This PR allows sharing the pubkey data between validators by using a
thread-local cache for pubkey data, netting about a 400mb mem usage
reduction on holesky due to us keeping 3 permanent + several ephemeral
state copies in memory at all times and each state copy holding a full
validator.
The PR also introduces a hash cache for the key which gives ~14% speedup
for a full state `hash_tree_root` - the key makes up for a large part of
the `Validator` htr time.
Finally, the time it takes to copy a state goes down as well from ~80m
ms to ~60, for reasons similar to htr.
We use a `ptr` even if a `ref` could in theory have been used - there is
not much practical benefit to a `ref` (given it's mutable) while a `ptr`
is cheaper and easier to copy (when copying temporary states).
We could go further and cache a cooked pubkey but it turns out this is
quite intrusive - in all the relevant places, we're already using a
cooked key from the immutable validator data so there are no immediate
performance gains of doing so while managing the compressed -> cooked
key mapping would become more difficult - something for a future PR
perhaps.
Co-authored-by: Etan Kissling <etan@status.im>
* track latest duration instead of total in new timing metrics
Change `db_checkpoint_seconds` and `state_replay_seconds` metrics to
record the latest duration instead of the total. `nim-metrics` already
synthesizes a `_total` metric from these implicitly.
* still have to use inc, metrics only synthesizes the name not the sum
* prefix with `beacon_dag`
Validator monitoring gained 2 new metrics for tracking when blocks are
included or not on the head chain.
Similar to attestations, if the block is produced in epoch N, reporting
will use the state when switching to epoch N+2 to do the reporting (so
as to reasonably stabilise the block inclusion in the face of reorgs).
Database checkpointing can take seconds, e.g., while Geth is syncing.
Add a debug log + metric for it, and also info log if it takes longer
than 250ms, same as for the existing `State replayed` log. If the log
shows up for a user while the system is not overloaded, it may point
to slow disk speed or thermal issue.
* compute post-merge randao mix without loading state
* avoid copying state on shuffling computation and compute epochref
* speed up state copy for block production
With checkpoint sync, the checkpoint block is typically unavailable at
the start, and only backfilled later. To avoid treating it as having
zero hash, execution disabled in some contexts, wrap the result of
`loadExecutionBlockHash` in `Opt` and handle block hash being unknown.
---------
Co-authored-by: Jacek Sieka <jacek@status.im>
When syncing, we log a notice each time someone asks us for a block that
we haven't backfilled yet. This is quite verbose and not unexpected,
because the status message does not allow indicating backfill progress.
When using checkpoint sync, only checkpoint state is available, block is
not downloaded and backfilled later.
`dag.backfill` tracks latest filled `slot`, and latest `parent_root` for
which no block has been synced yet.
In checkpoint sync, this assumption is broken, because there, the start
`dag.backfill.slot` is set based on checkpoint state slot, and the block
is also not available.
However, sync manager in backward mode also requests `dag.backfill.slot`
and `block_clearance` then backfills the checkpoint block once it is
synced. But, there is no guarantee that a peer ever sends us that block.
They could send us all parent blocks and solely omit the checkpoint
block itself. In that situation, we would accept the parent blocks and
advance `dag.backfill`, and subsequently never request the checkpoint
block again, resulting in gap inside blocks DB that is never filled.
To mitigate that, the assumption is restored that `dag.backfill.slot`
is the latest filled `slot`, and `dag.backfill.parent_root` is the next
block that needs to be synced. By setting `slot` to `tail.slot + 1` and
`parent_root` to `tail.root`, we put a fake summary into `dag.backfill`
so that `block_clearance` only proceeds once checkpoint block exists.
After checkpoint sync, historical block IDs cannot yet be queried.
However, they are needed to compute dependent roots of `ShufflingRef`.
To allow lookup, enable `getBlockIdAtSlot` to answer from compatible
states in memory; as long as they descend from the finalized checkpoint
and the requested slot is sufficiently recent, `block_roots` contains
everything to recover `BlockSlotId` up to `SLOTS_PER_HISTORICAL_ROOT`.
This is similar to how `attester_dependent_root` etc. are computed.
This accelerates the first couple minutes of checkpoint sync on Mainnet,
especially the time until finality advances past the synced checkpoint.
Finish the rename started in #4809 to have a consistent naming.
`ExecutionPayloadHash` suggests hash over payload instead of block.
`BlockHash` is also the canonical name in engine API.
When using checkpoint sync, the initial block is missing in the DB.
Update the LC data collector initialization to account for that,
avoiding a spurious error message when it is incorrectly accessed:
```
ERR 2024-02-07 11:21:55.416+01:00 Block failed to load unexpectedly topics="chaindag_lc" bid=d30517a7:8257504 tail=8257504
```
Also fixes a regression from #5691 that resulted in similar messages
while importing the first few blocks after checkpoint sync.
Thanks to @arnetheduck for reporting this.
Full caches should not be used to mark blocks as unviable. The unviable
status is quite persistent and a block marked as such won't be processed
again once the cache empties. Problem originally introduced in #4808.
Changes here are more significant because of some good old tech debt in
block production which has grown quite hairy - the reduction in
exception handling at least provides some steps in the right direction.
`eth1_chain` no longer logs with `topics` since #5768, making it hard
to filter messages from this module. Re-add the `topics`, and also fix
outdated `topics` in `el_manager` (formerly `*_monitor`).
When BN clock is out of sync, VC sets BN status to `BrokenClock`. It is
only reset to `Offline` after restoring time sync. However, if VC fails
encounters an error while checking time, Nimbus extensions are assumed
to be unavailable and the BN is no longer checked for having a synced
clock. This means it is never reset back to `Offline` if errors start
occurring _after_ BN is already set to `BrokenClock`. This could be
because BN is changed from Nimbus to an alternative implementation,
or due to intermittent connection issues.
Ensure that BN status is reset back to `Offline` when Nimbus extensions
are disabled to ensure eventual connection recovery.
Jacek Sieka
Etan Kissling
tersec
Zahary Karadjov
Kim De Mey
Eugene Kabanov