EIP-7251: Stricter bound for consolidation queue

[dapplion]

MaxEB is the first feature to splice SSZ lists. This operator is expensive w.r.t to hashing and memory de-duplication. I propose to add more stricter bounds to the consolidation queue. Ensuring that the consolidations queue is as small as possible will significantly help those drawbacks.

A consolidation must remain in the queue at least MIN_VALIDATOR_WITHDRAWABILITY_DELAY + 1 + MAX_SEED_LOOKAHEAD. Additional time beyond this minimum is to cover the churn. Consolidations can wait for the churn either included inside the state or in the operation mem-pool. This PR forces consolidations to sit in the mem-pool if the churn exceeds the minimum delay.

Assuming a max supply of 130M ETH, all staked, the worst case for the consolidation queue with 16 ETH consolidations is 256 * 130000000/65536 / 16 = 31738, so if the assumptions of this PR are okay, we can reduce the max size of the list from the current value of 262144. In current mainnet params this change will likely result in a queue size of a few thousands during high demands of consolidation

[hwwhww]

all the state.pending_* = state.pending_*[next_index:] assignment concerned me too. agree that implementing a method to manage the frequency would be optimal.

[dapplion]

all the state.pending_* = state.pending_*[next_index:] assignment concerned me too. agree that implementing a method to manage the frequency would be optimal.

Let's review the other two:

• pending_balance_deposits: unbounded, but not much we can do as all deposits must be accounted. Maybe we can only process deposits every N epochs.
• pending_partial_withdrawals: should be bounded by 7002's exponential fee, but it's mutated every slot. Worth it to model the worst case

[fradamt]

all the state.pending_* = state.pending_*[next_index:] assignment concerned me too. agree that implementing a method to manage the frequency would be optimal.

Let's review the other two:

• pending_balance_deposits: unbounded, but not much we can do as all deposits must be accounted. Maybe we can only process deposits every N epochs.

We can also keep processing them every epoch but only clean up the already processed ones every N epochs

[fradamt]

I don't mind this approach, but I see the criterion of "churn exceeding minimum delay" as somewhat arbitrary. I don't see why we couldn't have epoch = get_current_epoch(state) + MAX_EPOCHS_CONSOLIDATION_QUEUE for any reasonable value of the latter.

[ralexstokes]

have we considered using a (small) ring buffer for these? would avoid any concerns around a "splicing" operation shuffling data around and resulting implications for hashing overhead

[dapplion]

have we considered using a (small) ring buffer for these? would avoid any concerns around a "splicing" operation shuffling data around and resulting implications for hashing overhead

You have to size the ring buffer to a sufficiently forward compatible big size, which would be inefficient as serialized size and you need to track the pointer.