PC3 Phase 2.1: server-authoritative sim + prediction/reconciliation with replay
Made the multiplayer core actually work (it had never run a real sim): - project.godot: physics pinned to 30 Hz -- the shared sim/snapshot/prediction rate. - server.gd: loads the world's level (real ground + collision), spawns a server Player per peer at the level's PlayerSpawn, broadcasts a full snapshot every physics frame with each peer's state AND last_seq (the last input seq applied for that peer). Accepts `--port/--world/--status-file` args (status file = player count for a future idle reaper); defaults keep `godot res://scenes/server.tscn` working standalone. - player_spawner.gd (server Player): rewrote -- drains its whole input queue each physics frame in seq order, integrates gravity/walk/jump with the shared constants, move_and_slide(), tracks last_applied_seq. No more no-op process_server_input(). - network_client.gd: rewrote reconciliation -- drops inputs the server has acked (last_seq), re-simulates the still-in-flight ones from the authoritative state, snaps only on real drift. Exposes get_remote_states() (peer_id -> latest snapshot state) for the game to render other players. predict_and_send() now also sends body yaw. - proxies trimmed to the RPC set the rewrite actually uses. - scripts/mp_test_client.gd + run_multiplayer_2p_test.sh: two real client processes vs one real server -- client A sees client B and its own movement round-trips. PASS. - run_checkpoint_tests.py: also run the 2p shell test. smoke_test_network_client 4/4, player 15/15, avatar_in_world 10/10, all green. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015jtxAiu5dAW4WxCNdGkrkJ
This commit is contained in:
+139
-133
@@ -1,75 +1,62 @@
|
||||
class_name NetworkClient
|
||||
extends Node
|
||||
## Client-side networked player: input prediction + server reconciliation.
|
||||
## Client-side multiplayer: ENet connection, local-player input prediction +
|
||||
## server reconciliation (with input replay), and remote-player state for the
|
||||
## game to render.
|
||||
##
|
||||
## Implements docs/architecture/networking.md §5 (Prediction & Reconciliation)
|
||||
## for the *local* player only -- remote-entity interpolation (§5.3) and full
|
||||
## delta-compressed multi-entity snapshots (§6) are explicitly out of scope
|
||||
## for this first pass; the server already only ever has one player in PC2's
|
||||
## single-peer testing, and this task's own description scopes it to "the
|
||||
## player controller built in the dependency task" (singular).
|
||||
## RPC path-matching: Godot resolves an rpc_id() call's target by matching the
|
||||
## sending node's NodePath on the receiver. scripts/server.gd is the SceneTree
|
||||
## root named "Server"; it spawns each peer's PlayerNode as a child named
|
||||
## "Player_<peer_id>". So the client needs local nodes at /root/Server (for the
|
||||
## server-initiated _receive_snapshot) and /root/Server/Player_<own_id> (the
|
||||
## target of rpc_request_input). connect_to_server() builds that once the peer
|
||||
## id is known.
|
||||
##
|
||||
## RPC path-matching, the part that makes this actually work:
|
||||
## Godot's high-level rpc_id() resolves the *receiving* node by matching the
|
||||
## *sending* node's NodePath on the caller's side. The server (scripts/
|
||||
## server.gd) is the SceneTree root node named "Server" when the dedicated
|
||||
## server runs scenes/server.tscn as its main scene, and it spawns each
|
||||
## connected player as a *child* of itself named "Player_<peer_id>"
|
||||
## (scripts/server.gd's _on_peer_connected). For an RPC sent by the client to
|
||||
## actually land on the right node, the client needs a LOCAL node tree at the
|
||||
## exact same paths -- /root/Server for _receive_snapshot (called by the
|
||||
## server on itself), and /root/Server/Player_<own_peer_id> for
|
||||
## rpc_request_input (a real method on the server's PlayerNode). Those local
|
||||
## nodes don't need to run any server logic -- they only need to exist at the
|
||||
## right path so Godot's RPC routing finds them. connect_to_server() builds
|
||||
## that structure once the peer id is known (only available after the ENet
|
||||
## handshake completes).
|
||||
##
|
||||
## Usage: instantiate directly (not an autoload -- this only exists for the
|
||||
## lifetime of an actual multiplayer session, not on every client launch):
|
||||
## Usage:
|
||||
## var net := NetworkClient.new()
|
||||
## get_tree().root.add_child(net)
|
||||
## var ok := await net.connect_to_server("127.0.0.1", 7777)
|
||||
## if await net.connect_to_server(host, port):
|
||||
## # each physics frame, from the local player controller:
|
||||
## net.predict_and_send(delta, move_x, move_z, jump, body_yaw)
|
||||
## # read net.position / net.get_remote_states() to render
|
||||
|
||||
# Movement constants -- MUST match scripts/server.gd and player_spawner.gd.
|
||||
const WALK_SPEED := 5.0
|
||||
const JUMP_VELOCITY := 4.5
|
||||
const GRAVITY := 9.8
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Movement constants -- MUST match scripts/player_spawner.gd exactly, per
|
||||
# docs/architecture/networking.md §5.1 ("the same physics constants as the
|
||||
# server"). A mismatch here doesn't break anything outright, it just makes
|
||||
# reconciliation corrections more frequent/visible.
|
||||
# ---------------------------------------------------------------------------
|
||||
## Predicted vs authoritative drift beyond this (metres) snaps rather than
|
||||
## trusting the prediction.
|
||||
const RECONCILE_ERROR_THRESHOLD := 0.35
|
||||
|
||||
const WALK_SPEED: float = 5.0
|
||||
const JUMP_VELOCITY: float = 4.5
|
||||
const GRAVITY: float = 9.8
|
||||
|
||||
## How far predicted and authoritative position can drift before snapping
|
||||
## rather than trusting the prediction was "close enough." Per §5.2.
|
||||
const RECONCILE_ERROR_THRESHOLD: float = 0.5
|
||||
|
||||
## How many past predicted states to keep for reconciliation replay. Real
|
||||
## sizing per §5.4 is RTT-adaptive; a fixed generous buffer is the
|
||||
## appropriately-scoped first pass here -- 120 ticks at a 30Hz send rate is
|
||||
## 4 seconds of history, comfortably more than any realistic LAN/local RTT.
|
||||
const PREDICTION_BUFFER_SIZE: int = 120
|
||||
## Ring buffer cap. 120 frames at 30 Hz = 4 s of history, well past any
|
||||
## realistic LAN RTT.
|
||||
const PREDICTION_BUFFER_SIZE := 120
|
||||
|
||||
signal reconciled(correction_distance: float)
|
||||
signal connected_to_server
|
||||
signal connection_failed
|
||||
|
||||
signal server_disconnected
|
||||
|
||||
var peer: ENetMultiplayerPeer
|
||||
var _server_root: Node # local /root/Server proxy
|
||||
var _player_proxy: Node # local /root/Server/Player_<peer_id> proxy
|
||||
var my_peer_id: int = 0
|
||||
var _server_root: Node
|
||||
var _player_proxy: Node
|
||||
|
||||
var _seq: int = 0
|
||||
## Ring buffer of {seq, input, position, velocity} predicted at send time,
|
||||
## oldest first. Trimmed as the server acknowledges ticks via snapshots.
|
||||
## Ring buffer of {seq, move_x, move_z, jump, delta, position, velocity},
|
||||
## oldest first. Entries are dropped as the server acks their seq.
|
||||
var _prediction_buffer: Array[Dictionary] = []
|
||||
|
||||
## Local predicted player state.
|
||||
var position: Vector3 = Vector3.ZERO
|
||||
var velocity: Vector3 = Vector3.ZERO
|
||||
var _on_floor: bool = false
|
||||
var _on_floor: bool = true
|
||||
|
||||
## peer_id -> latest authoritative state dict from the most recent snapshot,
|
||||
## for every peer that isn't us. The game reads this to render other players.
|
||||
var _remote_states: Dictionary = {}
|
||||
var _last_snapshot_tick: int = -1
|
||||
|
||||
|
||||
func connect_to_server(host: String, port: int) -> bool:
|
||||
@@ -81,28 +68,24 @@ func connect_to_server(host: String, port: int) -> bool:
|
||||
return false
|
||||
|
||||
multiplayer.multiplayer_peer = peer
|
||||
multiplayer.connected_to_server.connect(_on_connected)
|
||||
multiplayer.connection_failed.connect(func():
|
||||
connection_failed.emit()
|
||||
)
|
||||
multiplayer.server_disconnected.connect(func(): server_disconnected.emit())
|
||||
|
||||
for i in 300: # ~5s at 60fps
|
||||
for _i in 300: # ~5 s at 60 fps
|
||||
await get_tree().process_frame
|
||||
var status := peer.get_connection_status()
|
||||
if status == MultiplayerPeer.CONNECTION_CONNECTED:
|
||||
return true
|
||||
elif status == MultiplayerPeer.CONNECTION_DISCONNECTED:
|
||||
connection_failed.emit()
|
||||
return false
|
||||
match peer.get_connection_status():
|
||||
MultiplayerPeer.CONNECTION_CONNECTED:
|
||||
_on_connected()
|
||||
return true
|
||||
MultiplayerPeer.CONNECTION_DISCONNECTED:
|
||||
connection_failed.emit()
|
||||
return false
|
||||
connection_failed.emit()
|
||||
return false
|
||||
|
||||
|
||||
func _on_connected() -> void:
|
||||
var my_id := multiplayer.get_unique_id()
|
||||
my_peer_id = multiplayer.get_unique_id()
|
||||
|
||||
# Build the local proxy tree at the exact paths the server's own RPC
|
||||
# calls target -- see the class docstring for why this has to exist.
|
||||
_server_root = Node.new()
|
||||
_server_root.name = "Server"
|
||||
_server_root.set_script(preload("res://scripts/network_receiver_proxy.gd"))
|
||||
@@ -110,7 +93,7 @@ func _on_connected() -> void:
|
||||
_server_root.set("owner_client", self)
|
||||
|
||||
_player_proxy = Node.new()
|
||||
_player_proxy.name = "Player_%d" % my_id
|
||||
_player_proxy.name = "Player_%d" % my_peer_id
|
||||
_player_proxy.set_script(preload("res://scripts/network_player_proxy.gd"))
|
||||
_server_root.add_child(_player_proxy)
|
||||
|
||||
@@ -123,60 +106,56 @@ func disconnect_from_server() -> void:
|
||||
if _server_root and is_instance_valid(_server_root):
|
||||
_server_root.queue_free()
|
||||
multiplayer.multiplayer_peer = null
|
||||
_prediction_buffer.clear()
|
||||
_remote_states.clear()
|
||||
|
||||
|
||||
func is_connected_to_server() -> bool:
|
||||
return peer != null and peer.get_connection_status() == MultiplayerPeer.CONNECTION_CONNECTED
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Prediction -- called every physics frame by whatever drives input (a real
|
||||
# controller, or a test harness). Applies input locally immediately (so it
|
||||
# renders with zero perceived latency) and sends it to the server, per §5.1.
|
||||
# Prediction -- call every physics frame from the local player controller.
|
||||
# Applies input locally immediately (zero perceived latency) and sends it.
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
func predict_and_send(delta: float, move_x: float, move_z: float, jump: bool) -> void:
|
||||
func predict_and_send(delta: float, move_x: float, move_z: float, jump: bool, look_yaw: float = 0.0) -> void:
|
||||
_apply_movement(delta, move_x, move_z, jump)
|
||||
|
||||
_seq += 1
|
||||
var record := {
|
||||
_prediction_buffer.append({
|
||||
"seq": _seq,
|
||||
"input": {"move_x": move_x, "move_z": move_z, "jump": jump},
|
||||
"position": position,
|
||||
"velocity": velocity,
|
||||
}
|
||||
_prediction_buffer.append(record)
|
||||
"move_x": move_x, "move_z": move_z, "jump": jump,
|
||||
"delta": delta,
|
||||
"position": position, "velocity": velocity,
|
||||
})
|
||||
if _prediction_buffer.size() > PREDICTION_BUFFER_SIZE:
|
||||
_prediction_buffer.pop_front()
|
||||
|
||||
if _player_proxy and is_instance_valid(_player_proxy):
|
||||
_player_proxy.rpc_id(1, "rpc_request_input", record["input"], _seq)
|
||||
_player_proxy.rpc_id(1, "rpc_request_input",
|
||||
{"move_x": move_x, "move_z": move_z, "jump": jump, "look_yaw": look_yaw}, _seq)
|
||||
|
||||
|
||||
func _apply_movement(delta: float, move_x: float, move_z: float, jump: bool) -> void:
|
||||
# Mirrors scripts/player_spawner.gd's _process_input + gravity/jump,
|
||||
# intentionally without its ground-friction smoothing -- friction is a
|
||||
# server-authoritative-feel detail, not something worth risking a
|
||||
# prediction/server mismatch over in this first pass. Position still
|
||||
# reconciles against the server's real (frictioned) result every
|
||||
# snapshot, so this doesn't silently drift.
|
||||
if not _on_floor:
|
||||
velocity.y -= GRAVITY * delta
|
||||
elif jump:
|
||||
velocity.y = JUMP_VELOCITY
|
||||
|
||||
var direction := Vector3(move_x, 0.0, move_z)
|
||||
if direction.length() > 0.01:
|
||||
direction = direction.normalized()
|
||||
velocity.x = direction.x * WALK_SPEED
|
||||
velocity.z = direction.z * WALK_SPEED
|
||||
var dir := Vector3(move_x, 0.0, move_z)
|
||||
if dir.length() > 0.01:
|
||||
dir = dir.normalized()
|
||||
velocity.x = dir.x * WALK_SPEED
|
||||
velocity.z = dir.z * WALK_SPEED
|
||||
else:
|
||||
velocity.x = 0.0
|
||||
velocity.z = 0.0
|
||||
|
||||
if jump and _on_floor:
|
||||
velocity.y = JUMP_VELOCITY
|
||||
|
||||
position += velocity * delta
|
||||
# Headless prediction has no collision shape of its own to test
|
||||
# is_on_floor() against -- treat "at or below the spawn-plane ground"
|
||||
# as grounded. A real render-side controller would swap this for the
|
||||
# actual CharacterBody3D.is_on_floor() of whatever visual node this
|
||||
# drives; this class only owns the *networking* half.
|
||||
# Headless prediction has no collision shape; treat the spawn plane as the
|
||||
# floor. A render-side controller drives the real CharacterBody3D and reads
|
||||
# `position` back for reconciliation only.
|
||||
if position.y <= 0.0:
|
||||
position.y = 0.0
|
||||
velocity.y = 0.0
|
||||
@@ -186,49 +165,76 @@ func _apply_movement(delta: float, move_x: float, move_z: float, jump: bool) ->
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Reconciliation -- called by _ReceiverScript when a snapshot arrives.
|
||||
# Reconciliation -- _ReceiverProxy calls this when a snapshot arrives.
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
func _handle_snapshot(snapshot_data: Dictionary) -> void:
|
||||
var my_id := multiplayer.get_unique_id()
|
||||
var players: Dictionary = snapshot_data.get("players", {})
|
||||
# Server keys player_states by peer_id -- Dictionary keys from JSON-ish
|
||||
# RPC payloads can arrive as String or int depending on the transport
|
||||
# encoding, so check both rather than assuming.
|
||||
var my_state: Variant = players.get(my_id, players.get(str(my_id), null))
|
||||
if my_state == null or not (my_state is Dictionary):
|
||||
func _handle_snapshot(snap: Dictionary) -> void:
|
||||
var snap_tick: int = snap.get("tick", 0)
|
||||
if snap_tick <= _last_snapshot_tick:
|
||||
return # Stale / out-of-order (unreliable channel).
|
||||
_last_snapshot_tick = snap_tick
|
||||
|
||||
var players: Dictionary = snap.get("players", {})
|
||||
|
||||
# --- Remote players ---
|
||||
_remote_states.clear()
|
||||
for key in players:
|
||||
var pid := int(key)
|
||||
if pid == my_peer_id:
|
||||
continue
|
||||
_remote_states[pid] = players[key]
|
||||
|
||||
# --- Local player: reconcile ---
|
||||
var mine: Variant = players.get(my_peer_id, players.get(str(my_peer_id), null))
|
||||
if not (mine is Dictionary):
|
||||
return
|
||||
var auth_pos: Vector3 = mine.get("position", position)
|
||||
var auth_vel: Vector3 = mine.get("velocity", velocity)
|
||||
var acked_seq: int = int(mine.get("last_seq", 0))
|
||||
|
||||
var authoritative_pos: Vector3 = (my_state as Dictionary).get("position", position)
|
||||
# Drop inputs the server has already applied.
|
||||
while not _prediction_buffer.is_empty() and _prediction_buffer[0]["seq"] <= acked_seq:
|
||||
_prediction_buffer.pop_front()
|
||||
|
||||
# Real limitation, noted explicitly rather than silently faked: the
|
||||
# server's snapshot_state (scripts/server.gd's _broadcast_snapshot) does
|
||||
# not currently echo back "last input seq processed for this peer" --
|
||||
# its `tick` field is the server's own internal simulation tick counter,
|
||||
# a completely different clock from this client's per-input `_seq`. True
|
||||
# tick-correlated rollback (docs/architecture/networking.md §5.2 step 4,
|
||||
# "re-simulate any inputs that occurred after the server's tick") needs
|
||||
# that ack echoed back per-peer -- a real, scoped follow-up, not
|
||||
# something to fake here. Until then, reconcile against the MOST RECENT
|
||||
# prediction (not a tick-matched one): still a real error measurement
|
||||
# and a real correction, just without precise replay-from-exact-tick.
|
||||
if _prediction_buffer.is_empty():
|
||||
return
|
||||
var latest_entry: Dictionary = _prediction_buffer[-1]
|
||||
# Re-simulate the still-in-flight inputs from the authoritative state.
|
||||
var replay_pos := auth_pos
|
||||
var replay_vel := auth_vel
|
||||
var replay_floor := replay_pos.y <= 0.0
|
||||
for entry in _prediction_buffer:
|
||||
var d: float = entry["delta"]
|
||||
if not replay_floor:
|
||||
replay_vel.y -= GRAVITY * d
|
||||
elif entry["jump"]:
|
||||
replay_vel.y = JUMP_VELOCITY
|
||||
var dir := Vector3(entry["move_x"], 0.0, entry["move_z"])
|
||||
if dir.length() > 0.01:
|
||||
dir = dir.normalized()
|
||||
replay_vel.x = dir.x * WALK_SPEED
|
||||
replay_vel.z = dir.z * WALK_SPEED
|
||||
else:
|
||||
replay_vel.x = 0.0
|
||||
replay_vel.z = 0.0
|
||||
replay_pos += replay_vel * d
|
||||
if replay_pos.y <= 0.0:
|
||||
replay_pos.y = 0.0
|
||||
replay_vel.y = 0.0
|
||||
replay_floor = true
|
||||
else:
|
||||
replay_floor = false
|
||||
|
||||
var error: float = (latest_entry["position"] as Vector3).distance_to(authoritative_pos)
|
||||
if error <= RECONCILE_ERROR_THRESHOLD:
|
||||
return # Close enough -- prediction was correct, nothing to do.
|
||||
var error := position.distance_to(replay_pos)
|
||||
if error > RECONCILE_ERROR_THRESHOLD:
|
||||
position = replay_pos
|
||||
velocity = replay_vel
|
||||
_on_floor = replay_floor
|
||||
reconciled.emit(error)
|
||||
|
||||
# Correct: snap to the authoritative position. Without a real ack-seq
|
||||
# from the server there's no reliable "replay inputs after this point"
|
||||
# window to recompute, so the buffer is cleared and prediction resumes
|
||||
# fresh from the corrected position on the next input -- still visibly
|
||||
# responsive (no waiting for a round trip before the *next* input
|
||||
# applies), just without this pass replaying in-flight inputs.
|
||||
position = authoritative_pos
|
||||
velocity = Vector3.ZERO
|
||||
_prediction_buffer.clear()
|
||||
|
||||
reconciled.emit(error)
|
||||
# ---------------------------------------------------------------------------
|
||||
# Remote-player state for rendering
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
## peer_id -> {position, rotation_y, velocity, on_floor, last_seq} from the most
|
||||
## recent snapshot. Read this each frame to spawn/update remote avatars.
|
||||
func get_remote_states() -> Dictionary:
|
||||
return _remote_states
|
||||
|
||||
Reference in New Issue
Block a user