Bootstrap and replication fixes

docs/p2p-replication-and-routing-changes-2026-01-13.md

@@ -0,0 +1,224 @@
+# P2P Replication + Routing Eligibility Changes (2026-01-13)
+
+This document describes all  changes in the `https://github.com/LumeraProtocol/supernode/pull/259`  that affect the P2P/Kademlia subsystem, including the replication/backlog fixes and routing-table eligibility gating.
+
+## Context / Symptoms Observed
+
+These changes were prompted by two production symptoms:
+
+1. P2P metrics error (status/telemetry path):
+   - Log: `failed to get p2p records count`
+   - Error: `failed to get count of records: context deadline exceeded`
+   - Typical attributes included `host=<supernode identity>`, `service=<ip:port>`
+
+2. Replication producing huge key payloads (replication worker path):
+   - Log: `replicate batching keys`
+   - Example attributes:
+     - `batch_size=5000`
+     - `keys=15934902`
+     - `batches=3187`
+     - `correlation_id=supernode-start`
+     - `module=p2p`
+     - `rep-id=<peer supernode account>`
+     - `rep-ip=<peer ip>`
+
+Historically, replication had a bug that prevented progress when payloads became oversized; batching fixed request size, but exposed/created a large backlog catch-up scenario where a single replication cycle attempted to send millions of keys.
+
+## High-Level Outcomes
+
+The combined changes aim to:
+
+- Make replication “catch-up” bounded and incremental (steady progress instead of huge one-shot attempts).
+- Reduce replication complexity (avoid per-peer re-scans of the same key set).
+- Advance replication cursors safely (avoid skipping keys when operating in bounded windows).
+- Reduce noisy error logging for best-effort DB metrics under load.
+- Gate routing table membership to chain-active supernodes (avoid non-active peers participating in routing and replication decisions).
+
+## Change Inventory (Files Modified)
+
+The following files are modified:
+
+- `p2p/kademlia/bootstrap.go`
+- `p2p/kademlia/dht.go`
+- `p2p/kademlia/node_activity.go`
+- `p2p/kademlia/redundant_data.go`
+- `p2p/kademlia/replication.go`
+- `p2p/kademlia/store.go`
+- `p2p/kademlia/store/mem/mem.go`
+- `p2p/kademlia/store/sqlite/replication.go`
+- `p2p/kademlia/store/sqlite/sqlite.go`
+
+## Detailed Changes
+
+### 1) Routing table eligibility gating (chain-active only)
+
+#### `p2p/kademlia/bootstrap.go`
+
+**What changed**
+
+- `loadBootstrapCandidatesFromChain(...)` now returns two values:
+  1) `map[string]*Node` candidates keyed by `ip:port`
+  2) `map[[32]byte]struct{}` allowlist of *active* supernode IDs, stored as `blake3(supernodeAccount)` keys
+- The supernode account string is trimmed, validated non-empty, and used consistently as `node.ID`.
+- `SyncBootstrapOnce(...)` now:
+  - calls `setRoutingAllowlist(...)` with the active ID hash-set
+  - calls `pruneIneligibleRoutingPeers(...)` to remove already-admitted peers that are now ineligible
+
+**Why**
+
+- Kademlia routing decisions (closest nodes) should only consider chain-active supernodes. Without this, postponed/disabled/stopped nodes can:
+  - be admitted via inbound traffic
+  - appear in `FindNodeResponse.Closest`
+  - skew replication responsibility calculations
+
+#### `p2p/kademlia/dht.go`
+
+**What changed**
+
+- Added an in-memory allowlist gate to `DHT`:
+  - `routingAllow map[[32]byte]struct{}`
+  - `routingAllowReady atomic.Bool`
+  - `routingAllowCount atomic.Int64`
+  - guarded by `routingAllowMu`
+- New helpers:
+  - `setRoutingAllowlist(ctx, allow)`:
+    - refuses empty allowlists (to avoid locking out due to transient chain issues)
+    - no-ops when `INTEGRATION_TEST=true`
+  - `eligibleForRouting(node)`:
+    - returns true when allowlist isn’t ready (bootstrap safety)
+    - returns true when `INTEGRATION_TEST=true`
+    - otherwise checks `blake3(node.ID)` membership in the allowlist
+  - `filterEligibleNodes(nodes)` filters node slices returned by the network (FindNode/FindValue paths)
+  - `pruneIneligibleRoutingPeers(ctx)` walks the current routing table and removes ineligible nodes
+- Integrated eligibility checks into:
+  - iterative lookup handling: `nl.AddNodes(s.filterEligibleNodes(v.Closest))`
+  - response handling for `FindNode`/`StoreData`/`FindValue`
+  - `addNode(...)`:
+    - early nil/empty-ID returns
+    - invalid IP rejection log level reduced to Debug
+    - chain-state gating: rejects peers not in allowlist
+  - `addKnownNodes(...)`:
+    - skips nil/empty-ID nodes
+    - skips nodes failing `eligibleForRouting(...)`
+
+**Why**
+
+- Prevent non-active peers from entering buckets, appearing in closest-sets, or being considered as replication targets.
+- Keep performance acceptable by keeping gating as a fast memory lookup, updated on the bootstrap refresh cadence.
+
+#### `p2p/kademlia/node_activity.go`
+
+**What changed**
+
+- Node activity checks now skip pinging/promoting nodes that are not eligible for routing.
+- If a node is currently marked Active but becomes ineligible:
+  - it is removed from routing (`removeNode`)
+  - replication info is updated to inactive via `store.UpdateIsActive(..., false, false)`
+- `handlePingSuccess` now also refuses to “promote” an ineligible node to active/routing.
+
+**Why**
+
+- Avoid spending cycles on peers that the chain says should not participate.
+- Avoid “re-activating” a peer solely due to network reachability when it’s ineligible by chain state.
+
+### 2) Replication backlog + oversized payload mitigation (bounded windows + one-pass assignment)
+
+#### Root-cause (before)
+
+- Replication used a time window `[globalFrom, now]` and fetched *all* keys from sqlite in that window.
+- For each peer, it then:
+  - found the slice index for that peer’s `lastReplicatedAt` within the global key list, then
+  - iterated over all remaining keys to filter “keys this peer should hold”.
+- After batching was added, a peer could still be responsible for millions of keys, which became thousands of batch RPCs.
+- Cursor update (`lastReplicatedAt`) happened only after all batches succeeded; a single failure in thousands of batches meant zero progress and repeated attempts next cycle.
+
+#### `p2p/kademlia/store.go`
+
+**What changed**
+
+- Store interface method signature changed:
+  - from: `GetKeysForReplication(ctx, from, to)`
+  - to: `GetKeysForReplication(ctx, from, to, maxKeys int)` where `maxKeys <= 0` means “unlimited”.
+
+**Why**
+
+- Allows callers (replication worker) to bound the per-cycle key scan to a manageable size.
+
+#### `p2p/kademlia/store/sqlite/replication.go`
+
+**What changed**
+
+- Added support for bounded key scans:
+  - Base query now orders deterministically: `ORDER BY createdAt ASC, key ASC`
+  - When `maxKeys > 0`, a `LIMIT ?` is applied.
+  - If the limit is hit, the query is extended to include all additional rows where `createdAt` equals the last row’s `createdAt` (by fetching `createdAt = boundAt AND key > boundKey`).
+- `ctx.Err()` is checked post-query; cancellation/deadline is treated as failure (returns nil).
+
+**Why**
+
+- Ordering by `(createdAt, key)` provides a stable cursor for a bounded window.
+- Including “same `createdAt`” rows at the limit boundary prevents the replication cursor from skipping keys that share the same timestamp.
+
+#### `p2p/kademlia/store/mem/mem.go`
+
+**What changed**
+
+- Updated `GetKeysForReplication` signature to match interface (`maxKeys` unused).
+
+#### `p2p/kademlia/replication.go`
+
+**What changed**
+
+- Introduced `replicateKeysScanMax = 200000` to bound the per-cycle DB scan.
+- Replication now fetches:
+  - `replicationKeys := store.GetKeysForReplication(globalFrom, to, replicateKeysScanMax)`
+- Defines a stable replication “window end”:
+  - `windowEnd = lastKey.CreatedAt` if keys exist, else `windowEnd = to`
+- Replaces per-peer filtering loops with a one-pass assignment:
+  - Build `peerStart[peerID] = lastReplicatedAt (or historicStart)`
+  - For each key:
+    - compute `closestIDs := closestContactsWithIncludingNode(Alpha, key, ignores, self)`
+    - for each closest ID that is an active peer, if `key.CreatedAt.After(peerStart[id])`, append key into `assignedKeys[id]`
+- Each active peer then:
+  - sends `assignedKeys[peerID]` in 5000-key batches (existing batching)
+  - on full success, advances `lastReplicatedAt` to `windowEnd` (not `time.Now()`)
+  - if no keys assigned for that peer, still advances `lastReplicatedAt` to `windowEnd`
+- `adjustNodeKeys` switched to the new `GetKeysForReplication` signature (unbounded).
+
+**Why**
+
+- Bounding the scan turns replication into incremental progress rather than an unbounded backlog dump.
+- `windowEnd` ensures cursor advancement is aligned to what was actually scanned/processed.
+- One-pass assignment reduces CPU/memory pressure from O(peers * keys) scanning to O(keys * Alpha) assignment.
+
+**Operational expectations**
+
+- Nodes with large backlogs will now “catch up” over multiple intervals (e.g., 16M keys at 200k keys/cycle ≈ 80 cycles).
+- Per-interval replication logs should show materially smaller `keys` and `batches` counts.
+
+### 3) Redundant data cleanup worker compatibility
+
+#### `p2p/kademlia/redundant_data.go`
+
+- Updated call signature for `GetKeysForReplication(..., 0)` to mean “unlimited” in the redundant-data cleanup path.
+
+### 4) P2P DB stats logging noise reduction
+
+#### `p2p/kademlia/store/sqlite/sqlite.go`
+
+**What changed**
+
+- When `Store.Count(ctx)` fails inside `Store.Stats(ctx)`:
+  - If the error is `context.DeadlineExceeded` or `context.Canceled`, it logs at Info instead of Error.
+  - Other errors remain Error.
+
+**Why**
+
+- In production the count query (`SELECT COUNT(*) FROM data`) can legitimately time out under heavy DB load or when metrics collection has a short deadline.
+- The count is typically a best-effort metric; logging it as ERROR creates noise and can mask real faults.
+
+## Notes / Known Limitations
+
+- Replication assignment uses the current routing table’s `closestContactsWithIncludingNode(...)`. If an otherwise active peer is not present in routing-table-derived closest sets, it may receive zero keys for a window and still have its `lastReplicatedAt` advanced to `windowEnd`. This behavior existed in the prior approach as well (advancing on “no closest keys”), but bounding makes it more visible as replication progresses window-by-window.
+- `replicateKeysScanMax` is a constant. If you need environment-specific tuning, consider making it configurable (config file/env var), but this change intentionally keeps scope minimal.
+- The bounded sqlite query may return slightly more than `maxKeys` due to the “same createdAt extension” for correctness.

p2p/kademlia/bootstrap.go

@@ -13,6 +13,7 @@ import (
 	"github.com/LumeraProtocol/supernode/v2/pkg/errors"
 	"github.com/LumeraProtocol/supernode/v2/pkg/logtrace"
 	ltc "github.com/LumeraProtocol/supernode/v2/pkg/net/credentials"
+	"github.com/LumeraProtocol/supernode/v2/pkg/utils"
 )
 
 const (
@@ -111,12 +112,13 @@ func (s *DHT) setBootstrapNodesFromConfigVar(ctx context.Context, bootstrapNodes
 
 // loadBootstrapCandidatesFromChain returns active supernodes (by latest state)
 // mapped by "ip:port". No pings here.
-func (s *DHT) loadBootstrapCandidatesFromChain(ctx context.Context, selfAddress string) (map[string]*Node, error) {
+func (s *DHT) loadBootstrapCandidatesFromChain(ctx context.Context, selfAddress string) (map[string]*Node, map[[32]byte]struct{}, error) {
 	resp, err := s.options.LumeraClient.SuperNode().ListSuperNodes(ctx)
 	if err != nil {
-		return nil, fmt.Errorf("failed to list supernodes: %w", err)
+		return nil, nil, fmt.Errorf("failed to list supernodes: %w", err)
 	}
 
+	activeIDs := make(map[[32]byte]struct{}, len(resp.Supernodes))
 	mapNodes := make(map[string]*Node, len(resp.Supernodes))
 	for _, sn := range resp.Supernodes {
 		if len(sn.States) == 0 {
@@ -134,6 +136,16 @@ func (s *DHT) loadBootstrapCandidatesFromChain(ctx context.Context, selfAddress
 			continue
 		}
 
+		id := strings.TrimSpace(sn.SupernodeAccount)
+		if id == "" {
+			continue
+		}
+		if h, err := utils.Blake3Hash([]byte(id)); err == nil && len(h) == 32 {
+			var key [32]byte
+			copy(key[:], h)
+			activeIDs[key] = struct{}{}
+		}
+
 		// latest IP by height
 		var latestIP string
 		var maxHeight int64 = -1
@@ -170,10 +182,10 @@ func (s *DHT) loadBootstrapCandidatesFromChain(ctx context.Context, selfAddress
 			})
 			continue
 		}
-		node.ID = []byte(sn.SupernodeAccount)
+		node.ID = []byte(id)
 		mapNodes[full] = node
 	}
-	return mapNodes, nil
+	return mapNodes, activeIDs, nil
 }
 
 // upsertBootstrapNode inserts/updates replication_info for the discovered node (Active=false).
@@ -248,11 +260,16 @@ func (s *DHT) SyncBootstrapOnce(ctx context.Context, bootstrapNodes string) erro
 	}
 	selfAddress := fmt.Sprintf("%s:%d", parseSupernodeAddress(supernodeAddr), s.options.Port)
 
-	cands, err := s.loadBootstrapCandidatesFromChain(ctx, selfAddress)
+	cands, activeIDs, err := s.loadBootstrapCandidatesFromChain(ctx, selfAddress)
 	if err != nil {
 		return err
 	}
 
+	// Update eligibility gate from chain Active state and prune any peers that slipped in via
+	// inbound traffic before the last bootstrap refresh.
+	s.setRoutingAllowlist(ctx, activeIDs)
+	s.pruneIneligibleRoutingPeers(ctx)
+
 	// Upsert candidates to replication_info
 	seen := make(map[string]struct{}, len(cands))
 	s.options.BootstrapNodes = s.options.BootstrapNodes[:0]