of which the head section is build up. (Fatal) -
ErrIdleTooLong: Idle too long. This error is raised if the best peer does not send a new block message after an idle period. (Fatal) -
ErrIncorrectTD: Incorrect ID. Raised when the peer is found to advertise incorrect TD for their head block. (Fatal)
Fatal errors lead to the peer getting disconnected and suspended for a period of PeerSuspensionInterval for the duration of which they are not allowed to reconnect.