File size: 36,745 Bytes
12439fa 654c8c7 12439fa 654c8c7 cf2697b 12439fa 654c8c7 12439fa 654c8c7 12439fa 654c8c7 12439fa f55f75f 07e80ad 12439fa 07e80ad 12439fa 52a986a b5e5650 654c8c7 12439fa 654c8c7 74770e8 654c8c7 12439fa b5e5650 07e80ad 236abb8 654c8c7 07e80ad 3550e4d 07e80ad 16d8704 3a871a0 4b5c463 2244d0f 4b5c463 2244d0f 4b5c463 b5e5650 654c8c7 12439fa 654c8c7 4b5c463 12439fa 4b5c463 12439fa 654c8c7 12439fa 654c8c7 12439fa 654c8c7 279ccf2 bba6f8a 52a986a b5e5650 654c8c7 12439fa 654c8c7 12439fa 654c8c7 4b5c463 279ccf2 4b5c463 654c8c7 279ccf2 f656047 52a986a b5e5650 654c8c7 12439fa 654c8c7 cf2697b 654c8c7 5144b7e 94aef6f 654c8c7 f656047 07e80ad 52a986a 654c8c7 3a871a0 279ccf2 654c8c7 12439fa 07e80ad b5e5650 07e80ad b5e5650 07e80ad b5e5650 07e80ad 12439fa 4b5c463 52a986a 4b5c463 12439fa 654c8c7 94aef6f 654c8c7 94aef6f 654c8c7 07e80ad 52a986a 3a871a0 279ccf2 654c8c7 279ccf2 654c8c7 bba6f8a 279ccf2 bba6f8a 654c8c7 f55f75f 12439fa 654c8c7 d062bfb 12439fa f55f75f bba6f8a d062bfb 654c8c7 bba6f8a f55f75f bba6f8a 12439fa f55f75f 654c8c7 f55f75f 654c8c7 bba6f8a d062bfb f55f75f d062bfb 654c8c7 07e80ad f55f75f 654c8c7 3a871a0 f55f75f 3a871a0 654c8c7 279ccf2 f55f75f 279ccf2 654c8c7 f55f75f 654c8c7 12439fa 654c8c7 bba6f8a 654c8c7 b5e5650 279ccf2 654c8c7 b5e5650 654c8c7 52a986a b5e5650 279ccf2 b5e5650 52a986a 654c8c7 bba6f8a 12439fa 654c8c7 b5e5650 12439fa 654c8c7 b5e5650 12439fa 07e80ad 654c8c7 12439fa b5e5650 12439fa 07e80ad 52a986a b5e5650 07e80ad 654c8c7 12439fa 07e80ad b5e5650 07e80ad b5e5650 07e80ad 5144b7e 07e80ad 5144b7e 07e80ad 12439fa 07e80ad 5144b7e 07e80ad 5144b7e 07e80ad 5144b7e 236abb8 5144b7e 236abb8 07e80ad b5e5650 07e80ad b5e5650 07e80ad 654c8c7 b5e5650 654c8c7 279ccf2 654c8c7 279ccf2 654c8c7 279ccf2 654c8c7 279ccf2 654c8c7 12439fa 654c8c7 12439fa 52a986a 654c8c7 52a986a 12439fa 52a986a 12439fa 52a986a 654c8c7 12439fa 52a986a b5e5650 52a986a 12439fa 77ede9e 52a986a b5e5650 52a986a b5e5650 52a986a b5e5650 52a986a b5e5650 52a986a b5e5650 52a986a b5e5650 52a986a 77ede9e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
AntiAtropos Core Simulation Physics.
A discrete-time fluid-queue model simulating a 5-node microservice cluster.
Each node has stateful queues, capacities, and failure probabilities.
Dynamic traffic is injected per tick, and management actions shift capacity
and routing parameters.
"""
import random
from dataclasses import dataclass, field
from enum import Enum
from typing import Optional
# ---------------------------------------------------------------------------
# Simulator Constants (World Physics)
# ---------------------------------------------------------------------------
DEFAULT_CAPACITY: float = 3.0 # Baseline service rate (μ) per node
MAX_CAPACITY: float = 999.0 # High sentinel — cost function (not an artificial cap) limits scaling. Real K8s has no per-node pod ceiling; RL must learn cost discipline.
MAX_SCALING_STEP: int = 3 # Largest allowed SCALE_UP param (was 5)
BOOT_DELAY_TICKS: int = 5 # Time it takes to bring up infrastructure
BASE_LATENCY_MS: float = 20.0 # Minimum processing time
OVERLOAD_THRESHOLD: int = 80 # Request count where node begins to "fail" (DEGRADED)
LATENCY_STEEPNESS: float = 2.0 # Increased to ensure SLA violations before death
FATAL_FAIL_THRESHOLD: int = 200 # Hard cap on queue depth (catastrophic failure boundary)
CASCADE_WINDOW_TICKS: int = 3 # Ticks after a failure to check for cascade effects
CASCADE_QUEUE_MULTIPLIER: float = 1.2 # Queue must exceed FATAL_FAIL_THRESHOLD * this to cascade
NODE_RECOVERY_TICKS: int = 20 # Ticks before a FAILED node auto-recovers
BACKPRESSURE_THRESHOLD: float = 60.0 # Queue depth that triggers backpressure
BACKPRESSURE_MAX_FACTOR: float = 0.4 # Maximum service rate reduction (40%)
SENSOR_DROPOUT_PROB: float = 0.05 # P(node.queue, latency reports 0 or -1.0)
NODE_FAILURE_PROB: float = 0.00 # P(node fails naturally) — largely driven by task profile
# Cost model -- two-tier: cheap for capacity needed to handle current
# traffic, expensive for idling excess (punishes over-provisioning
# without discouraging necessary scaling).
COST_PER_CAPACITY_UNIT_PER_HOUR: float = 0.05 # Base rate for needed capacity
OVERPROVISION_COST_PER_UNIT: float = 1.0 # Penalty rate for excess (20x base)
# Task Profiles (Domain Randomization)
# Task 1: Start at 92-99% of ingress capacity (randomised in _randomize_domain).
# DAG ingress capacity = 2 ingress nodes * DEFAULT_CAPACITY * 15 = 90 req/tick.
# lambda_init ≈ 83-89 so each ingress node sees ~41-44 req/tick (just under 45 capacity).
T1_INITIAL_LAMBDA: float = 86.0 # midpoint of [82.8, 89.1]; overridden by _randomize_domain
T1_RAMP_SLOPE: float = 0.5 # +0.5 req/tick globally per tick
# Task 2: lambda at 100-110% of ingress capacity — guarantees immediate ingress overload.
T2_INITIAL_LAMBDA: float = 95.0 # midpoint of [90, 99]; overridden by _randomize_domain
T2_FAIL_TICK: int = 20
T3_INITIAL_LAMBDA: float = 60.0
# Task 3 surge parameters — base window, jitter applied per episode
T3_SURGE_CYCLE: int = 60 # Cycle length (ticks)
T3_SURGE_BASE_START: int = 30 # Nominal start of surge within cycle
T3_SURGE_BASE_END: int = 40 # Nominal end of surge within cycle
T3_SURGE_JITTER: int = 10 # ±jitter applied to start/end each episode
T3_SURGE_MAGNITUDE: float = 60.0 # Extra req/tick added to node-1 and node-2 (4 SCALE_UPs at delta=1 covers it: 5x15=75 = 15 DAG + 60 surge)
# Only the VIP payment gateway (node-0) is hard-protected from shedding --
# this is a safety guardrail, not a learned behavior. The agent must learn
# from the reward signal (error_rate -> sla_penalty at gamma=6.0) that
# shedding surge nodes (node-1/2) is rarely optimal.
CRITICAL_NODES: list[str] = ["node-0"]
# VIP / business-critical node weights.
# node-0 is the payment portal, so its queue growth or failure matters more.
# Reduced from 4.0 → 2.0 to prevent reward gradient from creating
# a local optimum where the agent only scales node-0.
# At 2×, node-0 is still prioritized but other nodes remain viable targets.
VIP_NODE_WEIGHTS: dict[str, float] = {
"node-0": 2.0,
}
# ---------------------------------------------------------------------------
# Graph Topology (DAG — fixed 5-node cluster architecture)
# ---------------------------------------------------------------------------
# Directed edges: parent -> list of direct children.
# node-0 (payments/VIP) is the primary ingress; node-4 (auth) is independent.
CLUSTER_TOPOLOGY: dict[str, list[str]] = {
"node-0": ["node-1", "node-2"],
"node-1": [],
"node-2": ["node-3"],
"node-3": [],
"node-4": [],
}
# Nodes that receive raw external traffic directly.
EXTERNAL_TRAFFIC_NODES: set[str] = {"node-0", "node-4"}
# 50/50 external λ split between the two ingress nodes.
# node-0 (payments/VIP) and node-4 (auth) each receive half of total_lambda.
# total_lambda is the cluster-wide external arrival rate (req/tick).
# Each ingress node therefore sees total_lambda * 0.5 req/tick at its input.
EXTERNAL_LAMBDA_FRACTION: float = 0.5
# Default upstream-to-downstream routing weights (parent → child splits).
# These represent the baseline traffic split before agent rerouting.
DEFAULT_ROUTING_SPLIT: dict[str, dict[str, float]] = {
"node-0": {"node-1": 0.5, "node-2": 0.5},
"node-2": {"node-3": 1.0},
}
# Pre-computed topological order (Kahn's BFS on CLUSTER_TOPOLOGY).
# Ensures parents are always processed before their children in _inject_traffic().
# Order: node-0, node-4 (roots) → node-1, node-2 (node-0 children) → node-3 (node-2 child).
_TOPOLOGICAL_ORDER: tuple[str, ...] = ("node-0", "node-4", "node-1", "node-2", "node-3")
class NodeStatus(str, Enum):
HEALTHY = "HEALTHY"
DEGRADED = "DEGRADED"
FAILED = "FAILED"
@dataclass
class NodeState:
node_id: str
status: NodeStatus = NodeStatus.HEALTHY
is_vip: bool = False
# Physics parameters
capacity: float = DEFAULT_CAPACITY
importance_weight: float = 1.0
queue_depth: float = 0.0
latency_ms: float = BASE_LATENCY_MS
incoming_request_rate: float = 0.0
cpu_utilization: float = 0.0
# Per-tick accounting (reset each tick)
dropped_requests: float = 0.0
shed_fraction: float = 0.0 # Fraction of incoming traffic to drop this tick
pending_capacity_queue: list[int] = field(default_factory=list)
recovery_timer: int = 0 # Countdown to auto-recovery from FAILED status
is_scripted_failure: bool = False # True if failed due to task scripting (no auto-recovery)
outflow_rate: float = 0.0 # Requests/tick actually dispatched downstream (DAG edge signal)
# Derived (recomputed whenever capacity or status changes)
@property
def service_rate(self) -> float:
"""The total service capacity (μ) in requests per tick."""
if self.status == NodeStatus.FAILED:
return 0.0
return self.capacity * 15.0 # Base unit = 15 req/tick processing power
def to_dict(self) -> dict:
return {
"node_id": self.node_id,
"status": self.status,
"is_vip": self.is_vip,
"capacity": self.capacity,
"importance_weight": self.importance_weight,
"queue_depth": int(self.queue_depth),
"latency_ms": round(self.latency_ms, 2),
"incoming_request_rate": round(self.incoming_request_rate, 2),
"cpu_utilization": round(min(1.0, self.cpu_utilization), 4),
"dropped_requests": int(self.dropped_requests),
"shed_fraction": round(self.shed_fraction, 4),
"capacity_units": int(self.capacity),
"pending_capacity_units": int(len(self.pending_capacity_queue)),
"recovery_timer": self.recovery_timer,
"is_scripted_failure": self.is_scripted_failure,
"outflow_rate": round(self.outflow_rate, 2),
}
class ClusterSimulator:
"""
Multi-node fluid queue simulator.
Operates in discrete ticks.
1. Action: Control plane actions (Scaling/Routing/Shedding) are applied.
2. Tick: Physics engine updates queues based on λ (incoming) and μ (service rate).
3. Failure Logic: Queue overflows trigger status degradation/node death.
"""
def __init__(self, n_nodes: int = 5, task_id: str = "task-1", seed: Optional[int] = None):
self._n_nodes = n_nodes
self._task_id = task_id
# Default to non-deterministic RNG seeding so fresh simulator instances
# do not replay identical domain-randomization sequences.
# Pass an explicit seed for reproducible experiments.
self._seed: Optional[int] = seed
self._rng = random.Random(seed)
self._tick_count: int = 0
self._failed_node_id: Optional[str] = None
self._t1_ramp_slope: float = T1_RAMP_SLOPE
self._t1_init_lambda: float = T1_INITIAL_LAMBDA
self._t2_fail_tick: int = T2_FAIL_TICK
self._t2_init_lambda: float = T2_INITIAL_LAMBDA
# Task-3 surge window — randomised per episode in _randomize_domain()
self._t3_surge_start: int = T3_SURGE_BASE_START
self._t3_surge_end: int = T3_SURGE_BASE_END
# Per-node reroute weights for REROUTE_TRAFFIC (node_id → fraction)
self._reroute_weights: dict[str, float] = {}
self._cascade_tick: int = 0 # Tick counter for cascade detection window
self._cascade_triggered: bool = False # Set True when a NEW overload failure occurs
self._nodes: list[NodeState] = []
self.invalid_action_count: int = 0
self._randomize_domain()
self._reset_nodes()
def _randomize_domain(self) -> None:
"""Apply domain randomization for RL robustness across tasks."""
self._t1_ramp_slope = self._rng.uniform(0.8, 2.0)
# DAG calibration: total_lambda is split across 2 ingress nodes (node-0, node-4).
# Each ingress node's capacity is DEFAULT_CAPACITY * 15 req/tick.
# Ingress cluster capacity = len(EXTERNAL_TRAFFIC_NODES) * DEFAULT_CAPACITY * 15 = 90.
# Task 1: start between 92-99% of ingress capacity so the ingress nodes are
# near saturation immediately, producing rich early reward signal.
n_ingress = len(EXTERNAL_TRAFFIC_NODES) # 2
ingress_mu_total = n_ingress * DEFAULT_CAPACITY * 15.0 # 90 req/tick
self._t1_init_lambda = self._rng.uniform(
ingress_mu_total * 0.92, ingress_mu_total * 0.99
)
self._t2_fail_tick = self._rng.randint(10, 40)
# Task 2: guarantee immediate ingress overload (slightly above ingress saturation).
# Each ingress node sees total_lambda/2; target ~102% of individual ingress capacity.
self._t2_init_lambda = self._rng.uniform(
ingress_mu_total * 1.00, ingress_mu_total * 1.10
)
# Task 3: jitter the surge window so the LLM can't memorise it.
jitter = self._rng.randint(-T3_SURGE_JITTER, T3_SURGE_JITTER)
self._t3_surge_start = T3_SURGE_BASE_START + jitter
self._t3_surge_end = T3_SURGE_BASE_END + jitter
# Reset reroute weights to uniform on each new episode
self._reroute_weights = {}
def _reset_nodes(self) -> None:
self._nodes = [
NodeState(
node_id=f"node-{i}",
is_vip=f"node-{i}" in VIP_NODE_WEIGHTS,
importance_weight=VIP_NODE_WEIGHTS.get(f"node-{i}", 1.0),
is_scripted_failure=False,
)
for i in range(self._n_nodes)
]
def reset(self, task_id: str = "task-1", seed: Optional[int] = None) -> None:
"""Restart the simulator for a fresh episode."""
if seed is not None:
self._seed = seed
# Reinitialize RNG so episode generation is reproducible for a given seed.
self._rng = random.Random(seed)
self._task_id = task_id
self._tick_count = 0
self._failed_node_id = None
self._reroute_weights = {}
self._cascade_tick = 0
self._cascade_triggered = False
self.invalid_action_count = 0
self._randomize_domain()
self._reset_nodes()
def state(self, for_agent: bool = True) -> list[dict]:
"""Return current per-node state as a list of plain dicts."""
state_list = []
for n in self._nodes:
d = n.to_dict()
if for_agent and self._rng.random() < SENSOR_DROPOUT_PROB:
d["queue_depth"] = 0
d["latency_ms"] = -1.0
state_list.append(d)
return state_list
def apply_action(self, action_model) -> bool:
"""Update simulator world-state parameters based on agent command.
Returns:
True if the action had a measurable effect on state, False if it
was silently discarded (e.g. SCALE_UP at capacity cap).
"""
at = action_model.action_type if hasattr(action_model, "action_type") else action_model["action_type"]
node_id = action_model.target_node_id if hasattr(action_model, "target_node_id") else action_model["target_node_id"]
param = action_model.parameter if hasattr(action_model, "parameter") else action_model["parameter"]
# 1. NO_OP always succeeds regardless of target node
if at == "NO_OP":
return True
# 2. Target node lookup (required for all other actions)
target = next((n for n in self._nodes if n.node_id == node_id), None)
if not target:
return False
# 3. Command implementation
if at == "SCALE_UP":
delta = max(1, int(param * MAX_SCALING_STEP))
for _ in range(delta):
target.pending_capacity_queue.append(BOOT_DELAY_TICKS)
# If scaling up, we forcefully 'repair' DEGRADED status (SRE manual intervention)
was_degraded = target.status == NodeStatus.DEGRADED
if was_degraded:
target.status = NodeStatus.HEALTHY
return True # Always has effect — adds pending capacity, costs money
elif at == "SCALE_DOWN":
delta = max(1, int(param * MAX_SCALING_STEP))
# First cancel any pending capacity (cancel boot queue entries)
# This is like canceling a VM launch — it hasn't served traffic yet.
cancelled = 0
while cancelled < delta and target.pending_capacity_queue:
target.pending_capacity_queue.pop() # Remove newest pending first
cancelled += 1
# If still need to remove more, reduce active capacity
remaining = delta - cancelled
old_capacity = target.capacity
if remaining > 0:
target.capacity = max(1, target.capacity - remaining)
# Had effect if we cancelled pending or reduced active capacity
return cancelled > 0 or target.capacity != old_capacity
elif at == "REROUTE_TRAFFIC":
# Physically offload traffic FROM the target node by proportion `param`.
# `param` ∈ [0, 1]: fraction of target node's share to redistribute.
# The shed amount is split evenly across healthy peer nodes.
frac = min(1.0, max(0.0, float(param)))
# Record a reroute weight for the target so _inject_traffic() can
# reduce its lambda and bump peers accordingly.
self._reroute_weights[node_id] = frac
return True # Always has effect (sets a routing weight)
elif at == "SHED_LOAD":
# Rule: Cannot shed critical nodes (database/control plane).
# This forces the agent to handle Task-3 surge via Scaling/Rerouting.
if node_id in CRITICAL_NODES:
self.invalid_action_count += 1
return False # Action rejected on critical node
frac = min(1.0, param)
target.shed_fraction = frac
return True # Had effect (set shed fraction)
# Note: physically applied in _update_queues() to incoming traffic
# NO_OP or unrecognised action types — always intentional, count as having effect
return True
def tick(self) -> None:
"""
Advance simulated time by one step.
Evaluates physics: Capacity growth, Traffic injection, Queue processing.
"""
self._tick_count += 1
self._update_capacity()
self._inject_traffic()
# Save original capacities; backpressure temporarily reduces service_rate
# for this tick only. Restore after _update_queues so the reduction does
# not compound across ticks and permanently cripple parent nodes.
saved_capacities = {n.node_id: n.capacity for n in self._nodes}
self._apply_backpressure()
# Reset per-tick shed counters before physics update
for node in self._nodes:
node.dropped_requests = 0.0
self._update_queues()
# Restore capacities so the next tick starts from the true provisioned level
for n in self._nodes:
n.capacity = saved_capacities.get(n.node_id, n.capacity)
self._update_derived_metrics()
self._update_statuses()
self._cascade_failures()
self._process_recovery()
# Decay shed fractions gradually
# *= 0.5 retains 50% per tick (fast decay).
for node in self._nodes:
node.shed_fraction *= 0.5
if node.shed_fraction < 0.05:
node.shed_fraction = 0.0
def _update_capacity(self) -> None:
"""Process pending capacity from SCALE_UP actions"""
for node in self._nodes:
for i in range(len(node.pending_capacity_queue)):
node.pending_capacity_queue[i] -= 1
# Move ready capacity to live
ready = sum(1 for delay in node.pending_capacity_queue if delay <= 0)
node.capacity += ready
node.pending_capacity_queue = [delay for delay in node.pending_capacity_queue if delay > 0]
def _inject_traffic(self) -> None:
"""
Distribute traffic through the cluster DAG in three phases.
Phase 1 — Task lambda + scripted events:
Compute total external λ for this tick and apply any task-specific
mutations (node failure scripting, surge flags). No early returns;
all branches fall through to the shared DAG in Phase 2.
Phase 2 — Topological DAG distribution:
Traverse _TOPOLOGICAL_ORDER (roots first). Each parent's
processed outflow (min(incoming, service_rate)) is split across
its children via DEFAULT_ROUTING_SPLIT. A FAILED node has
service_rate=0, so outflow=0 and its children are naturally
starved — this is the causal failure chain the RL agent must
learn to route around.
Phase 3 — Reroute weight correction:
Apply REROUTE_TRAFFIC weight adjustments post-DAG, then decay
weights. Keeps reroute semantics identical to pre-DAG behaviour.
"""
# -------------------------------------------------------------------
# Phase 1: task-specific lambda + scripted events (no early returns)
# -------------------------------------------------------------------
total_lambda: float = 0.0
# direct_injections: extra traffic added directly to a node ON TOP OF
# the DAG distribution. Used for Task-3 surge bursts that model a
# side-channel load source (e.g. bulk import hitting checkout/catalog
# directly), while the base λ still travels through node-0 as ingress.
direct_injections: dict[str, float] = {}
if self._task_id == "task-1":
# Linear ramp — starts near cluster capacity
total_lambda = self._t1_init_lambda + (self._t1_ramp_slope * self._tick_count)
elif self._task_id == "task-2":
total_lambda = self._t2_init_lambda
# Scripted node failure fires at the configured tick
if self._tick_count >= self._t2_fail_tick and not self._failed_node_id:
self._failed_node_id = self._rng.choice(
[n.node_id for n in self._nodes if n.node_id != "node-0"]
)
target = next((n for n in self._nodes if n.node_id == self._failed_node_id), None)
if target:
target.is_scripted_failure = True
# No early return: DAG distributes traffic to the failed node normally.
# The dead node's service_rate=0 means outflow=0, so its children are
# starved. _update_queues() converts all its incoming traffic to
# dropped_requests. The agent must issue REROUTE_TRAFFIC to shift
# the parent's split away from the dead child.
elif self._task_id == "task-3":
total_lambda = T3_INITIAL_LAMBDA
phase = self._tick_count % T3_SURGE_CYCLE
if self._t3_surge_start <= phase <= self._t3_surge_end:
# Surge is modelled as a direct external burst arriving at the
# checkout (node-1) and catalog (node-2) services from a side
# channel that bypasses the payment gateway ingress.
# Base λ still routes through the DAG; the surge is overlaid so
# CRITICAL_NODE protections (no SHED_LOAD on node-1/2) still apply.
for nid in ["node-1", "node-2"]:
direct_injections[nid] = T3_SURGE_MAGNITUDE
# -------------------------------------------------------------------
# Phase 2: DAG topological distribution
# -------------------------------------------------------------------
node_incoming: dict[str, float] = {n.node_id: 0.0 for n in self._nodes}
node_map: dict[str, "NodeState"] = {n.node_id: n for n in self._nodes}
# Seed ingress nodes with their share of external λ
node_incoming["node-0"] = total_lambda * EXTERNAL_LAMBDA_FRACTION
node_incoming["node-4"] = total_lambda * (1.0 - EXTERNAL_LAMBDA_FRACTION)
# Overlay task-specific direct injections (Task-3 surge)
for nid, extra in direct_injections.items():
node_incoming[nid] = node_incoming.get(nid, 0.0) + extra
# Propagate outflow through the graph in topological order
for parent_id in _TOPOLOGICAL_ORDER:
parent = node_map.get(parent_id)
if parent is None:
continue
parent.incoming_request_rate = node_incoming[parent_id]
# Outflow = requests the parent actually forwards downstream.
# FAILED nodes have service_rate=0 → outflow=0 → children starved.
outflow = min(parent.incoming_request_rate, parent.service_rate)
parent.outflow_rate = outflow
for child_id, split in DEFAULT_ROUTING_SPLIT.get(parent_id, {}).items():
node_incoming[child_id] = node_incoming.get(child_id, 0.0) + outflow * split
# -------------------------------------------------------------------
# Phase 3: REROUTE_TRAFFIC weight corrections (post-DAG)
# -------------------------------------------------------------------
self._apply_reroute_weights()
# Recalculate outflow after reroute so the agent sees accurate
# per-node dispatch rates. Without this, a node whose incoming was
# halved by reroute would still report its pre-reroute outflow.
for n in self._nodes:
n.outflow_rate = min(n.incoming_request_rate, n.service_rate)
def _apply_reroute_weights(self) -> None:
"""
Apply REROUTE_TRAFFIC adjustments.
For each node with a reroute weight w, reduce its incoming_request_rate
by (w × its current rate) and distribute that offloaded load equally
across healthy peer nodes.
Important: rerouting must still work when the target is FAILED (Task 2).
We therefore allow offload FROM failed targets. Only absorber nodes are
constrained to non-failed nodes.
Decays each reroute weight by 50% per tick so the effect must be
re-issued to be maintained (forces the agent to act).
"""
if not self._reroute_weights:
return
total_overflow: float = 0.0
rerouted_ids = set(self._reroute_weights.keys())
for n in self._nodes:
w = self._reroute_weights.get(n.node_id, 0.0)
if w <= 0.0:
continue
# Offload from the target regardless of node status.
offload = n.incoming_request_rate * w
if offload <= 0.0:
continue
n.incoming_request_rate -= offload
total_overflow += offload
# Spread overflow across all non-rerouted healthy nodes
if total_overflow > 0:
absorbers = [
n for n in self._nodes
if n.node_id not in rerouted_ids
and n.status != NodeStatus.FAILED
]
# If every healthy node is rerouted this tick, fall back to all
# healthy non-rerouted nodes first, then finally all healthy nodes.
# CRITICAL: A FAILED node should NEVER be an absorber.
if not absorbers:
absorbers = [n for n in self._nodes if n.status != NodeStatus.FAILED and n.node_id not in rerouted_ids]
if not absorbers:
absorbers = [n for n in self._nodes if n.status != NodeStatus.FAILED]
if absorbers:
share = total_overflow / len(absorbers)
for n in absorbers:
n.incoming_request_rate += share
# Decay weights — agent must keep re-issuing to maintain effect
# *= 0.5 retains 50% per tick (fast decay).
for nid in list(self._reroute_weights.keys()):
self._reroute_weights[nid] *= 0.5
if self._reroute_weights[nid] < 0.05:
del self._reroute_weights[nid]
def _apply_backpressure(self) -> None:
"""Reduce parent service rate when children are overloaded."""
for parent_id, children in CLUSTER_TOPOLOGY.items():
parent = next((n for n in self._nodes if n.node_id == parent_id), None)
if not children or not parent or parent.status == NodeStatus.FAILED:
continue
# Compute pressure from overloaded children
total_pressure = 0.0
for child_id in children:
child = next((n for n in self._nodes if n.node_id == child_id), None)
if child:
excess = max(0.0, child.queue_depth - BACKPRESSURE_THRESHOLD)
total_pressure += excess / FATAL_FAIL_THRESHOLD # normalise to [0, 1]
# Reduce parent's effective capacity proportionally
pressure_factor = min(BACKPRESSURE_MAX_FACTOR, total_pressure * 0.6)
parent.capacity = max(1.0, parent.capacity * (1.0 - pressure_factor))
def _update_queues(self) -> None:
"""
Fluid-queue update for all nodes.
Q_i(t+1) = max(Q_i(t) + λ_eff_i(t) − μ_i(t), 0)
where λ_eff = λ_incoming * (1 - shed_fraction).
"""
for node in self._nodes:
if node.status == NodeStatus.FAILED:
node.queue_depth = 0.0
node.dropped_requests = node.incoming_request_rate
continue
# Apply shedding to incoming traffic
lambda_eff = node.incoming_request_rate * (1.0 - node.shed_fraction)
node.dropped_requests = node.incoming_request_rate - lambda_eff
excess = lambda_eff - node.service_rate
node.queue_depth = max(0.0, node.queue_depth + excess)
def _update_derived_metrics(self) -> None:
"""Compute CPU and Latency based on current queue and status."""
for n in self._nodes:
if n.status == NodeStatus.FAILED:
n.cpu_utilization = 0.0
n.latency_ms = float("inf")
continue
# Utilization = Ratio of λ to μ
service_rate = n.service_rate
n.cpu_utilization = n.incoming_request_rate / service_rate if service_rate > 0 else 1.0
# Latency: Hybrid M/M/1 + backlog term
# M/M/1 gives exponential blow-up as utilization->1 (the "hockey stick")
# Backlog term ensures queue_depth still contributes signal even when
# utilization is capped at 0.99, preventing the flattening problem.
utilization = min(0.99, n.cpu_utilization) # cap to prevent infinity
mm1_latency = BASE_LATENCY_MS / (1.0 - utilization)
backlog_latency = n.queue_depth * LATENCY_STEEPNESS
n.latency_ms = mm1_latency + backlog_latency
def _update_statuses(self) -> None:
"""Transition node health based on queue boundaries.
Recovery rules:
- Scripted failures (Task 2 forced node kill): permanent, never auto-recover.
Marked by is_scripted_failure=True, recovery_timer=0.
- Overload failures (queue > FATAL_FAIL_THRESHOLD): auto-recover after
NODE_RECOVERY_TICKS. The agent can learn to reroute away and let the
node heal.
"""
for n in self._nodes:
# Scripted (task-forced) failures are permanent
if n.is_scripted_failure:
n.status = NodeStatus.FAILED
n.recovery_timer = 0
continue
if n.queue_depth > FATAL_FAIL_THRESHOLD:
if n.status != NodeStatus.FAILED:
n.status = NodeStatus.FAILED
n.recovery_timer = NODE_RECOVERY_TICKS
n.capacity = 0.5 # starts at half capacity when recovery begins
self._cascade_triggered = True # Signal cascade detection
elif n.queue_depth > OVERLOAD_THRESHOLD:
n.status = NodeStatus.DEGRADED
elif n.status == NodeStatus.DEGRADED and n.queue_depth < (OVERLOAD_THRESHOLD / 2):
n.status = NodeStatus.HEALTHY
def _cascade_failures(self) -> None:
"""Detect cascading failure: if a peer node's queue exceeds a heightened
threshold within CASCADE_WINDOW_TICKS of a *new* failure, degrade it.
Guardrails:
- Only triggers when a NEW failure occurred this tick (not any failed node).
- Graph-bounded: cascade only propagates along edges (parents or children).
- Scripted failures (Task 2) do not trigger cascades.
"""
if not self._cascade_triggered:
self._cascade_tick = 0
return
self._cascade_tick += 1
if self._cascade_tick > CASCADE_WINDOW_TICKS:
self._cascade_triggered = False
self._cascade_tick = 0
return
# Find all currently failed nodes
failed_ids = {n.node_id for n in self._nodes if n.status == NodeStatus.FAILED}
# Build set of nodes adjacent to any failed node (upstream or downstream)
at_risk = set()
for failed_id in failed_ids:
# Downstream children of the failed node
at_risk.update(CLUSTER_TOPOLOGY.get(failed_id, []))
# Upstream parents of the failed node
for parent_id, children in CLUSTER_TOPOLOGY.items():
if failed_id in children:
at_risk.add(parent_id)
cascade_threshold = FATAL_FAIL_THRESHOLD * CASCADE_QUEUE_MULTIPLIER
for n in self._nodes:
if n.node_id not in at_risk:
continue # Not adjacent to failure — cannot cascade
if n.status == NodeStatus.FAILED or n.is_scripted_failure:
continue
if n.queue_depth > cascade_threshold:
n.status = NodeStatus.DEGRADED
def _process_recovery(self) -> None:
"""Count down recovery timers and bring FAILED nodes back online.
Only overload-failed nodes (recovery_timer > 0) can recover.
Scripted failures (is_scripted_failure=True) are excluded.
"""
RECOVERY_RAMP_PER_TICK: float = 0.5 # capacity added per tick during recovery
for n in self._nodes:
if n.is_scripted_failure:
continue
# Check recovery_timer > 0, not status == FAILED: the first recovery
# tick transitions the node to DEGRADED, but the timer must keep
# counting until it reaches 0 and the node becomes HEALTHY.
if n.recovery_timer > 0:
n.recovery_timer -= 1
if n.recovery_timer <= 0:
n.status = NodeStatus.HEALTHY
n.queue_depth = 0.0
n.latency_ms = BASE_LATENCY_MS
n.cpu_utilization = 0.0
# capacity stays at whatever it ramped to
else:
# Still in recovery: ramp capacity up, stay DEGRADED
n.capacity = min(DEFAULT_CAPACITY, n.capacity + RECOVERY_RAMP_PER_TICK)
n.status = NodeStatus.DEGRADED # not HEALTHY until fully ramped
def reconcile_state(self, telemetry_map: dict) -> None:
"""
Reconcile internal simulator state with external telemetry signals.
Used in 'hybrid' or 'live' modes to align the physics engine with reality.
telemetry_map: node_id -> TelemetryRecord (or dict)
"""
for node in self._nodes:
if node.node_id in telemetry_map:
record = telemetry_map[node.node_id]
# If record is an object (TelemetryRecord), access fields; otherwise treat as dict
if hasattr(record, "queue_depth"):
q_ext = float(record.queue_depth)
r_ext = float(record.request_rate)
c_ext = float(record.cpu_utilization)
e_ext = float(record.error_rate)
l_ext = float(record.latency_ms)
else:
q_ext = float(record.get("queue_depth", node.queue_depth))
r_ext = float(record.get("request_rate", node.incoming_request_rate))
c_ext = float(record.get("cpu_utilization", node.cpu_utilization))
e_ext = float(record.get("error_rate", 0.0))
l_ext = float(record.get("latency_ms", node.latency_ms))
# Smoothly blend the external state into physics to prevent step jumps
# trust: 0.7 (reality) / 0.3 (simulation prediction)
node.queue_depth = (node.queue_depth * 0.3) + (q_ext * 0.7)
node.incoming_request_rate = (node.incoming_request_rate * 0.3) + (r_ext * 0.7)
node.cpu_utilization = (node.cpu_utilization * 0.3) + (c_ext * 0.7)
node.latency_ms = (node.latency_ms * 0.3) + (l_ext * 0.7)
# Status reconciliation: Fail the node if error rate is high
if e_ext > 0.5:
node.status = NodeStatus.FAILED
elif e_ext > 0.1 and node.status == NodeStatus.HEALTHY:
node.status = NodeStatus.DEGRADED
elif e_ext <= 0.05 and node.status == NodeStatus.DEGRADED:
# Allow recovery if telemetry says it's clean (physics will still check queue)
node.status = NodeStatus.HEALTHY
# Crucial: re-derive metrics so latency/cpu are consistent (except for the blended values)
# Note: We blend latency/cpu above, but _update_derived_metrics might overwrite them.
# So we update them after blending if we want physics to win, or before if telemetry wins.
# Usually, for SRE dashboard, we want the blended 'reality'.
# However, _update_derived_metrics is used to compute 'current' state in pure sim.
# We'll skip it if we just reconciled to keep the blended values, OR refine it.
# For now, let's just make sure statuses are updated based on new queue depths.
self._update_statuses()
|