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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711
// Copyright Materialize, Inc. and contributors. All rights reserved.
//
// Use of this software is governed by the Business Source License
// included in the LICENSE file.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0.
use std::collections::BTreeMap;
use std::future::Future;
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};
use std::{env, fmt};
use anyhow::{anyhow, Context};
use async_trait::async_trait;
use chrono::Utc;
use clap::ArgEnum;
use cloud_resource_controller::KubernetesResourceReader;
use futures::stream::{BoxStream, StreamExt};
use futures::TryFutureExt;
use k8s_openapi::api::apps::v1::{StatefulSet, StatefulSetSpec};
use k8s_openapi::api::core::v1::{
Affinity, Capabilities, Container, ContainerPort, EnvVar, EnvVarSource, EphemeralVolumeSource,
NodeAffinity, NodeSelector, NodeSelectorRequirement, NodeSelectorTerm, ObjectFieldSelector,
ObjectReference, PersistentVolumeClaim, PersistentVolumeClaimSpec,
PersistentVolumeClaimTemplate, Pod, PodAffinity, PodAffinityTerm, PodAntiAffinity,
PodSecurityContext, PodSpec, PodTemplateSpec, PreferredSchedulingTerm, ResourceRequirements,
SeccompProfile, Secret, SecurityContext, Service as K8sService, ServicePort, ServiceSpec,
Toleration, TopologySpreadConstraint, Volume, VolumeMount, VolumeResourceRequirements,
WeightedPodAffinityTerm,
};
use k8s_openapi::apimachinery::pkg::api::resource::Quantity;
use k8s_openapi::apimachinery::pkg::apis::meta::v1::{
LabelSelector, LabelSelectorRequirement, OwnerReference,
};
use kube::api::{Api, DeleteParams, ObjectMeta, Patch, PatchParams};
use kube::client::Client;
use kube::error::Error as K8sError;
use kube::runtime::{watcher, WatchStreamExt};
use kube::ResourceExt;
use maplit::btreemap;
use mz_cloud_resources::crd::vpc_endpoint::v1::VpcEndpoint;
use mz_cloud_resources::AwsExternalIdPrefix;
use mz_orchestrator::{
scheduling_config::*, DiskLimit, LabelSelectionLogic, LabelSelector as MzLabelSelector,
NamespacedOrchestrator, OfflineReason, Orchestrator, Service, ServiceConfig, ServiceEvent,
ServiceProcessMetrics, ServiceStatus,
};
use mz_ore::retry::Retry;
use mz_ore::task::AbortOnDropHandle;
use serde::Deserialize;
use sha2::{Digest, Sha256};
use tokio::sync::{mpsc, oneshot};
use tracing::{info, warn};
pub mod cloud_resource_controller;
pub mod secrets;
pub mod util;
const FIELD_MANAGER: &str = "environmentd";
const NODE_FAILURE_THRESHOLD_SECONDS: i64 = 30;
const POD_TEMPLATE_HASH_ANNOTATION: &str = "environmentd.materialize.cloud/pod-template-hash";
/// Configures a [`KubernetesOrchestrator`].
#[derive(Debug, Clone)]
pub struct KubernetesOrchestratorConfig {
/// The name of a Kubernetes context to use, if the Kubernetes configuration
/// is loaded from the local kubeconfig.
pub context: String,
/// The name of a non-default Kubernetes scheduler to use, if any.
pub scheduler_name: Option<String>,
/// Labels to install on every service created by the orchestrator.
pub service_labels: BTreeMap<String, String>,
/// Node selector to install on every service created by the orchestrator.
pub service_node_selector: BTreeMap<String, String>,
/// The service account that each service should run as, if any.
pub service_account: Option<String>,
/// The image pull policy to set for services created by the orchestrator.
pub image_pull_policy: KubernetesImagePullPolicy,
/// An AWS external ID prefix to use when making AWS operations on behalf
/// of the environment.
pub aws_external_id_prefix: Option<AwsExternalIdPrefix>,
/// Whether to use code coverage mode or not. Always false for production.
pub coverage: bool,
/// The Kubernetes StorageClass to use for the ephemeral volume attached to
/// services that request disk.
///
/// If unspecified, the orchestrator will refuse to create services that
/// request disk.
pub ephemeral_volume_storage_class: Option<String>,
/// The optional fs group for service's pods' `securityContext`.
pub service_fs_group: Option<i64>,
}
/// Specifies whether Kubernetes should pull Docker images when creating pods.
#[derive(ArgEnum, Debug, Clone, Copy)]
pub enum KubernetesImagePullPolicy {
/// Always pull the Docker image from the registry.
Always,
/// Pull the Docker image only if the image is not present.
IfNotPresent,
/// Never pull the Docker image.
Never,
}
impl fmt::Display for KubernetesImagePullPolicy {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
match self {
KubernetesImagePullPolicy::Always => f.write_str("Always"),
KubernetesImagePullPolicy::IfNotPresent => f.write_str("IfNotPresent"),
KubernetesImagePullPolicy::Never => f.write_str("Never"),
}
}
}
/// An orchestrator backed by Kubernetes.
pub struct KubernetesOrchestrator {
client: Client,
kubernetes_namespace: String,
config: KubernetesOrchestratorConfig,
secret_api: Api<Secret>,
vpc_endpoint_api: Api<VpcEndpoint>,
namespaces: Mutex<BTreeMap<String, Arc<dyn NamespacedOrchestrator>>>,
resource_reader: Arc<KubernetesResourceReader>,
}
impl fmt::Debug for KubernetesOrchestrator {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
f.debug_struct("KubernetesOrchestrator").finish()
}
}
impl KubernetesOrchestrator {
/// Creates a new Kubernetes orchestrator from the provided configuration.
pub async fn new(
config: KubernetesOrchestratorConfig,
) -> Result<KubernetesOrchestrator, anyhow::Error> {
let (client, kubernetes_namespace) = util::create_client(config.context.clone()).await?;
let resource_reader =
Arc::new(KubernetesResourceReader::new(config.context.clone()).await?);
Ok(KubernetesOrchestrator {
client: client.clone(),
kubernetes_namespace,
config,
secret_api: Api::default_namespaced(client.clone()),
vpc_endpoint_api: Api::default_namespaced(client),
namespaces: Mutex::new(BTreeMap::new()),
resource_reader,
})
}
}
impl Orchestrator for KubernetesOrchestrator {
fn namespace(&self, namespace: &str) -> Arc<dyn NamespacedOrchestrator> {
let mut namespaces = self.namespaces.lock().expect("lock poisoned");
Arc::clone(namespaces.entry(namespace.into()).or_insert_with(|| {
let (command_tx, command_rx) = mpsc::unbounded_channel();
let worker = OrchestratorWorker {
metrics_api: Api::default_namespaced(self.client.clone()),
custom_metrics_api: Api::default_namespaced(self.client.clone()),
service_api: Api::default_namespaced(self.client.clone()),
stateful_set_api: Api::default_namespaced(self.client.clone()),
pod_api: Api::default_namespaced(self.client.clone()),
owner_references: vec![],
command_rx,
}
.spawn(format!("kubernetes-orchestrator-worker:{namespace}"));
Arc::new(NamespacedKubernetesOrchestrator {
pod_api: Api::default_namespaced(self.client.clone()),
kubernetes_namespace: self.kubernetes_namespace.clone(),
namespace: namespace.into(),
config: self.config.clone(),
// TODO(guswynn): make this configurable.
scheduling_config: Default::default(),
service_infos: std::sync::Mutex::new(BTreeMap::new()),
command_tx,
_worker: worker,
})
}))
}
}
#[derive(Clone, Copy)]
struct ServiceInfo {
scale: u16,
disk: bool,
disk_limit: Option<DiskLimit>,
}
struct NamespacedKubernetesOrchestrator {
pod_api: Api<Pod>,
kubernetes_namespace: String,
namespace: String,
config: KubernetesOrchestratorConfig,
scheduling_config: std::sync::RwLock<ServiceSchedulingConfig>,
service_infos: std::sync::Mutex<BTreeMap<String, ServiceInfo>>,
command_tx: mpsc::UnboundedSender<WorkerCommand>,
_worker: AbortOnDropHandle<()>,
}
impl fmt::Debug for NamespacedKubernetesOrchestrator {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
f.debug_struct("NamespacedKubernetesOrchestrator")
.field("kubernetes_namespace", &self.kubernetes_namespace)
.field("namespace", &self.namespace)
.field("config", &self.config)
.finish()
}
}
/// Commands sent from a [`NamespacedKubernetesOrchestrator`] to its
/// [`OrchestratorWorker`].
///
/// Commands for which the caller expects a result include a `result_tx` on which the
/// [`OrchestratorWorker`] will deliver the result.
enum WorkerCommand {
EnsureService {
desc: ServiceDescription,
},
DropService {
name: String,
},
ListServices {
namespace: String,
result_tx: oneshot::Sender<Vec<String>>,
},
FetchServiceMetrics {
name: String,
info: ServiceInfo,
result_tx: oneshot::Sender<Vec<ServiceProcessMetrics>>,
},
}
/// A description of a service to be created by an [`OrchestratorWorker`].
#[derive(Debug, Clone)]
struct ServiceDescription {
name: String,
scale: u16,
service: K8sService,
stateful_set: StatefulSet,
pod_template_hash: String,
}
/// A task executing blocking work for a [`NamespacedKubernetesOrchestrator`] in the background.
///
/// This type exists to enable making [`NamespacedKubernetesOrchestrator::ensure_service`] and
/// [`NamespacedKubernetesOrchestrator::drop_service`] non-blocking, allowing invocation of these
/// methods in latency-sensitive contexts.
///
/// Note that, apart from `ensure_service` and `drop_service`, this worker also handles blocking
/// orchestrator calls that query service state (such as `list_services`). These need to be
/// sequenced through the worker loop to ensure they linearize as expected. For example, we want to
/// ensure that a `list_services` result contains exactly those services that were previously
/// created with `ensure_service` and not yet dropped with `drop_service`.
struct OrchestratorWorker {
metrics_api: Api<PodMetrics>,
custom_metrics_api: Api<MetricValueList>,
service_api: Api<K8sService>,
stateful_set_api: Api<StatefulSet>,
pod_api: Api<Pod>,
owner_references: Vec<OwnerReference>,
command_rx: mpsc::UnboundedReceiver<WorkerCommand>,
}
#[derive(Deserialize, Clone, Debug)]
pub struct PodMetricsContainer {
pub name: String,
pub usage: PodMetricsContainerUsage,
}
#[derive(Deserialize, Clone, Debug)]
pub struct PodMetricsContainerUsage {
pub cpu: Quantity,
pub memory: Quantity,
}
#[derive(Deserialize, Clone, Debug)]
pub struct PodMetrics {
pub metadata: ObjectMeta,
pub timestamp: String,
pub window: String,
pub containers: Vec<PodMetricsContainer>,
}
impl k8s_openapi::Resource for PodMetrics {
const GROUP: &'static str = "metrics.k8s.io";
const KIND: &'static str = "PodMetrics";
const VERSION: &'static str = "v1beta1";
const API_VERSION: &'static str = "metrics.k8s.io/v1beta1";
const URL_PATH_SEGMENT: &'static str = "pods";
type Scope = k8s_openapi::NamespaceResourceScope;
}
impl k8s_openapi::Metadata for PodMetrics {
type Ty = ObjectMeta;
fn metadata(&self) -> &Self::Ty {
&self.metadata
}
fn metadata_mut(&mut self) -> &mut Self::Ty {
&mut self.metadata
}
}
// Note that these types are very weird. We are `get`-ing a
// `List` object, and lying about it having an `ObjectMeta`
// (it deserializes as empty, but we don't need it). The custom
// metrics API is designed this way, which is very non-standard.
// A discussion in the `kube` channel in the `tokio` discord
// confirmed that this layout + using `get_subresource` is the
// best way to handle this.
#[derive(Deserialize, Clone, Debug)]
pub struct MetricIdentifier {
#[serde(rename = "metricName")]
pub name: String,
// We skip `selector` for now, as we don't use it
}
#[derive(Deserialize, Clone, Debug)]
pub struct MetricValue {
#[serde(rename = "describedObject")]
pub described_object: ObjectReference,
#[serde(flatten)]
pub metric_identifier: MetricIdentifier,
pub timestamp: String,
pub value: Quantity,
// We skip `windowSeconds`, as we don't need it
}
#[derive(Deserialize, Clone, Debug)]
pub struct MetricValueList {
pub metadata: ObjectMeta,
pub items: Vec<MetricValue>,
}
impl k8s_openapi::Resource for MetricValueList {
const GROUP: &'static str = "custom.metrics.k8s.io";
const KIND: &'static str = "MetricValueList";
const VERSION: &'static str = "v1beta1";
const API_VERSION: &'static str = "custom.metrics.k8s.io/v1beta1";
const URL_PATH_SEGMENT: &'static str = "persistentvolumeclaims";
type Scope = k8s_openapi::NamespaceResourceScope;
}
impl k8s_openapi::Metadata for MetricValueList {
type Ty = ObjectMeta;
fn metadata(&self) -> &Self::Ty {
&self.metadata
}
fn metadata_mut(&mut self) -> &mut Self::Ty {
&mut self.metadata
}
}
impl NamespacedKubernetesOrchestrator {
fn service_name(&self, id: &str) -> String {
format!("{}-{id}", self.namespace)
}
/// Return a `watcher::Config` instance that limits results to the namespace
/// assigned to this orchestrator.
fn watch_pod_params(&self) -> watcher::Config {
let ns_selector = format!(
"environmentd.materialize.cloud/namespace={}",
self.namespace
);
watcher::Config::default().labels(&ns_selector)
}
/// Convert a higher-level label key to the actual one we
/// will give to Kubernetes
fn make_label_key(&self, key: &str) -> String {
format!("{}.environmentd.materialize.cloud/{}", self.namespace, key)
}
fn label_selector_to_k8s(
&self,
MzLabelSelector { label_name, logic }: MzLabelSelector,
) -> Result<LabelSelectorRequirement, anyhow::Error> {
let (operator, values) = match logic {
LabelSelectionLogic::Eq { value } => Ok(("In", vec![value])),
LabelSelectionLogic::NotEq { value } => Ok(("NotIn", vec![value])),
LabelSelectionLogic::Exists => Ok(("Exists", vec![])),
LabelSelectionLogic::NotExists => Ok(("DoesNotExist", vec![])),
LabelSelectionLogic::InSet { values } => {
if values.is_empty() {
Err(anyhow!(
"Invalid selector logic for {label_name}: empty `in` set"
))
} else {
Ok(("In", values))
}
}
LabelSelectionLogic::NotInSet { values } => {
if values.is_empty() {
Err(anyhow!(
"Invalid selector logic for {label_name}: empty `notin` set"
))
} else {
Ok(("NotIn", values))
}
}
}?;
let lsr = LabelSelectorRequirement {
key: self.make_label_key(&label_name),
operator: operator.to_string(),
values: Some(values),
};
Ok(lsr)
}
fn send_command(&self, cmd: WorkerCommand) {
self.command_tx.send(cmd).expect("worker task not dropped");
}
}
#[derive(Debug)]
struct ScaledQuantity {
integral_part: u64,
exponent: i8,
base10: bool,
}
impl ScaledQuantity {
pub fn try_to_integer(&self, scale: i8, base10: bool) -> Option<u64> {
if base10 != self.base10 {
return None;
}
let exponent = self.exponent - scale;
let mut result = self.integral_part;
let base = if self.base10 { 10 } else { 2 };
if exponent < 0 {
for _ in exponent..0 {
result /= base;
}
} else {
for _ in 0..exponent {
result = result.checked_mul(2)?;
}
}
Some(result)
}
}
// Parse a k8s `Quantity` object
// into a numeric value.
//
// This is intended to support collecting CPU and Memory data.
// Thus, there are a few that things Kubernetes attempts to do, that we don't,
// because I've never observed metrics-server specifically sending them:
// (1) Handle negative numbers (because it's not useful for that use-case)
// (2) Handle non-integers (because I have never observed them being actually sent)
// (3) Handle scientific notation (e.g. 1.23e2)
fn parse_k8s_quantity(s: &str) -> Result<ScaledQuantity, anyhow::Error> {
const DEC_SUFFIXES: &[(&str, i8)] = &[
("n", -9),
("u", -6),
("m", -3),
("", 0),
("k", 3), // yep, intentionally lowercase.
("M", 6),
("G", 9),
("T", 12),
("P", 15),
("E", 18),
];
const BIN_SUFFIXES: &[(&str, i8)] = &[
("", 0),
("Ki", 10),
("Mi", 20),
("Gi", 30),
("Ti", 40),
("Pi", 50),
("Ei", 60),
];
let (positive, s) = match s.chars().next() {
Some('+') => (true, &s[1..]),
Some('-') => (false, &s[1..]),
_ => (true, s),
};
if !positive {
anyhow::bail!("Negative numbers not supported")
}
fn is_suffix_char(ch: char) -> bool {
"numkMGTPEKi".contains(ch)
}
let (num, suffix) = match s.find(is_suffix_char) {
None => (s, ""),
Some(idx) => s.split_at(idx),
};
let num: u64 = num.parse()?;
let (exponent, base10) = if let Some((_, exponent)) =
DEC_SUFFIXES.iter().find(|(target, _)| suffix == *target)
{
(exponent, true)
} else if let Some((_, exponent)) = BIN_SUFFIXES.iter().find(|(target, _)| suffix == *target) {
(exponent, false)
} else {
anyhow::bail!("Unrecognized suffix: {suffix}");
};
Ok(ScaledQuantity {
integral_part: num,
exponent: *exponent,
base10,
})
}
#[async_trait]
impl NamespacedOrchestrator for NamespacedKubernetesOrchestrator {
async fn fetch_service_metrics(
&self,
id: &str,
) -> Result<Vec<ServiceProcessMetrics>, anyhow::Error> {
let info = if let Some(info) = self.service_infos.lock().expect("poisoned lock").get(id) {
*info
} else {
// This should have been set in `ensure_service`.
tracing::error!("Failed to get info for {id}");
anyhow::bail!("Failed to get info for {id}");
};
let (result_tx, result_rx) = oneshot::channel();
self.send_command(WorkerCommand::FetchServiceMetrics {
name: self.service_name(id),
info,
result_tx,
});
let metrics = result_rx.await.expect("worker task not dropped");
Ok(metrics)
}
fn ensure_service(
&self,
id: &str,
ServiceConfig {
image,
init_container_image,
args,
ports: ports_in,
memory_limit,
cpu_limit,
scale,
labels: labels_in,
availability_zones,
other_replicas_selector,
replicas_selector,
disk,
disk_limit,
node_selector,
}: ServiceConfig,
) -> Result<Box<dyn Service>, anyhow::Error> {
// This is extremely cheap to clone, so just look into the lock once.
let scheduling_config: ServiceSchedulingConfig =
self.scheduling_config.read().expect("poisoned").clone();
let disk = scheduling_config.always_use_disk || disk;
let name = self.service_name(id);
// The match labels should be the minimal set of labels that uniquely
// identify the pods in the stateful set. Changing these after the
// `StatefulSet` is created is not permitted by Kubernetes, and we're
// not yet smart enough to handle deleting and recreating the
// `StatefulSet`.
let match_labels = btreemap! {
"environmentd.materialize.cloud/namespace".into() => self.namespace.clone(),
"environmentd.materialize.cloud/service-id".into() => id.into(),
};
let mut labels = match_labels.clone();
for (key, value) in labels_in {
labels.insert(self.make_label_key(&key), value);
}
labels.insert(self.make_label_key("scale"), scale.to_string());
for port in &ports_in {
labels.insert(
format!("environmentd.materialize.cloud/port-{}", port.name),
"true".into(),
);
}
for (key, value) in &self.config.service_labels {
labels.insert(key.clone(), value.clone());
}
let mut limits = BTreeMap::new();
if let Some(memory_limit) = memory_limit {
limits.insert(
"memory".into(),
Quantity(memory_limit.0.as_u64().to_string()),
);
}
if let Some(cpu_limit) = cpu_limit {
limits.insert(
"cpu".into(),
Quantity(format!("{}m", cpu_limit.as_millicpus())),
);
}
let service = K8sService {
metadata: ObjectMeta {
name: Some(name.clone()),
..Default::default()
},
spec: Some(ServiceSpec {
ports: Some(
ports_in
.iter()
.map(|port| ServicePort {
port: port.port_hint.into(),
name: Some(port.name.clone()),
..Default::default()
})
.collect(),
),
cluster_ip: None,
selector: Some(match_labels.clone()),
..Default::default()
}),
status: None,
};
let hosts = (0..scale)
.map(|i| {
format!(
"{name}-{i}.{name}.{}.svc.cluster.local",
self.kubernetes_namespace
)
})
.collect::<Vec<_>>();
let ports = ports_in
.iter()
.map(|p| (p.name.clone(), p.port_hint))
.collect::<BTreeMap<_, _>>();
let listen_addrs = ports_in
.iter()
.map(|p| (p.name.clone(), format!("0.0.0.0:{}", p.port_hint)))
.collect::<BTreeMap<_, _>>();
let mut args = args(&listen_addrs);
// This constrains the orchestrator (for those orchestrators that support
// anti-affinity, today just k8s) to never schedule pods for different replicas
// of the same cluster on the same node. Pods from the _same_ replica are fine;
// pods from different clusters are also fine.
//
// The point is that if pods of two replicas are on the same node, that node
// going down would kill both replicas, and so the replication factor of the
// cluster in question is illusory.
let anti_affinity = Some({
let label_selector_requirements = other_replicas_selector
.clone()
.into_iter()
.map(|ls| self.label_selector_to_k8s(ls))
.collect::<Result<Vec<_>, _>>()?;
let ls = LabelSelector {
match_expressions: Some(label_selector_requirements),
..Default::default()
};
let pat = PodAffinityTerm {
label_selector: Some(ls),
topology_key: "kubernetes.io/hostname".to_string(),
..Default::default()
};
if !scheduling_config.soften_replication_anti_affinity {
PodAntiAffinity {
required_during_scheduling_ignored_during_execution: Some(vec![pat]),
..Default::default()
}
} else {
PodAntiAffinity {
preferred_during_scheduling_ignored_during_execution: Some(vec![
WeightedPodAffinityTerm {
weight: scheduling_config.soften_replication_anti_affinity_weight,
pod_affinity_term: pat,
},
]),
..Default::default()
}
}
});
let pod_affinity = if let Some(weight) = scheduling_config.multi_pod_az_affinity_weight {
// `match_labels` sufficiently selects pods in the same replica.
let ls = LabelSelector {
match_labels: Some(match_labels.clone()),
..Default::default()
};
let pat = PodAffinityTerm {
label_selector: Some(ls),
topology_key: "topology.kubernetes.io/zone".to_string(),
..Default::default()
};
Some(PodAffinity {
preferred_during_scheduling_ignored_during_execution: Some(vec![
WeightedPodAffinityTerm {
weight,
pod_affinity_term: pat,
},
]),
..Default::default()
})
} else {
None
};
let topology_spread = if scheduling_config.topology_spread.enabled {
let config = &scheduling_config.topology_spread;
if !config.ignore_non_singular_scale || scale <= 1 {
let label_selector_requirements = (if config.ignore_non_singular_scale {
let mut replicas_selector_ignoring_scale = replicas_selector.clone();
replicas_selector_ignoring_scale.push(mz_orchestrator::LabelSelector {
label_name: "scale".into(),
logic: mz_orchestrator::LabelSelectionLogic::Eq {
value: "1".to_string(),
},
});
replicas_selector_ignoring_scale
} else {
replicas_selector
})
.into_iter()
.map(|ls| self.label_selector_to_k8s(ls))
.collect::<Result<Vec<_>, _>>()?;
let ls = LabelSelector {
match_expressions: Some(label_selector_requirements),
..Default::default()
};
let constraint = TopologySpreadConstraint {
label_selector: Some(ls),
min_domains: None,
max_skew: config.max_skew,
topology_key: "topology.kubernetes.io/zone".to_string(),
when_unsatisfiable: if config.soft {
"ScheduleAnyway".to_string()
} else {
"DoNotSchedule".to_string()
},
// TODO(guswynn): restore these once they are supported.
// Consider node affinities when calculating topology spread. This is the
// default: <https://docs.rs/k8s-openapi/latest/k8s_openapi/api/core/v1/struct.TopologySpreadConstraint.html#structfield.node_affinity_policy>,
// made explicit.
// node_affinity_policy: Some("Honor".to_string()),
// Do not consider node taints when calculating topology spread. This is the
// default: <https://docs.rs/k8s-openapi/latest/k8s_openapi/api/core/v1/struct.TopologySpreadConstraint.html#structfield.node_taints_policy>,
// made explicit.
// node_taints_policy: Some("Ignore".to_string()),
match_label_keys: None,
// Once the above are restorted, we should't have `..Default::default()` here because the specifics of these fields are
// subtle enough where we want compilation failures when we upgrade
..Default::default()
};
Some(vec![constraint])
} else {
None
}
} else {
None
};
let pod_annotations = btreemap! {
// Prevent the cluster-autoscaler (or karpenter) from evicting these pods in attempts to scale down
// and terminate nodes.
// This will cost us more money, but should give us better uptime.
// This does not prevent all evictions by Kubernetes, only the ones initiated by the
// cluster-autoscaler (or karpenter). Notably, eviction of pods for resource overuse is still enabled.
"cluster-autoscaler.kubernetes.io/safe-to-evict".to_owned() => "false".to_string(),
"karpenter.sh/do-not-evict".to_owned() => "true".to_string(),
// It's called do-not-disrupt in newer versions of karpenter, so adding for forward/backward compatibility
"karpenter.sh/do-not-disrupt".to_owned() => "true".to_string(),
};
let default_node_selector = if disk {
vec![("materialize.cloud/disk".to_string(), disk.to_string())]
} else {
// if the cluster doesn't require disk, we can omit the selector
// allowing it to be scheduled onto nodes with and without the
// selector
vec![]
};
let node_selector: BTreeMap<String, String> = default_node_selector
.into_iter()
.chain(self.config.service_node_selector.clone())
.chain(node_selector)
.collect();
let node_affinity = if let Some(availability_zones) = availability_zones {
let selector = NodeSelectorTerm {
match_expressions: Some(vec![NodeSelectorRequirement {
key: "materialize.cloud/availability-zone".to_string(),
operator: "In".to_string(),
values: Some(availability_zones),
}]),
match_fields: None,
};
if scheduling_config.soften_az_affinity {
Some(NodeAffinity {
preferred_during_scheduling_ignored_during_execution: Some(vec![
PreferredSchedulingTerm {
preference: selector,
weight: scheduling_config.soften_az_affinity_weight,
},
]),
required_during_scheduling_ignored_during_execution: None,
})
} else {
Some(NodeAffinity {
preferred_during_scheduling_ignored_during_execution: None,
required_during_scheduling_ignored_during_execution: Some(NodeSelector {
node_selector_terms: vec![selector],
}),
})
}
} else {
None
};
let container_name = image
.splitn(2, '/')
.skip(1)
.next()
.and_then(|name_version| name_version.splitn(2, ':').next())
.context("`image` is not ORG/NAME:VERSION")?
.to_string();
let container_security_context = if scheduling_config.security_context_enabled {
Some(SecurityContext {
privileged: Some(false),
run_as_non_root: Some(true),
allow_privilege_escalation: Some(false),
seccomp_profile: Some(SeccompProfile {
type_: "RuntimeDefault".to_string(),
..Default::default()
}),
capabilities: Some(Capabilities {
drop: Some(vec!["ALL".to_string()]),
..Default::default()
}),
..Default::default()
})
} else {
None
};
let init_containers = init_container_image.map(|image| {
vec![Container {
name: "init".to_string(),
image: Some(image),
image_pull_policy: Some(self.config.image_pull_policy.to_string()),
resources: Some(ResourceRequirements {
claims: None,
// Set both limits and requests to the same values, to ensure a
// `Guaranteed` QoS class for the pod.
limits: Some(limits.clone()),
requests: Some(limits.clone()),
}),
security_context: container_security_context.clone(),
env: Some(vec![
EnvVar {
name: "MZ_NAMESPACE".to_string(),
value_from: Some(EnvVarSource {
field_ref: Some(ObjectFieldSelector {
field_path: "metadata.namespace".to_string(),
..Default::default()
}),
..Default::default()
}),
..Default::default()
},
EnvVar {
name: "MZ_POD_NAME".to_string(),
value_from: Some(EnvVarSource {
field_ref: Some(ObjectFieldSelector {
field_path: "metadata.name".to_string(),
..Default::default()
}),
..Default::default()
}),
..Default::default()
},
EnvVar {
name: "MZ_NODE_NAME".to_string(),
value_from: Some(EnvVarSource {
field_ref: Some(ObjectFieldSelector {
field_path: "spec.nodeName".to_string(),
..Default::default()
}),
..Default::default()
}),
..Default::default()
},
]),
..Default::default()
}]
});
let env = if self.config.coverage {
Some(vec![EnvVar {
name: "LLVM_PROFILE_FILE".to_string(),
value: Some(format!("/coverage/{}-%p-%9m%c.profraw", self.namespace)),
..Default::default()
}])
} else {
None
};
let mut volume_mounts = vec![];
if self.config.coverage {
volume_mounts.push(VolumeMount {
name: "coverage".to_string(),
mount_path: "/coverage".to_string(),
..Default::default()
})
}
let volumes = match (disk, &self.config.ephemeral_volume_storage_class) {
(true, Some(ephemeral_volume_storage_class)) => {
volume_mounts.push(VolumeMount {
name: "scratch".to_string(),
mount_path: "/scratch".to_string(),
..Default::default()
});
args.push("--scratch-directory=/scratch".into());
Some(vec![Volume {
name: "scratch".to_string(),
ephemeral: Some(EphemeralVolumeSource {
volume_claim_template: Some(PersistentVolumeClaimTemplate {
spec: PersistentVolumeClaimSpec {
access_modes: Some(vec!["ReadWriteOnce".to_string()]),
storage_class_name: Some(
ephemeral_volume_storage_class.to_string(),
),
resources: Some(VolumeResourceRequirements {
requests: Some(BTreeMap::from([(
"storage".to_string(),
Quantity(
disk_limit
.unwrap_or(DiskLimit::ARBITRARY)
.0
.as_u64()
.to_string(),
),
)])),
..Default::default()
}),
..Default::default()
},
..Default::default()
}),
..Default::default()
}),
..Default::default()
}])
}
(true, None) => {
return Err(anyhow!(
"service requested disk but no ephemeral volume storage class was configured"
));
}
(false, _) => None,
};
let volume_claim_templates = if self.config.coverage {
Some(vec![PersistentVolumeClaim {
metadata: ObjectMeta {
name: Some("coverage".to_string()),
..Default::default()
},
spec: Some(PersistentVolumeClaimSpec {
access_modes: Some(vec!["ReadWriteOnce".to_string()]),
resources: Some(VolumeResourceRequirements {
requests: Some(BTreeMap::from([(
"storage".to_string(),
Quantity("10Gi".to_string()),
)])),
..Default::default()
}),
..Default::default()
}),
..Default::default()
}])
} else {
None
};
let security_context = if let Some(fs_group) = self.config.service_fs_group {
Some(PodSecurityContext {
fs_group: Some(fs_group),
run_as_user: Some(fs_group),
run_as_group: Some(fs_group),
..Default::default()
})
} else {
None
};
let tolerations = Some(vec![
// When the node becomes `NotReady` it indicates there is a problem
// with the node. By default Kubernetes waits 300s (5 minutes)
// before descheduling the pod, but we tune this to 30s for faster
// recovery in the case of node failure.
Toleration {
effect: Some("NoExecute".into()),
key: Some("node.kubernetes.io/not-ready".into()),
operator: Some("Exists".into()),
toleration_seconds: Some(NODE_FAILURE_THRESHOLD_SECONDS),
value: None,
},
Toleration {
effect: Some("NoExecute".into()),
key: Some("node.kubernetes.io/unreachable".into()),
operator: Some("Exists".into()),
toleration_seconds: Some(NODE_FAILURE_THRESHOLD_SECONDS),
value: None,
},
]);
let mut pod_template_spec = PodTemplateSpec {
metadata: Some(ObjectMeta {
labels: Some(labels.clone()),
annotations: Some(pod_annotations), // Do not delete, we insert into it below.
..Default::default()
}),
spec: Some(PodSpec {
init_containers,
containers: vec![Container {
name: container_name,
image: Some(image),
args: Some(args),
image_pull_policy: Some(self.config.image_pull_policy.to_string()),
ports: Some(
ports_in
.iter()
.map(|port| ContainerPort {
container_port: port.port_hint.into(),
name: Some(port.name.clone()),
..Default::default()
})
.collect(),
),
security_context: container_security_context.clone(),
resources: Some(ResourceRequirements {
claims: None,
// Set both limits and requests to the same values, to ensure a
// `Guaranteed` QoS class for the pod.
limits: Some(limits.clone()),
requests: Some(limits),
}),
volume_mounts: if !volume_mounts.is_empty() {
Some(volume_mounts)
} else {
None
},
env,
..Default::default()
}],
volumes,
security_context,
node_selector: Some(node_selector),
scheduler_name: self.config.scheduler_name.clone(),
service_account: self.config.service_account.clone(),
affinity: Some(Affinity {
pod_anti_affinity: anti_affinity,
pod_affinity,
node_affinity,
..Default::default()
}),
topology_spread_constraints: topology_spread,
tolerations,
// Setting a 0s termination grace period has the side effect of
// automatically starting a new pod when the previous pod is
// currently terminating. This enables recovery from a node
// failure with no manual intervention. Without this setting,
// the StatefulSet controller will refuse to start a new pod
// until the failed node is manually removed from the Kubernetes
// cluster.
//
// The Kubernetes documentation strongly advises against this
// setting, as StatefulSets attempt to provide "at most once"
// semantics [0]--that is, the guarantee that for a given pod in
// a StatefulSet there is *at most* one pod with that identity
// running in the cluster.
//
// Materialize services, however, are carefully designed to
// *not* rely on this guarantee. In fact, we do not believe that
// correct distributed systems can meaningfully rely on
// Kubernetes's guarantee--network packets from a pod can be
// arbitrarily delayed, long past that pod's termination.
//
// [0]: https://kubernetes.io/docs/tasks/run-application/force-delete-stateful-set-pod/#statefulset-considerations
termination_grace_period_seconds: Some(0),
..Default::default()
}),
};
let pod_template_json = serde_json::to_string(&pod_template_spec).unwrap();
let mut hasher = Sha256::new();
hasher.update(pod_template_json);
let pod_template_hash = format!("{:x}", hasher.finalize());
pod_template_spec
.metadata
.as_mut()
.unwrap()
.annotations
.as_mut()
.unwrap()
.insert(
POD_TEMPLATE_HASH_ANNOTATION.to_owned(),
pod_template_hash.clone(),
);
let stateful_set = StatefulSet {
metadata: ObjectMeta {
name: Some(name.clone()),
..Default::default()
},
spec: Some(StatefulSetSpec {
selector: LabelSelector {
match_labels: Some(match_labels),
..Default::default()
},
service_name: name.clone(),
replicas: Some(scale.into()),
template: pod_template_spec,
pod_management_policy: Some("Parallel".to_string()),
volume_claim_templates,
..Default::default()
}),
status: None,
};
self.send_command(WorkerCommand::EnsureService {
desc: ServiceDescription {
name,
scale,
service,
stateful_set,
pod_template_hash,
},
});
self.service_infos.lock().expect("poisoned lock").insert(
id.to_string(),
ServiceInfo {
scale,
disk,
disk_limit,
},
);
Ok(Box::new(KubernetesService { hosts, ports }))
}
/// Drops the identified service, if it exists.
fn drop_service(&self, id: &str) -> Result<(), anyhow::Error> {
fail::fail_point!("kubernetes_drop_service", |_| Err(anyhow!("failpoint")));
self.service_infos.lock().expect("poisoned lock").remove(id);
self.send_command(WorkerCommand::DropService {
name: self.service_name(id),
});
Ok(())
}
/// Lists the identifiers of all known services.
async fn list_services(&self) -> Result<Vec<String>, anyhow::Error> {
let (result_tx, result_rx) = oneshot::channel();
self.send_command(WorkerCommand::ListServices {
namespace: self.namespace.clone(),
result_tx,
});
let list = result_rx.await.expect("worker task not dropped");
Ok(list)
}
fn watch_services(&self) -> BoxStream<'static, Result<ServiceEvent, anyhow::Error>> {
fn into_service_event(pod: Pod) -> Result<ServiceEvent, anyhow::Error> {
let process_id = pod.name_any().split('-').last().unwrap().parse()?;
let service_id_label = "environmentd.materialize.cloud/service-id";
let service_id = pod
.labels()
.get(service_id_label)
.ok_or_else(|| anyhow!("missing label: {service_id_label}"))?
.clone();
let oomed = pod
.status
.as_ref()
.and_then(|status| status.container_statuses.as_ref())
.map(|container_statuses| {
container_statuses.iter().any(|cs| {
// The container might have already transitioned from "terminated" to
// "waiting"/"running" state, in which case we need to check its previous
// state to find out why it terminated.
let current_state = cs.state.as_ref().and_then(|s| s.terminated.as_ref());
let last_state = cs.last_state.as_ref().and_then(|s| s.terminated.as_ref());
let termination_state = current_state.or(last_state);
// The interesting exit codes are 135 (SIGBUS) and 137 (SIGKILL). SIGKILL
// occurs when the OOM killer terminates the container, SIGBUS occurs when
// the container runs out of disk. We treat the latter as an OOM condition
// too since we only use disk for spilling memory.
let exit_code = termination_state.map(|s| s.exit_code);
exit_code == Some(135) || exit_code == Some(137)
})
})
.unwrap_or(false);
let (pod_ready, last_probe_time) = pod
.status
.and_then(|status| status.conditions)
.and_then(|conditions| conditions.into_iter().find(|c| c.type_ == "Ready"))
.map(|c| (c.status == "True", c.last_probe_time))
.unwrap_or((false, None));
let status = if pod_ready {
ServiceStatus::Online
} else {
ServiceStatus::Offline(oomed.then_some(OfflineReason::OomKilled))
};
let time = if let Some(time) = last_probe_time {
time.0
} else {
Utc::now()
};
Ok(ServiceEvent {
service_id,
process_id,
status,
time,
})
}
let stream = watcher(self.pod_api.clone(), self.watch_pod_params())
.touched_objects()
.filter_map(|object| async move {
match object {
Ok(pod) => Some(into_service_event(pod)),
Err(error) => {
// We assume that errors returned by Kubernetes are usually transient, so we
// just log a warning and ignore them otherwise.
tracing::warn!("service watch error: {error}");
None
}
}
});
Box::pin(stream)
}
fn update_scheduling_config(&self, config: ServiceSchedulingConfig) {
*self.scheduling_config.write().expect("poisoned") = config;
}
}
impl OrchestratorWorker {
fn spawn(self, name: String) -> AbortOnDropHandle<()> {
mz_ore::task::spawn(|| name, self.run()).abort_on_drop()
}
async fn run(mut self) {
{
info!("initializing Kubernetes orchestrator worker");
let start = Instant::now();
// Fetch the owner reference for our own pod (usually a
// StatefulSet), so that we can propagate it to the services we
// create.
let hostname = env::var("HOSTNAME").unwrap_or_else(|_| panic!("HOSTNAME environment variable missing or invalid; required for Kubernetes orchestrator"));
let orchestrator_pod = Retry::default()
.clamp_backoff(Duration::from_secs(10))
.retry_async(|_| self.pod_api.get(&hostname))
.await
.expect("always retries on error");
self.owner_references
.extend(orchestrator_pod.owner_references().into_iter().cloned());
info!(
"Kubernetes orchestrator worker initialized in {:?}",
start.elapsed()
);
}
while let Some(cmd) = self.command_rx.recv().await {
self.handle_command(cmd).await;
}
}
/// Handle a worker command.
///
/// If handling the command fails, it is automatically retried. All command handlers return
/// [`K8sError`], so we can reasonably assume that a failure is caused by issues communicating
/// with the K8S server and that retrying resolves them eventually.
async fn handle_command(&mut self, cmd: WorkerCommand) {
async fn retry<F, U, R>(f: F, cmd_type: &str) -> R
where
F: Fn() -> U,
U: Future<Output = Result<R, K8sError>>,
{
Retry::default()
.clamp_backoff(Duration::from_secs(10))
.retry_async(|_| {
f().map_err(
|error| tracing::error!(%cmd_type, "orchestrator call failed: {error}"),
)
})
.await
.expect("always retries on error")
}
use WorkerCommand::*;
match cmd {
EnsureService { desc } => {
retry(|| self.ensure_service(desc.clone()), "EnsureService").await
}
DropService { name } => retry(|| self.drop_service(&name), "DropService").await,
ListServices {
namespace,
result_tx,
} => {
let result = retry(|| self.list_services(&namespace), "ListServices").await;
let _ = result_tx.send(result);
}
FetchServiceMetrics {
name,
info,
result_tx,
} => {
let result = self.fetch_service_metrics(&name, &info).await;
let _ = result_tx.send(result);
}
}
}
async fn fetch_service_metrics(
&self,
name: &str,
info: &ServiceInfo,
) -> Vec<ServiceProcessMetrics> {
/// Get metrics for a particular service and process, converting them into a sane (i.e., numeric) format.
///
/// Note that we want to keep going even if a lookup fails for whatever reason,
/// so this function is infallible. If we fail to get cpu or memory for a particular pod,
/// we just log a warning and install `None` in the returned struct.
async fn get_metrics(
self_: &OrchestratorWorker,
service_name: &str,
i: usize,
disk: bool,
disk_limit: Option<DiskLimit>,
) -> ServiceProcessMetrics {
let name = format!("{service_name}-{i}");
let disk_usage_fut = async {
if disk {
Some(
self_
.custom_metrics_api
.get_subresource(
"kubelet_volume_stats_used_bytes",
// The CSI provider we use sets up persistentvolumeclaim's
// with `{pod name}-scratch` as the name. It also provides
// the metrics, and does so under the `persistentvolumeclaims`
// resource type, instead of `pods`.
&format!("{name}-scratch"),
)
.await,
)
} else {
None
}
};
let disk_capacity_fut = async {
if disk {
Some(
self_
.custom_metrics_api
.get_subresource(
"kubelet_volume_stats_capacity_bytes",
&format!("{name}-scratch"),
)
.await,
)
} else {
None
}
};
let (metrics, disk_usage, disk_capacity) = match futures::future::join3(
self_.metrics_api.get(&name),
disk_usage_fut,
disk_capacity_fut,
)
.await
{
(Ok(metrics), disk_usage, disk_capacity) => {
// TODO(guswynn): don't tolerate errors here, when we are more
// comfortable with `prometheus-adapter` in production
// (And we run it in ci).
let disk_usage = match disk_usage {
Some(Ok(disk_usage)) => Some(disk_usage),
Some(Err(e)) if !matches!(&e, K8sError::Api(e) if e.code == 404) => {
warn!(
"Failed to fetch `kubelet_volume_stats_used_bytes` for {name}: {e}"
);
None
}
_ => None,
};
let disk_capacity = match disk_capacity {
Some(Ok(disk_capacity)) => Some(disk_capacity),
Some(Err(e)) if !matches!(&e, K8sError::Api(e) if e.code == 404) => {
warn!("Failed to fetch `kubelet_volume_stats_capacity_bytes` for {name}: {e}");
None
}
_ => None,
};
(metrics, disk_usage, disk_capacity)
}
(Err(e), _, _) => {
warn!("Failed to get metrics for {name}: {e}");
return ServiceProcessMetrics::default();
}
};
let Some(PodMetricsContainer {
usage:
PodMetricsContainerUsage {
cpu: Quantity(cpu_str),
memory: Quantity(mem_str),
},
..
}) = metrics.containers.get(0)
else {
warn!("metrics result contained no containers for {name}");
return ServiceProcessMetrics::default();
};
let cpu = match parse_k8s_quantity(cpu_str) {
Ok(q) => match q.try_to_integer(-9, true) {
Some(i) => Some(i),
None => {
tracing::error!("CPU value {q:? }out of range");
None
}
},
Err(e) => {
tracing::error!("Failed to parse CPU value {cpu_str}: {e}");
None
}
};
let memory = match parse_k8s_quantity(mem_str) {
Ok(q) => match q.try_to_integer(0, false) {
Some(i) => Some(i),
None => {
tracing::error!("Memory value {q:?} out of range");
None
}
},
Err(e) => {
tracing::error!("Failed to parse memory value {mem_str}: {e}");
None
}
};
let disk_usage = match (disk_usage, disk_capacity) {
(Some(disk_usage_metrics), Some(disk_capacity_metrics)) => {
if !disk_usage_metrics.items.is_empty()
&& !disk_capacity_metrics.items.is_empty()
{
let disk_usage_str = &*disk_usage_metrics.items[0].value.0;
let disk_usage = match parse_k8s_quantity(disk_usage_str) {
Ok(q) => match q.try_to_integer(0, true) {
Some(i) => Some(i),
None => {
tracing::error!("Disk usage value {q:?} out of range");
None
}
},
Err(e) => {
tracing::error!(
"Failed to parse disk usage value {disk_usage_str}: {e}"
);
None
}
};
let disk_capacity_str = &*disk_capacity_metrics.items[0].value.0;
let disk_capacity = match parse_k8s_quantity(disk_capacity_str) {
Ok(q) => match q.try_to_integer(0, true) {
Some(i) => Some(i),
None => {
tracing::error!("Disk capacity value {q:?} out of range");
None
}
},
Err(e) => {
tracing::error!(
"Failed to parse disk capacity value {disk_capacity_str}: {e}"
);
None
}
};
// We only populate a `disk_usage` if we have all 5 of of:
// - a disk limit (so it must be an actual managed cluster with a real limit)
// - a reported disk capacity
// - a reported disk usage
//
// There are 3 weird cases we have to handle here:
// - Some instance types report a large disk capacity than the requested
// limit. We clamp those capacities to the limit, which means we can
// report 100% usage even if we have a bit of space left.
// - Other instances report a smaller disk capacity than the limit, due to
// overheads. In this case, we correct the usage by adding the overhead, so
// we report a more accurate usage number.
// - The disk limit can be more up-to-date (from `service_infos`) than the
// reported metric. In that case, we report the minimum of the usage
// and the limit, which means we can report 100% usage temporarily
// if a replica is sized down.
match (disk_usage, disk_capacity, disk_limit) {
(
Some(disk_usage),
Some(disk_capacity),
Some(DiskLimit(disk_limit)),
) => {
let disk_capacity = if disk_limit.0 < disk_capacity {
// We issue a debug message instead
// of a warning or error because all the
// above cases are relatively common.
tracing::debug!(
"disk capacity {} is larger than the disk limit {}; \
disk usage will indicate 100% full before the disk is truly full; \
as long as the delta is small this is not a cause for concern",
disk_capacity,
disk_limit.0
);
// Clamp to the limit.
disk_limit.0
} else {
disk_capacity
};
// If we overflow, just clamp to the disk limit
let disk_usage = (disk_limit.0 - disk_capacity)
.checked_add(disk_usage)
.unwrap_or(disk_limit.0);
// Clamp to the limit. Note that this can be clamped during
// replica resizes of if the disk usage is ABOVE the
// configured limit, as may occur on some instances.
Some(std::cmp::min(disk_usage, disk_limit.0))
}
_ => None,
}
} else {
None
}
}
_ => None,
};
ServiceProcessMetrics {
cpu_nano_cores: cpu,
memory_bytes: memory,
disk_usage_bytes: disk_usage,
}
}
let ret = futures::future::join_all(
(0..info.scale).map(|i| get_metrics(self, name, i.into(), info.disk, info.disk_limit)),
);
ret.await
}
async fn ensure_service(&self, mut desc: ServiceDescription) -> Result<(), K8sError> {
// We inject our own pod's owner references into the Kubernetes objects
// created for the service so that if the
// Deployment/StatefulSet/whatever that owns the pod running the
// orchestrator gets deleted, so do all services spawned by this
// orchestrator.
desc.service
.metadata
.owner_references
.get_or_insert(vec![])
.extend(self.owner_references.iter().cloned());
desc.stateful_set
.metadata
.owner_references
.get_or_insert(vec![])
.extend(self.owner_references.iter().cloned());
self.service_api
.patch(
&desc.name,
&PatchParams::apply(FIELD_MANAGER).force(),
&Patch::Apply(desc.service),
)
.await?;
self.stateful_set_api
.patch(
&desc.name,
&PatchParams::apply(FIELD_MANAGER).force(),
&Patch::Apply(desc.stateful_set),
)
.await?;
// Explicitly delete any pods in the stateful set that don't match the
// template. In theory, Kubernetes would do this automatically, but
// in practice we have observed that it does not.
// See: https://github.com/kubernetes/kubernetes/issues/67250
for pod_id in 0..desc.scale {
let pod_name = format!("{}-{pod_id}", desc.name);
let pod = match self.pod_api.get(&pod_name).await {
Ok(pod) => pod,
// Pod already doesn't exist.
Err(kube::Error::Api(e)) if e.code == 404 => continue,
Err(e) => return Err(e),
};
if pod.annotations().get(POD_TEMPLATE_HASH_ANNOTATION) != Some(&desc.pod_template_hash)
{
match self
.pod_api
.delete(&pod_name, &DeleteParams::default())
.await
{
Ok(_) => (),
// Pod got deleted while we were looking at it.
Err(kube::Error::Api(e)) if e.code == 404 => (),
Err(e) => return Err(e),
}
}
}
Ok(())
}
async fn drop_service(&self, name: &str) -> Result<(), K8sError> {
let res = self
.stateful_set_api
.delete(name, &DeleteParams::default())
.await;
match res {
Ok(_) => (),
Err(K8sError::Api(e)) if e.code == 404 => (),
Err(e) => return Err(e),
}
let res = self
.service_api
.delete(name, &DeleteParams::default())
.await;
match res {
Ok(_) => Ok(()),
Err(K8sError::Api(e)) if e.code == 404 => Ok(()),
Err(e) => Err(e),
}
}
async fn list_services(&self, namespace: &str) -> Result<Vec<String>, K8sError> {
let stateful_sets = self.stateful_set_api.list(&Default::default()).await?;
let name_prefix = format!("{namespace}-");
Ok(stateful_sets
.into_iter()
.filter_map(|ss| {
ss.metadata
.name
.unwrap()
.strip_prefix(&name_prefix)
.map(Into::into)
})
.collect())
}
}
#[derive(Debug, Clone)]
struct KubernetesService {
hosts: Vec<String>,
ports: BTreeMap<String, u16>,
}
impl Service for KubernetesService {
fn addresses(&self, port: &str) -> Vec<String> {
let port = self.ports[port];
self.hosts
.iter()
.map(|host| format!("{host}:{port}"))
.collect()
}
}