fix: address twoeths' review on cached PTCs by lodekeeper · Pull Request #12 · lodekeeper/lodestar

lodekeeper · 2026-03-18T17:02:00Z

Address twoeths' review feedback on ChainSafe#9047.

Changes

1. Replace `getHeadStateAtSlot` with `getHeadStateAtEpoch` (comments 1, 3, 4)

Gossip only validates current-slot attestations, so epoch-level state is sufficient
getHeadStateAtEpoch reuses checkpoint state cache (cheaper than per-slot regen)
Removes getHeadStateAtSlot entirely to prevent DOS via arbitrary slot lookups

2. Use `epochCtx` for PTC reads (comment 2)

Gossip validation: state.epochCtx.getPayloadTimelinessCommittee(data.slot) returns Uint32Array directly (no state tree traversal)
Block processing: getPayloadTimelinessCommittee() uses epochCtx for current epoch, falls back to state tree previousPtc for epoch boundary case (previous epoch's last slot)
Removes PtcCommitteeView abstraction, uses Uint32Array throughout

3. BLS verification priority (comment 5)

API path: {batchable: true, priority: true} (higher priority, like attestations)
Gossip path: {batchable: true, priority: false}

4. Documentation (comment 6)

Added comments on rotatePayloadTimelinessCommittees calls explaining that payloadTimelinessCommittees are computed in finalProcessEpoch

Verification

pnpm build ✅
pnpm lint ✅
pnpm check-types ✅ (all 19 packages)
pnpm vitest run upgradeState.test.ts ✅

- Use getHeadStateAtEpoch instead of getHeadStateAtSlot for gossip validation (cheaper, avoids per-slot state regen) - Remove getHeadStateAtSlot to prevent DOS via arbitrary slot lookups - Use epochCtx.getPayloadTimelinessCommittee() for PTC reads in gossip and current-epoch block processing (avoids state tree traversal) - Keep state tree fallback for previous-epoch PTC at epoch boundaries - Add batchable + priority BLS verification (API gets priority) - Add documentation comments on PTC rotation in state transition - Remove PtcCommitteeView type, use Uint32Array directly Co-authored-by: Nico Flaig <nflaig@users.noreply.github.com>

- Remove compute_ptc#gloas exception (function was renamed to get_ptc in spec) - Remove compute_ptc#gloas entry from functions.yml (no longer in gloas spec) - Update get_ptc#gloas spec hash and body to match v1.7.0-alpha.2 pyspec - Add get_ptc_assignment#gloas entry from new spec function

nflaig · 2026-03-18T17:15:16Z

packages/state-transition/src/util/gloas.ts

  for (let i = 0; i < ptc.length; i++) {
    if (payloadAttestation.aggregationBits.get(i)) {
-      attestingIndices.push(ptc.get(i));
+      attestingIndices.push(ptc[i]);


can we change this to use intersectValues again? see ChainSafe#9047 (comment)

Done — switched back to intersectValues, matches the pattern used everywhere else in the codebase.

nflaig · 2026-03-18T17:15:41Z

packages/state-transition/src/util/gloas.ts

+export function getPtcCommitteeIndex(ptc: Uint32Array, validatorIndex: ValidatorIndex): number {
+  return ptc.indexOf(validatorIndex);


this fucntion can be remove, no reason to wrap ptc.indexOf

Removed — callers now use ptc.indexOf(validatorIndex) directly.

nflaig · 2026-03-18T17:16:53Z

packages/beacon-node/src/chain/validation/payloadAttestationMessage.ts

  payloadAttestationMessage: gloas.PayloadAttestationMessage
 ): Promise<PayloadAttestationValidationResult> {
-  return validatePayloadAttestationMessage(chain, payloadAttestationMessage);
+  return validatePayloadAttestationMessage(chain, payloadAttestationMessage, true);


can you pass this as opts or assign a const before function call for readablity, having a magic boolean valeu like this that just says true/false is not that readable

see how we do it in other places like htis

Fixed — API path uses const prioritizeBls = true, gossip path omits the argument (defaults to false).

nflaig · 2026-03-18T17:19:39Z

packages/state-transition/src/util/gloas.ts

+    const ptc = new Uint32Array(state.previousPtc.length);
+    for (let i = 0; i < ptc.length; i++) {
+      ptc[i] = state.previousPtc.get(i);
+    }
+    return ptc;


isn't it better to use getAll() here, please see which approach is more performance cpu/memory wise

Switched to Uint32Array.from(state.previousPtc.getAll()) — single allocation, no manual loop.

nflaig · 2026-03-18T17:20:16Z

specrefs/functions.yml

please ignore specrefs for now, we need a new spec release for this first, revert changes to this file

Reverted — specrefs restored to base branch state.

- Use named const for prioritizeBls (readability) - Use getAll() for previousPtc fallback instead of manual loop - Remove getPtcCommitteeIndex wrapper (just use ptc.indexOf directly) - Use intersectValues for getIndexedPayloadAttestation (matches existing pattern) - Revert specrefs changes (needs new spec release first)

Add previousEpochLastSlotPtc to epochCache to eliminate mixed access pattern (epochCtx + state tree). At epoch transition, the last slot's PTC is saved before overwriting. This is the only previous-epoch PTC ever needed (processPayloadAttestation enforces data.slot + 1 === state.slot). Now all PTC reads go through epochCtx.getPayloadTimelinessCommittee() - current epoch: payloadTimelinessCommittees[slot % SLOTS_PER_EPOCH] - previous epoch last slot: previousEpochLastSlotPtc - no state tree traversal needed for reads

nflaig · 2026-03-18T17:37:31Z

packages/state-transition/src/util/gloas.ts

+export function getPayloadTimelinessCommittee(state: CachedBeaconStateGloas, slot: Slot): Uint32Array {
+  return state.epochCtx.getPayloadTimelinessCommittee(slot);


no need for getPayloadTimelinessCommittee now as it's just a simple wrapper function

Removed — callers now use state.epochCtx.getPayloadTimelinessCommittee(slot) directly.

nflaig · 2026-03-18T17:38:11Z

packages/state-transition/src/cache/epochCache.ts


-    return this.payloadTimelinessCommittees[slot % SLOTS_PER_EPOCH];
+    // Previous epoch: only the last slot is cached (for epoch-boundary payload attestation processing)
+    if (epoch === this.epoch - 1 && slot % SLOTS_PER_EPOCH === SLOTS_PER_EPOCH - 1 && this.previousEpochLastSlotPtc) {


this.previousEpochLastSlotPtc should never be undefined, we need to initialize it from state when we create the cache

Fixed — previousEpochLastSlotPtc is now initialized from state.previousPtc.getAll() in createFromState, so it's always available after checkpoint sync / restarts.

nflaig · 2026-03-18T17:38:46Z

packages/state-transition/src/cache/epochCache.ts

  // PTC for current epoch, computed eagerly at epoch transition
  payloadTimelinessCommittees: Uint32Array[];
+  // PTC for last slot of previous epoch, needed at epoch boundary for payload attestation processing
+  previousEpochLastSlotPtc: Uint32Array | null;


Suggested change

previousEpochLastSlotPtc: Uint32Array | null;

previousEpochLastSlotPtc: Uint32Array;

Done — changed to Uint32Array (non-nullable). Pre-gloas epochs get an empty new Uint32Array(0).

…t_ptc_assignment compute_ptc was renamed to get_ptc in pyspec v1.7.0-alpha.2. The stale compute_ptc#gloas entry caused check-specrefs to fail. Also updates get_ptc spec body (slot parameter added) and adds get_ptc_assignment#gloas (sources: [] — not yet implemented in TS).

- Remove getPayloadTimelinessCommittee wrapper (callers use epochCtx directly) - Initialize previousEpochLastSlotPtc from state.previousPtc in createFromState - Change type to non-nullable Uint32Array (always initialized)

…ates

nflaig · 2026-03-18T18:41:50Z

packages/beacon-node/src/chain/validation/payloadAttestationMessage.ts

+  const state = (await chain.getHeadStateAtEpoch(
+    epoch,
    RegenCaller.validateGossipPayloadAttestation
  )) as CachedBeaconStateGloas;


as CachedBeaconStateGloas can be removed

Removed — epochCtx is available on all cached states, no Gloas-specific cast needed.

nflaig · 2026-03-18T18:42:25Z

packages/state-transition/src/cache/epochCache.ts

+      previousEpochLastSlotPtc:
+        currentEpoch >= config.GLOAS_FORK_EPOCH
+          ? Uint32Array.from((state as CachedBeaconStateGloas).previousPtc.getAll())
+          : new Uint32Array(0),


ah maybe allowing it to be nullable was better

Reverted to Uint32Array | null, initialized as null in createFromState. afterProcessEpoch populates it at epoch transition before it's needed.

…tPtc - Remove 'as CachedBeaconStateGloas' cast (epochCtx available on all forks) - Revert previousEpochLastSlotPtc to nullable (Uint32Array | null) - Remove state-init in createFromState (afterProcessEpoch handles it)

compute_ptc was renamed to get_ptc in pyspec v1.7.0-alpha.2. The stale compute_ptc#gloas exception entry caused ethspecify 0.3.9 to error on load (invalid key validation). Changes: - Remove stale compute_ptc#gloas from exceptions and functions.yml - Align get_ptc#gloas spec body + hash with unstable base (ae15f761) - Align get_ptc#gloas sources to epochCache.ts (matches unstable) - Add get_ptc_assignment#gloas to exceptions (matches unstable) All spec refs now pass ethspecify process + check locally.

nflaig · 2026-03-27T19:12:00Z

@lodekeeper feel free to close this, I have closed the PR from the target branch since a different approach was merged on the spec side

lodekeeper · 2026-03-27T19:13:49Z

Closing as requested — the target branch (nflaig/cached-ptcs) has been closed on the spec side since a different approach was merged. Thanks!

lodekeeper and others added 2 commits March 18, 2026 17:01

nflaig reviewed Mar 18, 2026

View reviewed changes

lodekeeper added 2 commits March 18, 2026 17:23

nflaig reviewed Mar 18, 2026

View reviewed changes

lodekeeper added 4 commits March 18, 2026 17:43

fix: remove wrapper, init previousEpochLastSlotPtc from state

f630ba2

- Remove getPayloadTimelinessCommittee wrapper (callers use epochCtx directly) - Initialize previousEpochLastSlotPtc from state.previousPtc in createFromState - Change type to non-nullable Uint32Array (always initialized)

fix: revert specrefs to base branch (needs new spec release)

b3dde34

docs: clarify previousEpochLastSlotPtc init invariant on mid-epoch st…

990c980

…ates

nflaig reviewed Mar 18, 2026

View reviewed changes

lodekeeper added 2 commits March 18, 2026 18:46

lodekeeper closed this Mar 27, 2026

		export function getPtcCommitteeIndex(ptc: Uint32Array, validatorIndex: ValidatorIndex): number {
		return ptc.indexOf(validatorIndex);

		export function getPayloadTimelinessCommittee(state: CachedBeaconStateGloas, slot: Slot): Uint32Array {
		return state.epochCtx.getPayloadTimelinessCommittee(slot);

	previousEpochLastSlotPtc: Uint32Array \| null;
	previousEpochLastSlotPtc: Uint32Array;

Conversation

lodekeeper commented Mar 18, 2026

Changes

1. Replace getHeadStateAtSlot with getHeadStateAtEpoch (comments 1, 3, 4)

2. Use epochCtx for PTC reads (comment 2)

3. BLS verification priority (comment 5)

4. Documentation (comment 6)

Verification

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nflaig commented Mar 27, 2026

Uh oh!

lodekeeper commented Mar 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

1. Replace `getHeadStateAtSlot` with `getHeadStateAtEpoch` (comments 1, 3, 4)

2. Use `epochCtx` for PTC reads (comment 2)