fix(hnsw): de-duplicate dead-node UIDs on vector delete

[shaunpatterson]

Problem

When a vfloat (vector) value is deleted, addIndexMutations records the uid in the HNSW dead-node list (posting/index.go). That list is a single posting — DataKey(<pred>__vector_dead, 1) holding a JSON array — which is read, appended to, and re-marshalled in full on every delete:

deadNodes, err = hnsw.ParseEdges(string(deadData.Value.([]byte)))
...
deadNodes = append(deadNodes, uid)   // unconditional
deadNodesBytes, _ := json.Marshal(deadNodes)
pl.addMutation(ctx, txn, edge)

The uid is appended unconditionally, so a uid already recorded as dead — delete/reinsert churn, or a Raft re-apply of the same mutation — is appended again. That grows the posting with duplicates and re-marshals the whole blob for no effect. removeDeadNodes de-dupes on read (it builds a set), so the duplicates were functionally invisible but still bloated the posting and its parse cost.

Fix

Skip the rewrite when the uid is already present, via a small addDeadNode helper that reports whether the list changed:

func addDeadNode(deadNodes []uint64, uid uint64) ([]uint64, bool) {
	if slices.Contains(deadNodes, uid) {
		return deadNodes, false
	}
	return append(deadNodes, uid), true
}

Behaviour is unchanged for the first delete of a uid; repeated deletes become a no-op instead of appending a duplicate and rewriting the posting.

Scope

This intentionally does not address the deeper design issue that the dead list grows unbounded with distinct deletions and is never garbage-collected — that needs the index.Remove() path the existing // TODO look into better alternatives calls for, and is left as-is.

Test

TestAddDeadNode covers first-insert, distinct-append, and already-present (no-op) cases.

• go test ./posting/ — pass
• go vet / gofmt — clean (no new findings)
• golangci-lint (repo config) — no findings in the changed code

🤖 Generated with Claude Code