Files
lighthouse/beacon_node/store/src/state_cache.rs
Lion - dapplion dd98534158 Hierarchical state diffs in hot DB (#6750)
This PR implements https://github.com/sigp/lighthouse/pull/5978 (tree-states) but on the hot DB. It allows Lighthouse to massively reduce its disk footprint during non-finality and overall I/O in all cases.

Closes https://github.com/sigp/lighthouse/issues/6580

Conga into https://github.com/sigp/lighthouse/pull/6744

### TODOs

- [x] Fix OOM in CI https://github.com/sigp/lighthouse/pull/7176
- [x] optimise store_hot_state to avoid storing a duplicate state if the summary already exists (should be safe from races now that pruning is cleaner)
- [x] mispelled: get_ancenstor_state_root
- [x] get_ancestor_state_root should use state summaries
- [x] Prevent split from changing during ancestor calc
- [x] Use same hierarchy for hot and cold

### TODO Good optimization for future PRs

- [ ] On the migration, if the latest hot snapshot is aligned with the cold snapshot migrate the diffs instead of the full states.
```
align slot  time
10485760    Nov-26-2024
12582912    Sep-14-2025
14680064    Jul-02-2026
```

### TODO Maybe things good to have

- [ ] Rename anchor_slot https://github.com/sigp/lighthouse/compare/tree-states-hot-rebase-oom...dapplion:lighthouse:tree-states-hot-anchor-slot-rename?expand=1
- [ ] Make anchor fields not public such that they must be mutated through a method. To prevent un-wanted changes of the anchor_slot

### NOTTODO

- [ ] Use fork-choice and a new method [`descendants_of_checkpoint`](ca2388e196 (diff-046fbdb517ca16b80e4464c2c824cf001a74a0a94ac0065e635768ac391062a8)) to filter only the state summaries that descend of finalized checkpoint]
2025-06-19 02:43:25 +00:00

505 lines
18 KiB
Rust

use crate::hdiff::HDiffBuffer;
use crate::{
metrics::{self, HOT_METRIC},
Error,
};
use lru::LruCache;
use std::collections::{BTreeMap, HashMap, HashSet};
use std::num::NonZeroUsize;
use types::{BeaconState, ChainSpec, Epoch, EthSpec, Hash256, Slot};
/// Fraction of the LRU cache to leave intact during culling.
const CULL_EXEMPT_NUMERATOR: usize = 1;
const CULL_EXEMPT_DENOMINATOR: usize = 10;
/// States that are less than or equal to this many epochs old *could* become finalized and will not
/// be culled from the cache.
const EPOCH_FINALIZATION_LIMIT: u64 = 4;
#[derive(Debug)]
pub struct FinalizedState<E: EthSpec> {
state_root: Hash256,
state: BeaconState<E>,
}
/// Map from block_root -> slot -> state_root.
#[derive(Debug, Default)]
pub struct BlockMap {
blocks: HashMap<Hash256, SlotMap>,
}
/// Map from slot -> state_root.
#[derive(Debug, Default)]
pub struct SlotMap {
slots: BTreeMap<Slot, Hash256>,
}
#[derive(Debug)]
pub struct StateCache<E: EthSpec> {
finalized_state: Option<FinalizedState<E>>,
// Stores the tuple (state_root, state) as LruCache only returns the value on put and we need
// the state_root
states: LruCache<Hash256, (Hash256, BeaconState<E>)>,
block_map: BlockMap,
hdiff_buffers: HotHDiffBufferCache,
max_epoch: Epoch,
head_block_root: Hash256,
headroom: NonZeroUsize,
}
/// Cache of hdiff buffers for hot states.
///
/// This cache only keeps buffers prior to the finalized state, which are required by the
/// hierarchical state diff scheme to construct newer unfinalized states.
///
/// The cache always retains the hdiff buffer for the most recent snapshot so that even if the
/// cache capacity is 1, this snapshot never needs to be loaded from disk.
#[derive(Debug)]
pub struct HotHDiffBufferCache {
/// Cache of HDiffBuffers for states *prior* to the `finalized_state`.
///
/// Maps state_root -> (slot, buffer).
hdiff_buffers: LruCache<Hash256, (Slot, HDiffBuffer)>,
}
#[derive(Debug)]
pub enum PutStateOutcome {
/// State is prior to the cache's finalized state (lower slot) and was cached as an HDiffBuffer.
PreFinalizedHDiffBuffer,
/// State is equal to the cache's finalized state and was not inserted.
Finalized,
/// State was already present in the cache.
Duplicate,
/// State is new to the cache and was inserted.
///
/// Includes deleted states as a result of this insertion.
New(Vec<Hash256>),
}
#[allow(clippy::len_without_is_empty)]
impl<E: EthSpec> StateCache<E> {
pub fn new(
state_capacity: NonZeroUsize,
headroom: NonZeroUsize,
hdiff_capacity: NonZeroUsize,
) -> Self {
StateCache {
finalized_state: None,
states: LruCache::new(state_capacity),
block_map: BlockMap::default(),
hdiff_buffers: HotHDiffBufferCache::new(hdiff_capacity),
max_epoch: Epoch::new(0),
head_block_root: Hash256::ZERO,
headroom,
}
}
pub fn len(&self) -> usize {
self.states.len()
}
pub fn capacity(&self) -> usize {
self.states.cap().get()
}
pub fn num_hdiff_buffers(&self) -> usize {
self.hdiff_buffers.len()
}
pub fn hdiff_buffer_mem_usage(&self) -> usize {
self.hdiff_buffers.mem_usage()
}
pub fn update_finalized_state(
&mut self,
state_root: Hash256,
block_root: Hash256,
state: BeaconState<E>,
pre_finalized_slots_to_retain: &[Slot],
) -> Result<(), Error> {
if state.slot() % E::slots_per_epoch() != 0 {
return Err(Error::FinalizedStateUnaligned);
}
if self
.finalized_state
.as_ref()
.is_some_and(|finalized_state| state.slot() < finalized_state.state.slot())
{
return Err(Error::FinalizedStateDecreasingSlot);
}
// Add to block map.
self.block_map.insert(block_root, state.slot(), state_root);
// Prune block map.
let state_roots_to_prune = self.block_map.prune(state.slot());
// Prune HDiffBuffers that are no longer required by the hdiff grid of the finalized state.
// We need to do this prior to copying in any new hdiff buffers, because the cache
// preferences older slots.
// NOTE: This isn't perfect as it prunes by slot: there could be multiple buffers
// at some slots in the case of long forks without finality.
let new_hdiff_cache = HotHDiffBufferCache::new(self.hdiff_buffers.cap());
let old_hdiff_cache = std::mem::replace(&mut self.hdiff_buffers, new_hdiff_cache);
for (state_root, (slot, buffer)) in old_hdiff_cache.hdiff_buffers {
if pre_finalized_slots_to_retain.contains(&slot) {
self.hdiff_buffers.put(state_root, slot, buffer);
}
}
// Delete states.
for state_root in state_roots_to_prune {
if let Some((_, state)) = self.states.pop(&state_root) {
// Add the hdiff buffer for this state to the hdiff cache if it is now part of
// the pre-finalized grid. The `put` method will take care of keeping the most
// useful buffers.
let slot = state.slot();
if pre_finalized_slots_to_retain.contains(&slot) {
let hdiff_buffer = HDiffBuffer::from_state(state);
self.hdiff_buffers.put(state_root, slot, hdiff_buffer);
}
}
}
// Update finalized state.
self.finalized_state = Some(FinalizedState { state_root, state });
Ok(())
}
/// Update the state cache's view of the enshrined head block.
///
/// We never prune the unadvanced state for the head block.
pub fn update_head_block_root(&mut self, head_block_root: Hash256) {
self.head_block_root = head_block_root;
}
/// Rebase the given state on the finalized state in order to reduce its memory consumption.
///
/// This function should only be called on states that are likely not to already share tree
/// nodes with the finalized state, e.g. states loaded from disk.
///
/// If the finalized state is not initialized this function is a no-op.
pub fn rebase_on_finalized(
&self,
state: &mut BeaconState<E>,
spec: &ChainSpec,
) -> Result<(), Error> {
if let Some(finalized_state) = &self.finalized_state {
state.rebase_on(&finalized_state.state, spec)?;
}
Ok(())
}
/// Return a status indicating whether the state already existed in the cache.
pub fn put_state(
&mut self,
state_root: Hash256,
block_root: Hash256,
state: &BeaconState<E>,
) -> Result<PutStateOutcome, Error> {
if let Some(ref finalized_state) = self.finalized_state {
if finalized_state.state_root == state_root {
return Ok(PutStateOutcome::Finalized);
} else if state.slot() <= finalized_state.state.slot() {
// We assume any state being inserted into the cache is grid-aligned (it is the
// caller's responsibility to not feed us garbage) as we don't want to thread the
// hierarchy config through here. So any state received is converted to an
// HDiffBuffer and saved.
let hdiff_buffer = HDiffBuffer::from_state(state.clone());
self.hdiff_buffers
.put(state_root, state.slot(), hdiff_buffer);
return Ok(PutStateOutcome::PreFinalizedHDiffBuffer);
}
}
if self.states.peek(&state_root).is_some() {
return Ok(PutStateOutcome::Duplicate);
}
// Refuse states with pending mutations: we want cached states to be as small as possible
// i.e. stored entirely as a binary merkle tree with no updates overlaid.
if state.has_pending_mutations() {
return Err(Error::StateForCacheHasPendingUpdates {
state_root,
slot: state.slot(),
});
}
// Update the cache's idea of the max epoch.
self.max_epoch = std::cmp::max(state.current_epoch(), self.max_epoch);
// If the cache is full, use the custom cull routine to make room.
let mut deleted_states =
if let Some(over_capacity) = self.len().checked_sub(self.capacity()) {
// The `over_capacity` should always be 0, but we add it here just in case.
self.cull(over_capacity + self.headroom.get())
} else {
vec![]
};
// Insert the full state into the cache.
if let Some((deleted_state_root, _)) =
self.states.put(state_root, (state_root, state.clone()))
{
deleted_states.push(deleted_state_root);
}
// Record the connection from block root and slot to this state.
let slot = state.slot();
self.block_map.insert(block_root, slot, state_root);
Ok(PutStateOutcome::New(deleted_states))
}
pub fn get_by_state_root(&mut self, state_root: Hash256) -> Option<BeaconState<E>> {
if let Some(ref finalized_state) = self.finalized_state {
if state_root == finalized_state.state_root {
return Some(finalized_state.state.clone());
}
}
self.states.get(&state_root).map(|(_, state)| state.clone())
}
pub fn put_hdiff_buffer(&mut self, state_root: Hash256, slot: Slot, buffer: &HDiffBuffer) {
// Only accept HDiffBuffers prior to finalization. Later states should be stored as proper
// states, not HDiffBuffers.
if let Some(finalized_state) = &self.finalized_state {
if slot >= finalized_state.state.slot() {
return;
}
}
self.hdiff_buffers.put(state_root, slot, buffer.clone());
}
pub fn get_hdiff_buffer_by_state_root(&mut self, state_root: Hash256) -> Option<HDiffBuffer> {
if let Some(buffer) = self.hdiff_buffers.get(&state_root) {
metrics::inc_counter_vec(&metrics::STORE_BEACON_HDIFF_BUFFER_CACHE_HIT, HOT_METRIC);
let timer =
metrics::start_timer_vec(&metrics::BEACON_HDIFF_BUFFER_CLONE_TIME, HOT_METRIC);
let result = Some(buffer.clone());
drop(timer);
return result;
}
if let Some(buffer) = self
.get_by_state_root(state_root)
.map(HDiffBuffer::from_state)
{
metrics::inc_counter_vec(&metrics::STORE_BEACON_HDIFF_BUFFER_CACHE_HIT, HOT_METRIC);
return Some(buffer);
}
metrics::inc_counter_vec(&metrics::STORE_BEACON_HDIFF_BUFFER_CACHE_MISS, HOT_METRIC);
None
}
pub fn get_by_block_root(
&mut self,
block_root: Hash256,
slot: Slot,
) -> Option<(Hash256, BeaconState<E>)> {
let slot_map = self.block_map.blocks.get(&block_root)?;
// Find the state at `slot`, or failing that the most recent ancestor.
let state_root = slot_map
.slots
.iter()
.rev()
.find_map(|(ancestor_slot, state_root)| {
(*ancestor_slot <= slot).then_some(*state_root)
})?;
let state = self.get_by_state_root(state_root)?;
Some((state_root, state))
}
pub fn delete_state(&mut self, state_root: &Hash256) {
self.states.pop(state_root);
self.block_map.delete(state_root);
}
pub fn delete_block_states(&mut self, block_root: &Hash256) {
if let Some(slot_map) = self.block_map.delete_block_states(block_root) {
for state_root in slot_map.slots.values() {
self.states.pop(state_root);
}
}
}
/// Cull approximately `count` states from the cache.
///
/// States are culled LRU, with the following extra order imposed:
///
/// - Advanced states.
/// - Mid-epoch unadvanced states.
/// - Epoch-boundary states that are too old to be finalized.
/// - Epoch-boundary states that could be finalized.
pub fn cull(&mut self, count: usize) -> Vec<Hash256> {
let cull_exempt = std::cmp::max(
1,
self.len() * CULL_EXEMPT_NUMERATOR / CULL_EXEMPT_DENOMINATOR,
);
// Stage 1: gather states to cull.
let mut advanced_state_roots = vec![];
let mut mid_epoch_state_roots = vec![];
let mut old_boundary_state_roots = vec![];
let mut good_boundary_state_roots = vec![];
for (&state_root, (_, state)) in self.states.iter().skip(cull_exempt) {
let is_advanced = state.slot() > state.latest_block_header().slot;
let is_boundary = state.slot() % E::slots_per_epoch() == 0;
let could_finalize =
(self.max_epoch - state.current_epoch()) <= EPOCH_FINALIZATION_LIMIT;
if is_boundary {
if could_finalize {
good_boundary_state_roots.push(state_root);
} else {
old_boundary_state_roots.push(state_root);
}
} else if is_advanced {
advanced_state_roots.push(state_root);
} else if state.get_latest_block_root(state_root) != self.head_block_root {
// Never prune the head state
mid_epoch_state_roots.push(state_root);
}
// Terminate early in the common case where we've already found enough junk to cull.
if advanced_state_roots.len() == count {
break;
}
}
// Stage 2: delete.
// This could probably be more efficient in how it interacts with the block map.
let state_roots_to_delete = advanced_state_roots
.into_iter()
.chain(old_boundary_state_roots)
.chain(mid_epoch_state_roots)
.chain(good_boundary_state_roots)
.take(count)
.collect::<Vec<_>>();
for state_root in &state_roots_to_delete {
self.delete_state(state_root);
}
state_roots_to_delete
}
}
impl BlockMap {
fn insert(&mut self, block_root: Hash256, slot: Slot, state_root: Hash256) {
let slot_map = self.blocks.entry(block_root).or_default();
slot_map.slots.insert(slot, state_root);
}
fn prune(&mut self, finalized_slot: Slot) -> HashSet<Hash256> {
let mut pruned_states = HashSet::new();
self.blocks.retain(|_, slot_map| {
slot_map.slots.retain(|slot, state_root| {
let keep = *slot >= finalized_slot;
if !keep {
pruned_states.insert(*state_root);
}
keep
});
!slot_map.slots.is_empty()
});
pruned_states
}
fn delete(&mut self, state_root_to_delete: &Hash256) {
self.blocks.retain(|_, slot_map| {
slot_map
.slots
.retain(|_, state_root| state_root != state_root_to_delete);
!slot_map.slots.is_empty()
});
}
fn delete_block_states(&mut self, block_root: &Hash256) -> Option<SlotMap> {
self.blocks.remove(block_root)
}
}
impl HotHDiffBufferCache {
pub fn new(capacity: NonZeroUsize) -> Self {
Self {
hdiff_buffers: LruCache::new(capacity),
}
}
pub fn get(&mut self, state_root: &Hash256) -> Option<HDiffBuffer> {
self.hdiff_buffers
.get(state_root)
.map(|(_, buffer)| buffer.clone())
}
/// Put a value in the cache, making room for it if necessary.
///
/// If the value was inserted then `true` is returned.
pub fn put(&mut self, state_root: Hash256, slot: Slot, buffer: HDiffBuffer) -> bool {
// If the cache is not full, simply insert the value.
if self.hdiff_buffers.len() != self.hdiff_buffers.cap().get() {
self.hdiff_buffers.put(state_root, (slot, buffer));
return true;
}
// If the cache is full, it has room for this new entry if:
//
// - The capacity is greater than 1: we can retain the snapshot and the new entry, or
// - The capacity is 1 and the slot of the new entry is older than the min_slot in the
// cache. This is a simplified way of retaining the snapshot in the cache. We don't need
// to worry about inserting/retaining states older than the snapshot because these are
// pruned on finalization and never reinserted.
let Some(min_slot) = self.hdiff_buffers.iter().map(|(_, (slot, _))| *slot).min() else {
// Unreachable: cache is full so should have >0 entries.
return false;
};
if self.hdiff_buffers.cap().get() > 1 || slot < min_slot {
// Remove LRU value. Cache is now at size `cap - 1`.
let Some((removed_state_root, (removed_slot, removed_buffer))) =
self.hdiff_buffers.pop_lru()
else {
// Unreachable: cache is full so should have at least one entry to pop.
return false;
};
// Insert new value. Cache size is now at size `cap`.
self.hdiff_buffers.put(state_root, (slot, buffer));
// If the removed value had the min slot and we didn't intend to replace it (cap=1)
// then we reinsert it.
if removed_slot == min_slot && slot >= min_slot {
self.hdiff_buffers
.put(removed_state_root, (removed_slot, removed_buffer));
}
true
} else {
// No room.
false
}
}
pub fn cap(&self) -> NonZeroUsize {
self.hdiff_buffers.cap()
}
#[allow(clippy::len_without_is_empty)]
pub fn len(&self) -> usize {
self.hdiff_buffers.len()
}
pub fn mem_usage(&self) -> usize {
self.hdiff_buffers
.iter()
.map(|(_, (_, buffer))| buffer.size())
.sum()
}
}