We received a bug report of a node restarting custody backfill unnecessarily after upgrading to Lighthouse v8.1.1. What happened is:
- User started LH v8.0.1 many months ago, CGC updated 0 -> N but the CGC was not eagerly persisted.
- LH experienced an unclean shutdown (not sure of what type).
- Upon restarting (still running v8.0.1), the custody context read from disk contains CGC=0: `DEBUG Loaded persisted custody context custody_context: CustodyContext { validator_custody_count: 0, ...`).
- CGC updates again to N, retriggering custody backfill: `DEBUG Validator count at head updated old_count: 0, new_count: N`.
- Custody backfill does a bunch of downloading for no gain: `DEBUG Imported historical data columns epoch: Epoch(428433), total_imported: 0`
- While custody backfill is running user updated to v8.1.1, and we see logs for the CGC=N being peristed upon clean shutdown, and then correctly read on startup with v8.1.1.
- Custody backfill keeps running and downloading due to the CGC change still being considered in progress.
- Call `persist_custody_context` inside the `register_validators` handler so that it is written to disk eagerly whenever it changes. The performance impact of this should be minimal as the amount of data is very small and this call can only happen at most ~128 times (once for each change) in the entire life of a beacon node.
- Call `persist_custody_context` inside `BeaconChainBuilder::build` so that changes caused by CLI flags are persisted (otherwise starting a node with `--semi-supernode` and no validators, then shutting it down uncleanly would cause use to forget the CGC).
These changes greatly reduce the timespan during which an unclean shutdown can create inconsistency. In the worst case, we only lose backfill progress that runs concurrently with the `register_validators` handler (should be extremely minimal, nigh impossible).
Co-Authored-By: Michael Sproul <michael@sigmaprime.io>
