Add missing block and receipt message fields #9
Conversation
The response to `CurrentBlockHeaderRequest{}` is the peer's local
current header. This can be used to get a rough idea of what the latest
block is as viewed by other peers on the network. The header can be
compared with a subset of peers to determine an efficient way of pulling
blocks from the network while syncing.
Initially, we modified the iteration message to include a `latest`
option, however, since the iteration message is shared by multiple
requests it didn't make sense for each of the messages to have access to
it.
The block header is updated to allow for the P2P spec to be compatible with the RPC spec. Before these changes, p2p spec was not enough to serve to block RPC requests. Block bodies include Transactions and Receipts because RPC requires the block to have transactions and receipts. A peer could retrieve them in a separate request, however, since block bodies are retrieved over multiple messages I don't see much benefit in requesting them separately. Also, note Events were added to Receipt common because `BlockID` is not enough to determine which event belongs to which receipt. We require the events for the bloom filter.
IronGauntlets
requested review from
ittaysw,
Mirko-von-Leipzig,
CHr15F0x and
omerfirmak
p2p/proto/block.proto
Outdated
| StateDiff diff = 2; | ||
| Classes classes = 3; | ||
| BlockProof proof = 4; | ||
| Transactions transactions = 6; |
There was a problem hiding this comment.
transactoins and events were not supposed to be part of the block body. this can be discussed. the idea was that with stark proofs we don't need to validate transactions by executing them. This is when there's proof which is K blocks behind. Until then the idea was to rely on the consensus (similarly to how Ethereum has block times and epochs).
The protocol can be defined so that txs are mandatory until K and then are optional.
Similarly for receipts & events (which were separated).
minor comment: numbering jumps from 4 to 6
There was a problem hiding this comment.
Are all forms of nodes expected to serve the RPC spec? If so, then having transactions and receipts as optional until after K blocks seems unnecessary because all nodes would be fetching them separately to serve the RPC requests. And since block bodies are chunked I don't see any benefit in fetching them separately as that would increase network overhead.
minor comment: numbering jumps from 4 to 6
I will update
There was a problem hiding this comment.
It means that to have the data needed to sync with the chain, a node doesn't need to keep transactions before the block with the proof. They can clean their DB. If a node wants to, it can of course keep whatever it wants.
As mentioned, they are not part of the block body, so no need for a client to fetch them. Even if they are included for new blocks, they would be optional and with a separate endpoint to fetch them (from those nodes that keep them beyond the optional point)
There was a problem hiding this comment.
What defines whether something should be a part of the block body?
We don't need State Diff or Classes to be part of the block body because State Diff has a commitment which includes the Classes declared in the blocks, therefore, a peer can fetch them separately and verify the state diff commitment with the signature which is over the block hash and state diff commitment.
Note we will have to add BlockID and possibly Classes to State Diff. Also State diff commitment is over the state diff defined by feeder gateway not p2p state diff.
Why can't we include the BlockProof in the BlockHeader? Given other fields are so small 142KB+ other fields' size < 1MB.
If we make the above changes the block bodies can be removed.
There was a problem hiding this comment.
Right, they can fetch them separately. So the thought was what will peers have to fetch in order to be an operational node. They need the current state, they don't need past transactions.
I agree the proof may be part of the header, as optional. Note that we can think of returning many headers in one message (and then it makes sense to have the proof outside)
There was a problem hiding this comment.
So the thought was what will peers have to fetch in order to be an operational node.
Can you please elaborate on what you mean by an operational node and fetch which messages?
Also, fetching can be done in order but there is no guarantee the messages will arrive in order. So I am not sure what you mean.
There was a problem hiding this comment.
I was replying to "... includes the Classes declared in the blocks, therefore, a peer can fetch them separately". Operational means a node that can follow the chain. Such a node doesn't need transactions. If a client wants to check transactions state, they will enable that in a node, if they just want to track events, they will enable that. If they want to check state, they can do that.
What messages won't arrive in order? If we support out of order messages in a stream then each will have a reference to be able to order them, if needed.
There was a problem hiding this comment.
Operational means a node that can follow the chain. Such a node doesn't need transactions. If a client wants to check transactions state, they will enable that in a node, if they just want to track events, they will enable that. If they want to check state, they can do that.
Why is the RPC spec not sufficient to serve this? I don't see the need to replicate this behaviour on the p2p side. Also, how does a peer verify that the events (or any other date) received are correct?
What messages won't arrive in order? If we support out of order messages in a stream then each will have a reference to be able to order them, if needed.
Messages from a single peer will indeed come in order over a stream but libp2p doesn't allow for ordered messages from multiple peers on a single stream without further multiplexing. If a peer requests information for a block from multiple peers that peer will have multiple streams open, per peer, and thus there is no guarantee block messages part will arrive in order.
IronGauntlets
force-pushed
the
IronGauntlets/update-block-spec
branch
from
651dfb0
to
1692596
Compare
| Transactions transactions = 5; | ||
| Receipts receipts = 6; |
There was a problem hiding this comment.
Why should transactions and receipts be included in the response stream?
Why not just use separate queries for them (as in the current spec version)?
There was a problem hiding this comment.
I added the transactions and receipts as part of the block body for completeness as it was represented by the block returned by the feeder gateway. However, you are correct we can fetch them separately.
Based on the same reasoning I don't think we should have a body message at all as we have state commitment which can be used to verify StateDiff and Classes, while the BlockProof can be included in the Header. Also StateDiff, Classes and Block proof are already sent separately just wrapped in a body message.
This also ties in with having multiple protocols instead of one. I may be missing some context but I cannot see under which situation a node would be interested in just transactions/events/receipts/headers etc. How would the node verify the validity of the single message which they retrieved without also retrieving other information such as block header, state diff transactions etc? And why would RPC not be enough (cc:@joshklop)? We can decide to add web sockets for whatever data the user is interested in. In my opinion, we should have a single protocol.
Please see this convo for more context.
Please see commit messages for details.