Isthmus: Header Accumulator

[clabby]

Overview

Specifies an addition to block execution in the Holocene hardfork, which designates "header accumulation" blocks and an addition to the extraData field in "header accumulation" blocks' headers to enable this functionality.

TODO

• Explicit block validity rules.

Possible Ideas

• [✅] Rather than the receipts batches, why not store header batches? Could be useful for efficient historical state lookups, i.e. in the fault proof program. No less efficient to store the broader commitment, either. h/t @protolambda

[clabby]

Curious reviewers' thoughts. Should we also append the header batch root to this? Small bit extra data, though makes it so that we wouldn't need to reconstruct the batch tree off-chain for finding leaves of the accumulator tree. It should be a quick process (with a reasonable HEADER_BATCH_SIZE), and they'll need the full batch tree for inclusion proofs in the batch, though might be nice to have a quick lookup for batch roots as well.

[axelKingsley]

What's the size of the redundant data being attached? And would we do any validation that the batch-root included matches what is calculated?

[clabby]

If we did add this, it would be an extra 32 bytes per HEADER_BATCH_SIZE blocks in the extraData field. We would need to add a validation step, where we verify that the leaf (the header batch tree root) is verified to be included within the tree that the merkle stack represents.

[clabby]

• Holocene: Add L2ToL1MessagePasser account storage root to Header withdrawalsRoot #177
• Isthmus: Header Accumulator #259 👈
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

Join @clabby and the rest of your teammates on Graphite

[axelKingsley]

Our Design Doc Template has "Alternatives Considered" and "Risks and Uncertainties" sections which I've found really helpful in the past for my own understanding of the problem space.

Since we're in specs, maybe it doesn't apply the same way, but in lieu of that, a couple questions:

Seems like the problem we're trying to solve is the impending spike in compute requirements when there are many interoperating superchains that all require header-reading. It'd be more efficient to compute the headers all as a batch and to only emit/expect them once per batch-window.
Does this header-batch-window size inform the resolution at which interop can run? Seems like while you wait for the next header batch, you can't make use of any of the other blocks beyond doing the work (which is what we're trying to avoid).

My other question is around the selection of the Merkle Tree. Seems a good choice for compact storage, but this doc also points out that it only requires two merkle proofs as well as _n_ merkle patricia trie proofs to verify inclusion. This seems like an important feature, but I don't fully understand how that mechanism is used by nodes to achieve the goal of only a single L2 block at a height that includes _all receipts_ on a single chain that must be validated to resolve its dependents. Would you mind sketching that out a bit? This could be something very obvious to folks more familiar with Merkle Tree than I am.

[axelKingsley]

Suggested change

	OP Stack's state transition will increase. The current plan for proving the inclusion of remote logs within chains
	OP Stack's state transition will multiply(?). The current plan for proving the inclusion of remote logs within chains

Might be in this doc farther down, but what is the relationship number-of-chains, total gas, and compute required? I'm thinking it's linear with the total gas of all chains, but not sure.

[clabby]

I am not sure either, or at least not sure enough to give it a concrete time-complexity.

From what I do know, it will multiply, with a minima of linear complexity increase (scaling with # of chains in the dependency set). Though, due to us having to reproduce the L2 block for every block that a remote log is included in, the computational complexity also grows with the number of relayed messages in the current frame (note: I'm using "frame" here as a non-official term; What I mean is the "frame" in which we're resolving cross-chain dependencies.)

[ajsutton]

The maximum we have to execute in a single FP VM run is a single L2 block (from a single chain). We can extend the output root bisection game to avoid having to execute a block from every L2 chain in the set within a single VM execution. So the cost to having additional chains is pretty negligible. The issue is actually just how old a log can be referenced by an executing message because we have to walk back one block at a time (by hash) to retrieve that referenced log which gets expensive if it's a long way in the past.

[axelKingsley]

This comes out to a 32 leaf nodes, and I think that means a 32x performance optimization, is that correct? As the header concerns now happen only once per HEADER_BATCH_SIZE?

Do any behaviors change if we find that this needs to be increased further? Like, if we had 64 per batch, would we start to put pressure on other parts of the system in negative ways?

[clabby]

Yes, that is the optimization. HEADER_BATCH_SIZEx less data stored in extraData (since extraData is only expanded once every HEADER_BATCH_SIZE blocks)

I would advise not making HEADER_BATCH_SIZE > 256, since most execution layer clients drop state > 256 blocks old out of memory. With HEADER_BATCH_SIZE <= 256, the lookups for the header hashes should be very quick, since we can directly consult in-memory state at tip.

[axelKingsley]

What's the size of the redundant data being attached? And would we do any validation that the batch-root included matches what is calculated?

[clabby]

@axelKingsley

My other question is around the selection of the Merkle Tree. Seems a good choice for compact storage, but this doc also points out that it only requires two merkle proofs as well as n merkle patricia trie proofs to verify inclusion. This seems like an important feature, but I don't fully understand how that mechanism is used by nodes to achieve the goal of only a single L2 block at a height that includes all receipts on a single chain that must be validated to resolve its dependents. Would you mind sketching that out a bit? This could be something very obvious to folks more familiar with Merkle Tree than I am.

This choice was made to allow for efficient batching of data, and to reduce historical state expansion. By splitting the incremental tree, and including batch roots as leaves, we have a very nice property where only every HEADER_BATCH_SIZE blocks, we include a new merkle stack. In addition, we also reduce the size of the merkle stack in the extraData field by 5 * 20 bytes for every "accumulation block" (block.number % HEADER_BATCH_SIZE == 0), saving quite a bit of storage costs over time.

[ajsutton]

I'm having a bit of a hard time picturing how this winds up being used in practice to verify that a log exists on L2. It would be really useful to have a worked example showing how you get from a L2 block hash, back to a receipt from a historic block where all data is accessed by getting the preimage of a hash. Or I guess we could have a new PreimageOracle data type that verifies a provided merkle tree so you can supply a single proof.

[ajsutton]

We don't have to derive multiple L2 blocks in a single execution - we break that down as part of the modified output root bisection so that we are only executing a single L2 block at a time. The final consolidation step requires checking if the cross-chain dependencies are met by checking if a referenced log exists for each executing message. "Fetching" the referenced log message uses the same approach as for fetching L1 logs - we can walk backwards from the chain head via parentHash to the block, then down the receiptHash trie and get the log by hash.

So we don't need to execute any L2 blocks as part of this step, but currently we have to fetch each header back to the block the claimed log is from which, if the required block is quite old can be very expensive.

[clabby]

Good to know! There are currently no specifications for fault proofs post interop, but the output bisection modification makes sense to reduce derivation + execution work. Can we please get this written down somewhere in the specs?

"Fetching" the referenced log message uses the same approach as for fetching L1 logs - we can walk backwards from the chain head via parentHash to the block, then down the receiptHash trie and get the log by hash.

The proposal seeks to optimize this walkback on L2 during the consolidation step by making the lookup constant time, rather than scaling linearly with the depth of the state relative to tip. Especially for messages that are relayed far after they are originally sent. Interop, as far as I understand it, places no limitations on how old a message can be on a remote chain in order for it to be relayed. Relay protection is on the consumer of the relayed logs, meaning we still may have to prove relayed logs that are incredibly deep in historical state. Is this right?

[ajsutton]

Fault proving interop isn't happening for a while, but I made some notes in https://www.notion.so/oplabs/External-Interop-Fault-Dispute-Game-Notes-1537bf9fad054bcfb2245dea88d48d16?pvs=4 - that links to the current draft PR of the specs but it will need some changes. The key thing though is that applying blocks from multiple chains in a single cannon run is never going to work so we definitely won't be doing that. :)

There is a limit on the log you can reference, currently it's 180 days which is far too long to be fault provable without something like this: https://specs.optimism.io/interop/messaging.html#messaging-invariants

[ajsutton]

The maximum we have to execute in a single FP VM run is a single L2 block (from a single chain). We can extend the output root bisection game to avoid having to execute a block from every L2 chain in the set within a single VM execution. So the cost to having additional chains is pretty negligible. The issue is actually just how old a log can be referenced by an executing message because we have to walk back one block at a time (by hash) to retrieve that referenced log which gets expensive if it's a long way in the past.

[ajsutton]

I don't know of any place where we access historical information for the L2 chain - only for L1 which won't have this change.

[clabby]

We don't currently, but post-interop, we will have to in order to verify log inclusion for relayed logs where there was a significant delay between the log's emission on chain A and the log being relayed on chain B.

We also do have the walkback on L2 today in the program, though it's not very bad due to output bisection. When fetching PayloadByNumber, etc., we must walk back from the starting output root (the current safe head) when the program is initialized.

[ajsutton]

Yeah this will definitely help with interop, just the way I read this section it sounded like it was saying it would help with other exisiting behaviour for fault proofs today, but I don't believe there is anything that would benefit. The agreed starting block from the output root bisection is the unsafe, safe and finalized head initially so the L2 walkback just exits the loop immediately because it has already reached the finalized block.

[ajsutton]

One thing to note is that if we only have a truncated commitment, we can't use it to retrieve data from the preimage oracle - we need the full hash for that.

[clabby]

We can add a new PreimageKeyType for this, yes? Should be the same as the keccak route, just need to truncate the resulting digest.

[ajsutton]

Yeh I think that should work. This is where I think having those details fleshed out would be really helpful - we don't want to implement this spec and then find there's something that's suboptimal for actually using it in fault proofs, so it's worth making sure we're actually thinking through all the details of how it will be used.

[BlocksOnAChain]

As agreed on the ENG staff meeting, I'm assigning this to @protolambda as lead implementer that will own this side of work for Granite hardfork.