Add AttesterCache for attestation production (#2478)

## Issue Addressed

- Resolves #2169

## Proposed Changes

Adds the `AttesterCache` to allow validators to produce attestations for older slots. Presently, some arbitrary restrictions can force validators to receive an error when attesting to a slot earlier than the present one. This can cause attestation misses when there is excessive load on the validator client or time sync issues between the VC and BN.

## Additional Info

NA

beacon_node/beacon_chain/src/attester_cache.rs

@@ -0,0 +1,377 @@
+//! This module provides the `AttesterCache`, a cache designed for reducing state-reads when
+//! validators produce `AttestationData`.
+//!
+//! This cache is required *as well as* the `ShufflingCache` since the `ShufflingCache` does not
+//! provide any information about the `state.current_justified_checkpoint`. It is not trivial to add
+//! the justified checkpoint to the `ShufflingCache` since that cache is keyed by shuffling decision
+//! root, which is not suitable for the justified checkpoint. Whilst we can know the shuffling for
+//! epoch `n` during `n - 1`, we *cannot* know the justified checkpoint. Instead, we *must* perform
+//! `per_epoch_processing` to transform the state from epoch `n - 1` to epoch `n` so that rewards
+//! and penalties can be computed and the `state.current_justified_checkpoint` can be updated.
+
+use crate::{BeaconChain, BeaconChainError, BeaconChainTypes};
+use parking_lot::RwLock;
+use state_processing::state_advance::{partial_state_advance, Error as StateAdvanceError};
+use std::collections::HashMap;
+use std::ops::Range;
+use types::{
+    beacon_state::{
+        compute_committee_index_in_epoch, compute_committee_range_in_epoch, epoch_committee_count,
+    },
+    BeaconState, BeaconStateError, ChainSpec, Checkpoint, Epoch, EthSpec, Hash256, RelativeEpoch,
+    Slot,
+};
+
+type JustifiedCheckpoint = Checkpoint;
+type CommitteeLength = usize;
+type CommitteeIndex = u64;
+type CacheHashMap = HashMap<AttesterCacheKey, AttesterCacheValue>;
+
+/// The maximum number of `AttesterCacheValues` to be kept in memory.
+///
+/// Each `AttesterCacheValues` is very small (~16 bytes) and the cache will generally be kept small
+/// by pruning on finality.
+///
+/// The value provided here is much larger than will be used during ideal network conditions,
+/// however we make it large since the values are so small.
+const MAX_CACHE_LEN: usize = 1_024;
+
+#[derive(Debug)]
+pub enum Error {
+    BeaconState(BeaconStateError),
+    // Boxed to avoid an infinite-size recursion issue.
+    BeaconChain(Box<BeaconChainError>),
+    MissingBeaconState(Hash256),
+    FailedToTransitionState(StateAdvanceError),
+    CannotAttestToFutureState {
+        state_slot: Slot,
+        request_slot: Slot,
+    },
+    /// Indicates a cache inconsistency.
+    WrongEpoch {
+        request_epoch: Epoch,
+        epoch: Epoch,
+    },
+    InvalidCommitteeIndex {
+        committee_index: u64,
+    },
+    /// Indicates an inconsistency with the beacon state committees.
+    InverseRange {
+        range: Range<usize>,
+    },
+}
+
+impl From<BeaconStateError> for Error {
+    fn from(e: BeaconStateError) -> Self {
+        Error::BeaconState(e)
+    }
+}
+
+impl From<BeaconChainError> for Error {
+    fn from(e: BeaconChainError) -> Self {
+        Error::BeaconChain(Box::new(e))
+    }
+}
+
+/// Stores the minimal amount of data required to compute the committee length for any committee at any
+/// slot in a given `epoch`.
+struct CommitteeLengths {
+    /// The `epoch` to which the lengths pertain.
+    epoch: Epoch,
+    /// The length of the shuffling in `self.epoch`.
+    active_validator_indices_len: usize,
+}
+
+impl CommitteeLengths {
+    /// Instantiate `Self` using `state.current_epoch()`.
+    fn new<T: EthSpec>(state: &BeaconState<T>, spec: &ChainSpec) -> Result<Self, Error> {
+        let active_validator_indices_len = if let Ok(committee_cache) =
+            state.committee_cache(RelativeEpoch::Current)
+        {
+            committee_cache.active_validator_indices().len()
+        } else {
+            // Building the cache like this avoids taking a mutable reference to `BeaconState`.
+            let committee_cache = state.initialize_committee_cache(state.current_epoch(), spec)?;
+            committee_cache.active_validator_indices().len()
+        };
+
+        Ok(Self {
+            epoch: state.current_epoch(),
+            active_validator_indices_len,
+        })
+    }
+
+    /// Get the length of the committee at the given `slot` and `committee_index`.
+    fn get<T: EthSpec>(
+        &self,
+        slot: Slot,
+        committee_index: CommitteeIndex,
+        spec: &ChainSpec,
+    ) -> Result<CommitteeLength, Error> {
+        let slots_per_epoch = T::slots_per_epoch();
+        let request_epoch = slot.epoch(slots_per_epoch);
+
+        // Sanity check.
+        if request_epoch != self.epoch {
+            return Err(Error::WrongEpoch {
+                request_epoch,
+                epoch: self.epoch,
+            });
+        }
+
+        let slots_per_epoch = slots_per_epoch as usize;
+        let committees_per_slot =
+            T::get_committee_count_per_slot(self.active_validator_indices_len, spec)?;
+        let index_in_epoch = compute_committee_index_in_epoch(
+            slot,
+            slots_per_epoch,
+            committees_per_slot,
+            committee_index as usize,
+        );
+        let range = compute_committee_range_in_epoch(
+            epoch_committee_count(committees_per_slot, slots_per_epoch),
+            index_in_epoch,
+            self.active_validator_indices_len,
+        )
+        .ok_or(Error::InvalidCommitteeIndex { committee_index })?;
+
+        range
+            .end
+            .checked_sub(range.start)
+            .ok_or(Error::InverseRange { range })
+    }
+}
+
+/// Provides the following information for some epoch:
+///
+/// - The `state.current_justified_checkpoint` value.
+/// - The committee lengths for all indices and slots.
+///
+/// These values are used during attestation production.
+pub struct AttesterCacheValue {
+    current_justified_checkpoint: Checkpoint,
+    committee_lengths: CommitteeLengths,
+}
+
+impl AttesterCacheValue {
+    /// Instantiate `Self` using `state.current_epoch()`.
+    pub fn new<T: EthSpec>(state: &BeaconState<T>, spec: &ChainSpec) -> Result<Self, Error> {
+        let current_justified_checkpoint = state.current_justified_checkpoint();
+        let committee_lengths = CommitteeLengths::new(state, spec)?;
+        Ok(Self {
+            current_justified_checkpoint,
+            committee_lengths,
+        })
+    }
+
+    /// Get the justified checkpoint and committee length for some `slot` and `committee_index`.
+    fn get<T: EthSpec>(
+        &self,
+        slot: Slot,
+        committee_index: CommitteeIndex,
+        spec: &ChainSpec,
+    ) -> Result<(JustifiedCheckpoint, CommitteeLength), Error> {
+        self.committee_lengths
+            .get::<T>(slot, committee_index, spec)
+            .map(|committee_length| (self.current_justified_checkpoint, committee_length))
+    }
+}
+
+/// The `AttesterCacheKey` is fundamentally the same thing as the proposer shuffling decision root,
+/// however here we use it as an identity for both of the following values:
+///
+/// 1. The `state.current_justified_checkpoint`.
+/// 2. The attester shuffling.
+///
+/// This struct relies upon the premise that the `state.current_justified_checkpoint` in epoch `n`
+/// is determined by the root of the latest block in epoch `n - 1`. Notably, this is identical to
+/// how the proposer shuffling is keyed in `BeaconProposerCache`.
+///
+/// It is also safe, but not maximally efficient, to key the attester shuffling with the same
+/// strategy. For better shuffling keying strategies, see the `ShufflingCache`.
+#[derive(Eq, PartialEq, Hash, Clone, Copy)]
+pub struct AttesterCacheKey {
+    /// The epoch from which the justified checkpoint should be observed.
+    ///
+    /// Attestations which use `self.epoch` as `target.epoch` should use this key.
+    epoch: Epoch,
+    /// The root of the block at the last slot of `self.epoch - 1`.
+    decision_root: Hash256,
+}
+
+impl AttesterCacheKey {
+    /// Instantiate `Self` to key `state.current_epoch()`.
+    ///
+    /// The `latest_block_root` should be the latest block that has been applied to `state`. This
+    /// parameter is required since the state does not store the block root for any block with the
+    /// same slot as `state.slot()`.
+    ///
+    /// ## Errors
+    ///
+    /// May error if `epoch` is out of the range of `state.block_roots`.
+    pub fn new<T: EthSpec>(
+        epoch: Epoch,
+        state: &BeaconState<T>,
+        latest_block_root: Hash256,
+    ) -> Result<Self, Error> {
+        let slots_per_epoch = T::slots_per_epoch();
+        let decision_slot = epoch.start_slot(slots_per_epoch).saturating_sub(1_u64);
+
+        let decision_root = if decision_slot.epoch(slots_per_epoch) == epoch {
+            // This scenario is only possible during the genesis epoch. In this scenario, all-zeros
+            // is used as an alias to the genesis block.
+            Hash256::zero()
+        } else if epoch > state.current_epoch() {
+            // If the requested epoch is higher than the current epoch, the latest block will always
+            // be the decision root.
+            latest_block_root
+        } else {
+            *state.get_block_root(decision_slot)?
+        };
+
+        Ok(Self {
+            epoch,
+            decision_root,
+        })
+    }
+}
+
+/// Provides a cache for the justified checkpoint and committee length when producing an
+/// attestation.
+///
+/// See the module-level documentation for more information.
+#[derive(Default)]
+pub struct AttesterCache {
+    cache: RwLock<CacheHashMap>,
+}
+
+impl AttesterCache {
+    /// Get the justified checkpoint and committee length for the `slot` and `committee_index` in
+    /// the state identified by the cache `key`.
+    pub fn get<T: EthSpec>(
+        &self,
+        key: &AttesterCacheKey,
+        slot: Slot,
+        committee_index: CommitteeIndex,
+        spec: &ChainSpec,
+    ) -> Result<Option<(JustifiedCheckpoint, CommitteeLength)>, Error> {
+        self.cache
+            .read()
+            .get(key)
+            .map(|cache_item| cache_item.get::<T>(slot, committee_index, spec))
+            .transpose()
+    }
+
+    /// Cache the `state.current_epoch()` values if they are not already present in the state.
+    pub fn maybe_cache_state<T: EthSpec>(
+        &self,
+        state: &BeaconState<T>,
+        latest_block_root: Hash256,
+        spec: &ChainSpec,
+    ) -> Result<(), Error> {
+        let key = AttesterCacheKey::new(state.current_epoch(), state, latest_block_root)?;
+        let mut cache = self.cache.write();
+        if !cache.contains_key(&key) {
+            let cache_item = AttesterCacheValue::new(state, spec)?;
+            Self::insert_respecting_max_len(&mut cache, key, cache_item);
+        }
+        Ok(())
+    }
+
+    /// Read the state identified by `state_root` from the database, advance it to the required
+    /// slot, use it to prime the cache and return the values for the provided `slot` and
+    /// `committee_index`.
+    ///
+    /// ## Notes
+    ///
+    /// This function takes a write-lock on the internal cache. Prefer attempting a `Self::get` call
+    /// before running this function as `Self::get` only takes a read-lock and is therefore less
+    /// likely to create contention.
+    pub fn load_and_cache_state<T: BeaconChainTypes>(
+        &self,
+        state_root: Hash256,
+        key: AttesterCacheKey,
+        slot: Slot,
+        committee_index: CommitteeIndex,
+        chain: &BeaconChain<T>,
+    ) -> Result<(JustifiedCheckpoint, CommitteeLength), Error> {
+        let spec = &chain.spec;
+        let slots_per_epoch = T::EthSpec::slots_per_epoch();
+        let epoch = slot.epoch(slots_per_epoch);
+
+        // Take a write-lock on the cache before starting the state read.
+        //
+        // Whilst holding the write-lock during the state read will create contention, it prevents
+        // the scenario where multiple requests from separate threads cause duplicate state reads.
+        let mut cache = self.cache.write();
+
+        // Try the cache to see if someone has already primed it between the time the function was
+        // called and when the cache write-lock was obtained. This avoids performing duplicate state
+        // reads.
+        if let Some(value) = cache
+            .get(&key)
+            .map(|cache_item| cache_item.get::<T::EthSpec>(slot, committee_index, spec))
+            .transpose()?
+        {
+            return Ok(value);
+        }
+
+        let mut state: BeaconState<T::EthSpec> = chain
+            .get_state(&state_root, None)?
+            .ok_or(Error::MissingBeaconState(state_root))?;
+
+        if state.slot() > slot {
+            // This indicates an internal inconsistency.
+            return Err(Error::CannotAttestToFutureState {
+                state_slot: state.slot(),
+                request_slot: slot,
+            });
+        } else if state.current_epoch() < epoch {
+            // Only perform a "partial" state advance since we do not require the state roots to be
+            // accurate.
+            partial_state_advance(
+                &mut state,
+                Some(state_root),
+                epoch.start_slot(slots_per_epoch),
+                spec,
+            )
+            .map_err(Error::FailedToTransitionState)?;
+            state.build_committee_cache(RelativeEpoch::Current, spec)?;
+        }
+
+        let cache_item = AttesterCacheValue::new(&state, spec)?;
+        let value = cache_item.get::<T::EthSpec>(slot, committee_index, spec)?;
+        Self::insert_respecting_max_len(&mut cache, key, cache_item);
+        Ok(value)
+    }
+
+    /// Insert a value to `cache`, ensuring it does not exceed the maximum length.
+    ///
+    /// If the cache is already full, the item with the lowest epoch will be removed.
+    fn insert_respecting_max_len(
+        cache: &mut CacheHashMap,
+        key: AttesterCacheKey,
+        value: AttesterCacheValue,
+    ) {
+        while cache.len() >= MAX_CACHE_LEN {
+            if let Some(oldest) = cache
+                .iter()
+                .map(|(key, _)| *key)
+                .min_by_key(|key| key.epoch)
+            {
+                cache.remove(&oldest);
+            } else {
+                break;
+            }
+        }
+
+        cache.insert(key, value);
+    }
+
+    /// Remove all entries where the `key.epoch` is lower than the given `epoch`.
+    ///
+    /// Generally, the provided `epoch` should be the finalized epoch.
+    pub fn prune_below(&self, epoch: Epoch) {
+        self.cache.write().retain(|target, _| target.epoch >= epoch);
+    }
+}

beacon_node/beacon_chain/src/beacon_chain.rs

@@ -2,6 +2,7 @@ use crate::attestation_verification::{
     Error as AttestationError, SignatureVerifiedAttestation, VerifiedAggregatedAttestation,
     VerifiedUnaggregatedAttestation,
 };
+use crate::attester_cache::{AttesterCache, AttesterCacheKey};
 use crate::beacon_proposer_cache::BeaconProposerCache;
 use crate::block_verification::{
     check_block_is_finalized_descendant, check_block_relevancy, get_block_root,
@@ -289,6 +290,8 @@ pub struct BeaconChain<T: BeaconChainTypes> {
     pub beacon_proposer_cache: Mutex<BeaconProposerCache>,
     /// Caches a map of `validator_index -> validator_pubkey`.
     pub(crate) validator_pubkey_cache: TimeoutRwLock<ValidatorPubkeyCache<T>>,
+    /// A cache used when producing attestations.
+    pub(crate) attester_cache: Arc<AttesterCache>,
     /// A list of any hard-coded forks that have been disabled.
     pub disabled_forks: Vec<String>,
     /// Sender given to tasks, so that if they encounter a state in which execution cannot
@@ -1217,44 +1220,174 @@ impl<T: BeaconChainTypes> BeaconChain<T> {
     /// validator that is in the committee for `slot` and `index` in the canonical chain.
     ///
     /// Always attests to the canonical chain.
+    ///
+    /// ## Errors
+    ///
+    /// May return an error if the `request_slot` is too far behind the head state.
     pub fn produce_unaggregated_attestation(
         &self,
-        slot: Slot,
-        index: CommitteeIndex,
+        request_slot: Slot,
+        request_index: CommitteeIndex,
     ) -> Result<Attestation<T::EthSpec>, Error> {
-        // Note: we're taking a lock on the head. The work involved here should be trivial enough
-        // that the lock should not be held for long.
-        let head = self
-            .canonical_head
-            .try_read_for(HEAD_LOCK_TIMEOUT)
-            .ok_or(Error::CanonicalHeadLockTimeout)?;
+        let _total_timer = metrics::start_timer(&metrics::ATTESTATION_PRODUCTION_SECONDS);
 
-        if slot >= head.beacon_block.slot() {
-            self.produce_unaggregated_attestation_for_block(
-                slot,
-                index,
-                head.beacon_block_root,
-                Cow::Borrowed(&head.beacon_state),
-                head.beacon_state_root(),
-            )
+        let slots_per_epoch = T::EthSpec::slots_per_epoch();
+        let request_epoch = request_slot.epoch(slots_per_epoch);
+
+        /*
+         * Phase 1/2:
+         *
+         * Take a short-lived read-lock on the head and copy the necessary information from it.
+         *
+         * It is important that this first phase is as quick as possible; creating contention for
+         * the head-lock is not desirable.
+         */
+
+        let head_state_slot;
+        let beacon_block_root;
+        let beacon_state_root;
+        let target;
+        let current_epoch_attesting_info: Option<(Checkpoint, usize)>;
+        let attester_cache_key;
+        let head_timer = metrics::start_timer(&metrics::ATTESTATION_PRODUCTION_HEAD_SCRAPE_SECONDS);
+        if let Some(head) = self.canonical_head.try_read_for(HEAD_LOCK_TIMEOUT) {
+            let head_state = &head.beacon_state;
+            head_state_slot = head_state.slot();
+
+            // There is no value in producing an attestation to a block that is pre-finalization and
+            // it is likely to cause expensive and pointless reads to the freezer database. Exit
+            // early if this is the case.
+            let finalized_slot = head_state
+                .finalized_checkpoint()
+                .epoch
+                .start_slot(slots_per_epoch);
+            if request_slot < finalized_slot {
+                return Err(Error::AttestingToFinalizedSlot {
+                    finalized_slot,
+                    request_slot,
+                });
+            }
+
+            // This function will eventually fail when trying to access a slot which is
+            // out-of-bounds of `state.block_roots`. This explicit error is intended to provide a
+            // clearer message to the user than an ambiguous `SlotOutOfBounds` error.
+            let slots_per_historical_root = T::EthSpec::slots_per_historical_root() as u64;
+            let lowest_permissible_slot =
+                head_state.slot().saturating_sub(slots_per_historical_root);
+            if request_slot < lowest_permissible_slot {
+                return Err(Error::AttestingToAncientSlot {
+                    lowest_permissible_slot,
+                    request_slot,
+                });
+            }
+
+            if request_slot >= head_state.slot() {
+                // When attesting to the head slot or later, always use the head of the chain.
+                beacon_block_root = head.beacon_block_root;
+                beacon_state_root = head.beacon_state_root();
+            } else {
+                // Permit attesting to slots *prior* to the current head. This is desirable when
+                // the VC and BN are out-of-sync due to time issues or overloading.
+                beacon_block_root = *head_state.get_block_root(request_slot)?;
+                beacon_state_root = *head_state.get_state_root(request_slot)?;
+            };
+
+            let target_slot = request_epoch.start_slot(T::EthSpec::slots_per_epoch());
+            let target_root = if head_state.slot() <= target_slot {
+                // If the state is earlier than the target slot then the target *must* be the head
+                // block root.
+                beacon_block_root
+            } else {
+                *head_state.get_block_root(target_slot)?
+            };
+            target = Checkpoint {
+                epoch: request_epoch,
+                root: target_root,
+            };
+
+            current_epoch_attesting_info = if head_state.current_epoch() == request_epoch {
+                // When the head state is in the same epoch as the request, all the information
+                // required to attest is available on the head state.
+                Some((
+                    head_state.current_justified_checkpoint(),
+                    head_state
+                        .get_beacon_committee(request_slot, request_index)?
+                        .committee
+                        .len(),
+                ))
+            } else {
+                // If the head state is in a *different* epoch to the request, more work is required
+                // to determine the justified checkpoint and committee length.
+                None
+            };
+
+            // Determine the key for `self.attester_cache`, in case it is required later in this
+            // routine.
+            attester_cache_key =
+                AttesterCacheKey::new(request_epoch, head_state, beacon_block_root)?;
         } else {
-            // We disallow producing attestations *prior* to the current head since such an
-            // attestation would require loading a `BeaconState` from disk. Loading `BeaconState`
-            // from disk is very resource intensive and proposes a DoS risk from validator clients.
-            //
-            // Although we generally allow validator clients to do things that might harm us (i.e.,
-            // we trust them), sometimes we need to protect the BN from accidental errors which
-            // could cause it significant harm.
-            //
-            // This case is particularity harmful since the HTTP API can effectively call this
-            // function an unlimited amount of times. If `n` validators all happen to call it at
-            // the same time, we're going to load `n` states (and tree hash caches) into memory all
-            // at once. With `n >= 10` we're looking at hundreds of MB or GBs of RAM.
-            Err(Error::AttestingPriorToHead {
-                head_slot: head.beacon_block.slot(),
-                request_slot: slot,
-            })
+            return Err(Error::CanonicalHeadLockTimeout);
         }
+        drop(head_timer);
+
+        /*
+         *  Phase 2/2:
+         *
+         *  If the justified checkpoint and committee length from the head are suitable for this
+         *  attestation, use them. If not, try the attester cache. If the cache misses, load a state
+         *  from disk and prime the cache with it.
+         */
+
+        let cache_timer =
+            metrics::start_timer(&metrics::ATTESTATION_PRODUCTION_CACHE_INTERACTION_SECONDS);
+        let (justified_checkpoint, committee_len) =
+            if let Some((justified_checkpoint, committee_len)) = current_epoch_attesting_info {
+                // The head state is in the same epoch as the attestation, so there is no more
+                // required information.
+                (justified_checkpoint, committee_len)
+            } else if let Some(cached_values) = self.attester_cache.get::<T::EthSpec>(
+                &attester_cache_key,
+                request_slot,
+                request_index,
+                &self.spec,
+            )? {
+                // The suitable values were already cached. Return them.
+                cached_values
+            } else {
+                debug!(
+                    self.log,
+                    "Attester cache miss";
+                    "beacon_block_root" => ?beacon_block_root,
+                    "head_state_slot" => %head_state_slot,
+                    "request_slot" => %request_slot,
+                );
+
+                // Neither the head state, nor the attester cache was able to produce the required
+                // information to attest in this epoch. So, load a `BeaconState` from disk and use
+                // it to fulfil the request (and prime the cache to avoid this next time).
+                let _cache_build_timer =
+                    metrics::start_timer(&metrics::ATTESTATION_PRODUCTION_CACHE_PRIME_SECONDS);
+                self.attester_cache.load_and_cache_state(
+                    beacon_state_root,
+                    attester_cache_key,
+                    request_slot,
+                    request_index,
+                    &self,
+                )?
+            };
+        drop(cache_timer);
+
+        Ok(Attestation {
+            aggregation_bits: BitList::with_capacity(committee_len)?,
+            data: AttestationData {
+                slot: request_slot,
+                index: request_index,
+                beacon_block_root,
+                source: justified_checkpoint,
+                target,
+            },
+            signature: AggregateSignature::empty(),
+        })
     }
 
     /// Produces an "unaggregated" attestation for the given `slot` and `index` that attests to
@@ -2023,6 +2156,7 @@ impl<T: BeaconChainTypes> BeaconChain<T> {
         let block_root = fully_verified_block.block_root;
         let mut state = fully_verified_block.state;
         let current_slot = self.slot()?;
+        let current_epoch = current_slot.epoch(T::EthSpec::slots_per_epoch());
         let mut ops = fully_verified_block.confirmation_db_batch;
 
         let attestation_observation_timer =
@@ -2086,6 +2220,17 @@ impl<T: BeaconChainTypes> BeaconChain<T> {
             }
         }
 
+        // Apply the state to the attester cache, only if it is from the previous epoch or later.
+        //
+        // In a perfect scenario there should be no need to add previous-epoch states to the cache.
+        // However, latency between the VC and the BN might cause the VC to produce attestations at
+        // a previous slot.
+        if state.current_epoch().saturating_add(1_u64) >= current_epoch {
+            self.attester_cache
+                .maybe_cache_state(&state, block_root, &self.spec)
+                .map_err(BeaconChainError::from)?;
+        }
+
         let mut fork_choice = self.fork_choice.write();
 
         // Do not import a block that doesn't descend from the finalized root.
@@ -2936,6 +3081,9 @@ impl<T: BeaconChainTypes> BeaconChain<T> {
             self.head_tracker.clone(),
         )?;
 
+        self.attester_cache
+            .prune_below(new_finalized_checkpoint.epoch);
+
         if let Some(event_handler) = self.event_handler.as_ref() {
             if event_handler.has_finalized_subscribers() {
                 event_handler.register(EventKind::FinalizedCheckpoint(SseFinalizedCheckpoint {

beacon_node/beacon_chain/src/builder.rs

@@ -547,6 +547,7 @@ where
             shuffling_cache: TimeoutRwLock::new(ShufflingCache::new()),
             beacon_proposer_cache: <_>::default(),
             validator_pubkey_cache: TimeoutRwLock::new(validator_pubkey_cache),
+            attester_cache: <_>::default(),
             disabled_forks: self.disabled_forks,
             shutdown_sender: self
                 .shutdown_sender
@@ -561,6 +562,16 @@ where
             .head()
             .map_err(|e| format!("Failed to get head: {:?}", e))?;
 
+        // Prime the attester cache with the head state.
+        beacon_chain
+            .attester_cache
+            .maybe_cache_state(
+                &head.beacon_state,
+                head.beacon_block_root,
+                &beacon_chain.spec,
+            )
+            .map_err(|e| format!("Failed to prime attester cache: {:?}", e))?;
+
         // Only perform the check if it was configured.
         if let Some(wss_checkpoint) = beacon_chain.config.weak_subjectivity_checkpoint {
             if let Err(e) = beacon_chain.verify_weak_subjectivity_checkpoint(

beacon_node/beacon_chain/src/errors.rs

@@ -1,3 +1,4 @@
+use crate::attester_cache::Error as AttesterCacheError;
 use crate::beacon_chain::ForkChoiceError;
 use crate::beacon_fork_choice_store::Error as ForkChoiceStoreError;
 use crate::eth1_chain::Error as Eth1ChainError;
@@ -91,6 +92,7 @@ pub enum BeaconChainError {
     ObservedAttestationsError(ObservedAttestationsError),
     ObservedAttestersError(ObservedAttestersError),
     ObservedBlockProducersError(ObservedBlockProducersError),
+    AttesterCacheError(AttesterCacheError),
     PruningError(PruningError),
     ArithError(ArithError),
     InvalidShufflingId {
@@ -99,8 +101,12 @@ pub enum BeaconChainError {
     },
     WeakSubjectivtyVerificationFailure,
     WeakSubjectivtyShutdownError(TrySendError<ShutdownReason>),
-    AttestingPriorToHead {
-        head_slot: Slot,
+    AttestingToFinalizedSlot {
+        finalized_slot: Slot,
+        request_slot: Slot,
+    },
+    AttestingToAncientSlot {
+        lowest_permissible_slot: Slot,
         request_slot: Slot,
     },
     BadPreState {
@@ -137,6 +143,7 @@ easy_from_to!(NaiveAggregationError, BeaconChainError);
 easy_from_to!(ObservedAttestationsError, BeaconChainError);
 easy_from_to!(ObservedAttestersError, BeaconChainError);
 easy_from_to!(ObservedBlockProducersError, BeaconChainError);
+easy_from_to!(AttesterCacheError, BeaconChainError);
 easy_from_to!(BlockSignatureVerifierError, BeaconChainError);
 easy_from_to!(PruningError, BeaconChainError);
 easy_from_to!(ArithError, BeaconChainError);

beacon_node/beacon_chain/src/lib.rs

@@ -1,5 +1,6 @@
 #![recursion_limit = "128"] // For lazy-static
 pub mod attestation_verification;
+mod attester_cache;
 mod beacon_chain;
 mod beacon_fork_choice_store;
 mod beacon_proposer_cache;

beacon_node/beacon_chain/src/metrics.rs

@@ -210,18 +210,22 @@ lazy_static! {
     /*
      * Attestation Production
      */
-    pub static ref ATTESTATION_PRODUCTION_REQUESTS: Result<IntCounter> = try_create_int_counter(
-        "beacon_attestation_production_requests_total",
-        "Count of all attestation production requests"
-    );
-    pub static ref ATTESTATION_PRODUCTION_SUCCESSES: Result<IntCounter> = try_create_int_counter(
-        "beacon_attestation_production_successes_total",
-        "Count of attestations processed without error"
-    );
-    pub static ref ATTESTATION_PRODUCTION_TIMES: Result<Histogram> = try_create_histogram(
+    pub static ref ATTESTATION_PRODUCTION_SECONDS: Result<Histogram> = try_create_histogram(
         "beacon_attestation_production_seconds",
         "Full runtime of attestation production"
     );
+    pub static ref ATTESTATION_PRODUCTION_HEAD_SCRAPE_SECONDS: Result<Histogram> = try_create_histogram(
+        "attestation_production_head_scrape_seconds",
+        "Time taken to read the head state"
+    );
+    pub static ref ATTESTATION_PRODUCTION_CACHE_INTERACTION_SECONDS: Result<Histogram> = try_create_histogram(
+        "attestation_production_cache_interaction_seconds",
+        "Time spent interacting with the attester cache"
+    );
+    pub static ref ATTESTATION_PRODUCTION_CACHE_PRIME_SECONDS: Result<Histogram> = try_create_histogram(
+        "attestation_production_cache_prime_seconds",
+        "Time spent loading a new state from the disk due to a cache miss"
+    );
 }
 
 // Second lazy-static block is used to account for macro recursion limit.

beacon_node/beacon_chain/src/state_advance_timer.rs

@@ -321,6 +321,12 @@ fn advance_head<T: BeaconChainTypes>(
         );
     }
 
+    // Apply the state to the attester cache, if the cache deems it interesting.
+    beacon_chain
+        .attester_cache
+        .maybe_cache_state(&state, head_root, &beacon_chain.spec)
+        .map_err(BeaconChainError::from)?;
+
     let final_slot = state.slot();
 
     // Insert the advanced state back into the snapshot cache.

consensus/types/src/beacon_state.rs

@@ -21,7 +21,10 @@ use test_random_derive::TestRandom;
 use tree_hash::TreeHash;
 use tree_hash_derive::TreeHash;
 
-pub use self::committee_cache::CommitteeCache;
+pub use self::committee_cache::{
+    compute_committee_index_in_epoch, compute_committee_range_in_epoch, epoch_committee_count,
+    CommitteeCache,
+};
 pub use clone_config::CloneConfig;
 pub use eth_spec::*;
 pub use iter::BlockRootsIter;
@@ -1310,10 +1313,22 @@ impl<T: EthSpec> BeaconState<T> {
         let epoch = relative_epoch.into_epoch(self.current_epoch());
         let i = Self::committee_cache_index(relative_epoch);
 
-        *self.committee_cache_at_index_mut(i)? = CommitteeCache::initialized(&self, epoch, spec)?;
+        *self.committee_cache_at_index_mut(i)? = self.initialize_committee_cache(epoch, spec)?;
         Ok(())
     }
 
+    /// Initializes a new committee cache for the given `epoch`, regardless of whether one already
+    /// exists. Returns the committee cache without attaching it to `self`.
+    ///
+    /// To build a cache and store it on `self`, use `Self::build_committee_cache`.
+    pub fn initialize_committee_cache(
+        &self,
+        epoch: Epoch,
+        spec: &ChainSpec,
+    ) -> Result<CommitteeCache, Error> {
+        CommitteeCache::initialized(&self, epoch, spec)
+    }
+
     /// Advances the cache for this state into the next epoch.
     ///
     /// This should be used if the `slot` of this state is advanced beyond an epoch boundary.

consensus/types/src/beacon_state/committee_cache.rs

@@ -121,8 +121,12 @@ impl CommitteeCache {
             return None;
         }
 
-        let committee_index =
-            (slot.as_u64() % self.slots_per_epoch) * self.committees_per_slot + index;
+        let committee_index = compute_committee_index_in_epoch(
+            slot,
+            self.slots_per_epoch as usize,
+            self.committees_per_slot as usize,
+            index as usize,
+        );
         let committee = self.compute_committee(committee_index as usize)?;
 
         Some(BeaconCommittee {
@@ -219,7 +223,10 @@ impl CommitteeCache {
     ///
     /// Spec v0.12.1
     pub fn epoch_committee_count(&self) -> usize {
-        self.committees_per_slot as usize * self.slots_per_epoch as usize
+        epoch_committee_count(
+            self.committees_per_slot as usize,
+            self.slots_per_epoch as usize,
+        )
     }
 
     /// Returns the number of committees per slot for this cache's epoch.
@@ -242,16 +249,7 @@ impl CommitteeCache {
     ///
     /// Spec v0.12.1
     fn compute_committee_range(&self, index: usize) -> Option<Range<usize>> {
-        let count = self.epoch_committee_count();
-        if count == 0 || index >= count {
-            return None;
-        }
-
-        let num_validators = self.shuffling.len();
-        let start = (num_validators * index) / count;
-        let end = (num_validators * (index + 1)) / count;
-
-        Some(start..end)
+        compute_committee_range_in_epoch(self.epoch_committee_count(), index, self.shuffling.len())
     }
 
     /// Returns the index of some validator in `self.shuffling`.
@@ -264,6 +262,44 @@ impl CommitteeCache {
     }
 }
 
+/// Computes the position of the given `committee_index` with respect to all committees in the
+/// epoch.
+///
+/// The return result may be used to provide input to the `compute_committee_range_in_epoch`
+/// function.
+pub fn compute_committee_index_in_epoch(
+    slot: Slot,
+    slots_per_epoch: usize,
+    committees_per_slot: usize,
+    committee_index: usize,
+) -> usize {
+    (slot.as_usize() % slots_per_epoch) * committees_per_slot + committee_index
+}
+
+/// Computes the range for slicing the shuffled indices to determine the members of a committee.
+///
+/// The `index_in_epoch` parameter can be computed computed using
+/// `compute_committee_index_in_epoch`.
+pub fn compute_committee_range_in_epoch(
+    epoch_committee_count: usize,
+    index_in_epoch: usize,
+    shuffling_len: usize,
+) -> Option<Range<usize>> {
+    if epoch_committee_count == 0 || index_in_epoch >= epoch_committee_count {
+        return None;
+    }
+
+    let start = (shuffling_len * index_in_epoch) / epoch_committee_count;
+    let end = (shuffling_len * (index_in_epoch + 1)) / epoch_committee_count;
+
+    Some(start..end)
+}
+
+/// Returns the total number of committees in an epoch.
+pub fn epoch_committee_count(committees_per_slot: usize, slots_per_epoch: usize) -> usize {
+    committees_per_slot * slots_per_epoch
+}
+
 /// Returns a list of all `validators` indices where the validator is active at the given
 /// `epoch`.
 ///