An empirical analysis of slot timings on Ethereum
- Collect larger sample of block arrivals and estimate stats per slot -> this will be used for the attestation model
- Analyse blob arrivals and propagation (similar to what we did for blocks)
Here, we are focussed on the overall attestation arrivals, which includes bock propagation, block execution and validation and attestation propagation. Besides being a more realistic scenario than missed slots, we have a wider set of factor to account for late arrivals.
- Collect raw sample of attestations from two types of slots:
- Relay blocks
- Self-build blocks (still need to check how to get this data)
- Join with data on attestor - entity, country, client (using ethseer or validator_metadata?)
- Join with data on block - size, gas used, # transactions block propagation time (avg & p95)
- Analyze late arrivals (i.e. after 4.5s). Are they overrepresented across any attestor feature or block feature?
- Build regression model to predict attestation arrival and compute feature importance.
- Which validators have significantly different timings between missed slots and normal slots? These are likely the non-prysm validators as all the other clients don't wait for the 4s mark to send their attestations.
- Exclude blocks with single transactions using more than 16M gas units -> these blocks take longer on average to execute, and we already have an EIP that will fix the transaction size to 16M.
WIP