1from pydantic import BaseModel, ConfigDict, Field
2from typing import Optional, List, Literal
3
4
[docs]
5class Dimensions(BaseModel):
6 model_config = ConfigDict(extra="forbid", populate_by_name=True)
7
8 name: str = Field(description="CloudWatch Metric dimension name")
9 value: str = Field(description="CloudWatch Metric dimension value")
10
11
[docs]
12class CloudWatchTrigger(BaseModel):
13 model_config = ConfigDict(extra="forbid", populate_by_name=True)
14
15 """CloudWatch metric trigger to use for autoscaling"""
16
17
18 activationTargetValue: Optional[float] = Field(
19 default=0,
20 alias="activation_target_value",
21 description="Activation Value for CloudWatch metric to scale from 0 to 1. Only applicable if minReplicaCount = 0",
22 )
23 dimensions: Optional[List[Dimensions]] = Field(
24 default=None, description="Dimensions for Cloudwatch metrics"
25 )
26 metricCollectionPeriod: Optional[int] = Field(
27 default=300,
28 alias="metric_collection_period",
29 description="Defines the Period for CloudWatch query",
30 )
31 metricCollectionStartTime: Optional[int] = Field(
32 default=300,
33 alias="metric_collection_start_time",
34 description="Defines the StartTime for CloudWatch query",
35 )
36 metricName: Optional[str] = Field(
37 default=None,
38 alias="metric_name",
39 description="Metric name to query for Cloudwatch trigger",
40 )
41 metricStat: Optional[str] = Field(
42 default="Average",
43 alias="metric_stat",
44 description="Statistics metric to be used by Trigger. Used to define Stat for CloudWatch query. Default is Average.",
45 )
46 metricType: Optional[Literal["Value", "Average"]] = Field(
47 default="Average",
48 alias="metric_type",
49 description="The type of metric to be used by HPA. Enum: AverageValue - Uses average value of metric per pod, Value - Uses absolute metric value",
50 )
51 minValue: Optional[float] = Field(
52 default=0,
53 alias="min_value",
54 description="Minimum metric value used in case of empty response from CloudWatch. Default is 0.",
55 )
56 name: Optional[str] = Field(
57 default=None, description="Name for the CloudWatch trigger"
58 )
59 namespace: Optional[str] = Field(
60 default=None, description="AWS CloudWatch namespace for metric"
61 )
62 targetValue: Optional[float] = Field(
63 default=None,
64 alias="target_value",
65 description="TargetValue for CloudWatch metric",
66 )
67 useCachedMetrics: Optional[bool] = Field(
68 default=True,
69 alias="use_cached_metrics",
70 description="Enable caching of metric values during polling interval. Default is true",
71 )
72
73
[docs]
74class PrometheusTrigger(BaseModel):
75 model_config = ConfigDict(extra="forbid", populate_by_name=True)
76
77 """Prometheus metric trigger to use for autoscaling"""
78
79
80 activationTargetValue: Optional[float] = Field(
81 default=0,
82 alias="activation_target_value",
83 description="Activation Value for Prometheus metric to scale from 0 to 1. Only applicable if minReplicaCount = 0",
84 )
85 customHeaders: Optional[str] = Field(
86 default=None,
87 alias="custom_headers",
88 description="Custom headers to include while querying the prometheus endpoint.",
89 )
90 metricType: Optional[Literal["Value", "Average"]] = Field(
91 default="Average",
92 alias="metric_type",
93 description="The type of metric to be used by HPA. Enum: AverageValue - Uses average value of metric per pod, Value - Uses absolute metric value",
94 )
95 name: Optional[str] = Field(
96 default=None, description="Name for the Prometheus trigger"
97 )
98 namespace: Optional[str] = Field(
99 default=None, description="Namespace for namespaced queries"
100 )
101 query: Optional[str] = Field(
102 default=None, description="PromQLQuery for the metric."
103 )
104 serverAddress: Optional[str] = Field(
105 default=None,
106 alias="server_address",
107 description="Server address for AMP workspace",
108 )
109 targetValue: Optional[float] = Field(
110 default=None,
111 alias="target_value",
112 description="Target metric value for scaling",
113 )
114 useCachedMetrics: Optional[bool] = Field(
115 default=True,
116 alias="use_cached_metrics",
117 description="Enable caching of metric values during polling interval. Default is true",
118 )
119
120
[docs]
121class CloudWatchTriggerList(BaseModel):
122 model_config = ConfigDict(extra="forbid", populate_by_name=True)
123
124 activationTargetValue: Optional[float] = Field(
125 default=0,
126 alias="activation_target_value",
127 description="Activation Value for CloudWatch metric to scale from 0 to 1. Only applicable if minReplicaCount = 0",
128 )
129 dimensions: Optional[List[Dimensions]] = Field(
130 default=None, description="Dimensions for Cloudwatch metrics"
131 )
132 metricCollectionPeriod: Optional[int] = Field(
133 default=300,
134 alias="metric_collection_period",
135 description="Defines the Period for CloudWatch query",
136 )
137 metricCollectionStartTime: Optional[int] = Field(
138 default=300,
139 alias="metric_collection_start_time",
140 description="Defines the StartTime for CloudWatch query",
141 )
142 metricName: Optional[str] = Field(
143 default=None,
144 alias="metric_name",
145 description="Metric name to query for Cloudwatch trigger",
146 )
147 metricStat: Optional[str] = Field(
148 default="Average",
149 alias="metric_stat",
150 description="Statistics metric to be used by Trigger. Used to define Stat for CloudWatch query. Default is Average.",
151 )
152 metricType: Optional[Literal["Value", "Average"]] = Field(
153 default="Average",
154 alias="metric_type",
155 description="The type of metric to be used by HPA. Enum: AverageValue - Uses average value of metric per pod, Value - Uses absolute metric value",
156 )
157 minValue: Optional[float] = Field(
158 default=0,
159 alias="min_value",
160 description="Minimum metric value used in case of empty response from CloudWatch. Default is 0.",
161 )
162 name: Optional[str] = Field(
163 default=None, description="Name for the CloudWatch trigger"
164 )
165 namespace: Optional[str] = Field(
166 default=None, description="AWS CloudWatch namespace for metric"
167 )
168 targetValue: Optional[float] = Field(
169 default=None,
170 alias="target_value",
171 description="TargetValue for CloudWatch metric",
172 )
173 useCachedMetrics: Optional[bool] = Field(
174 default=True,
175 alias="use_cached_metrics",
176 description="Enable caching of metric values during polling interval. Default is true",
177 )
178
179
[docs]
180class PrometheusTriggerList(BaseModel):
181 model_config = ConfigDict(extra="forbid", populate_by_name=True)
182
183 activationTargetValue: Optional[float] = Field(
184 default=0,
185 alias="activation_target_value",
186 description="Activation Value for Prometheus metric to scale from 0 to 1. Only applicable if minReplicaCount = 0",
187 )
188 customHeaders: Optional[str] = Field(
189 default=None,
190 alias="custom_headers",
191 description="Custom headers to include while querying the prometheus endpoint.",
192 )
193 metricType: Optional[Literal["Value", "Average"]] = Field(
194 default="Average",
195 alias="metric_type",
196 description="The type of metric to be used by HPA. Enum: AverageValue - Uses average value of metric per pod, Value - Uses absolute metric value",
197 )
198 name: Optional[str] = Field(
199 default=None, description="Name for the Prometheus trigger"
200 )
201 namespace: Optional[str] = Field(
202 default=None, description="Namespace for namespaced queries"
203 )
204 query: Optional[str] = Field(
205 default=None, description="PromQLQuery for the metric."
206 )
207 serverAddress: Optional[str] = Field(
208 default=None,
209 alias="server_address",
210 description="Server address for AMP workspace",
211 )
212 targetValue: Optional[float] = Field(
213 default=None,
214 alias="target_value",
215 description="Target metric value for scaling",
216 )
217 useCachedMetrics: Optional[bool] = Field(
218 default=True,
219 alias="use_cached_metrics",
220 description="Enable caching of metric values during polling interval. Default is true",
221 )
222
223
[docs]
224class AutoScalingSpec(BaseModel):
225 model_config = ConfigDict(extra="forbid", populate_by_name=True)
226
227 cloudWatchTrigger: Optional[CloudWatchTrigger] = Field(
228 default=None,
229 alias="cloud_watch_trigger",
230 description="CloudWatch metric trigger to use for autoscaling",
231 )
232 cloudWatchTriggerList: Optional[List[CloudWatchTriggerList]] = Field(
233 default=None,
234 alias="cloud_watch_trigger_list",
235 description="Multiple CloudWatch metric triggers to use for autoscaling. Takes priority over CloudWatchTrigger if both are provided.",
236 )
237 cooldownPeriod: Optional[int] = Field(
238 default=300,
239 alias="cooldown_period",
240 description="The period to wait after the last trigger reported active before scaling the resource back to 0. Default 300 seconds.",
241 )
242 initialCooldownPeriod: Optional[int] = Field(
243 default=300,
244 alias="initial_cooldown_period",
245 description="The delay before the cooldownPeriod starts after the initial creation of the ScaledObject. Default 300 seconds.",
246 )
247 maxReplicaCount: Optional[int] = Field(
248 default=5,
249 alias="max_replica_count",
250 description="The maximum number of model pods to scale to. Default 5.",
251 )
252 minReplicaCount: Optional[int] = Field(
253 default=1,
254 alias="min_replica_count",
255 description="The minimum number of model pods to scale down to. Default 1.",
256 )
257 pollingInterval: Optional[int] = Field(
258 default=30,
259 alias="polling_interval",
260 description="This is the interval to check each trigger on. Default 30 seconds.",
261 )
262 prometheusTrigger: Optional[PrometheusTrigger] = Field(
263 default=None,
264 alias="prometheus_trigger",
265 description="Prometheus metric trigger to use for autoscaling",
266 )
267 prometheusTriggerList: Optional[List[PrometheusTriggerList]] = Field(
268 default=None,
269 alias="prometheus_trigger_list",
270 description="Multiple Prometheus metric triggers to use for autoscaling. Takes priority over PrometheusTrigger if both are provided.",
271 )
272 scaleDownStabilizationTime: Optional[int] = Field(
273 default=300,
274 alias="scale_down_stabilization_time",
275 description="The time window to stabilize for HPA before scaling down. Default 300 seconds.",
276 )
277 scaleUpStabilizationTime: Optional[int] = Field(
278 default=0,
279 alias="scale_up_stabilization_time",
280 description="The time window to stabilize for HPA before scaling up. Default 0 seconds.",
281 )
282
283
[docs]
284class EnvironmentVariables(BaseModel):
285 model_config = ConfigDict(extra="forbid", populate_by_name=True)
286
287 name: str
288 value: str
289
290
[docs]
291class ModelMetrics(BaseModel):
292 """Configuration for model container metrics scraping"""
293
294 model_config = ConfigDict(extra="forbid", populate_by_name=True)
295
296 path: Optional[str] = Field(
297 default="/metrics", description="Path where the model exposes metrics"
298 )
299 port: Optional[int] = Field(
300 default=8080,
301 description="Port where the model exposes metrics. If not specified, a default port will be used.",
302 )
303
304
[docs]
305class Metrics(BaseModel):
306 model_config = ConfigDict(extra="forbid", populate_by_name=True)
307
308 """Configuration for metrics collection and exposure"""
309
310
311 enabled: Optional[bool] = Field(
312 default=True, description="Enable metrics collection for this model deployment"
313 )
314 metricsScrapeIntervalSeconds: Optional[int] = Field(
315 default=15,
316 alias="metrics_scrape_interval_seconds",
317 description="Scrape interval in seconds for metrics collection from sidecar and model container.",
318 )
319 modelMetrics: Optional[ModelMetrics] = Field(
320 default=None,
321 alias="model_metrics",
322 description="Configuration for model container metrics scraping",
323 )
324
325
[docs]
326class AdditionalConfigs(BaseModel):
327 model_config = ConfigDict(extra="forbid", populate_by_name=True)
328
329 name: str
330 value: str
331
332
[docs]
333class Model(BaseModel):
334 model_config = ConfigDict(extra="forbid", populate_by_name=True)
335
336 acceptEula: bool = Field(
337 default=False,
338 alias="accept_eula",
339 description="For models that require a Model Access Config, specify True or False to indicate whether model terms of use have been accepted.",
340 )
341 additionalConfigs: Optional[List[AdditionalConfigs]] = Field(
342 default=None, alias="additional_configs"
343 )
344 gatedModelDownloadRole: Optional[str] = Field(
345 default=None,
346 alias="gated_model_download_role",
347 description="The Amazon Resource Name (ARN) of an IAM role that will be used to download gated model",
348 )
349 modelHubName: Optional[str] = Field(
350 default="SageMakerPublicHub",
351 alias="model_hub_name",
352 description="The name of the model hub content. Can be an ARN or a simple name.",
353 )
354 modelId: str = Field(
355 alias="model_id",
356 description="The unique identifier of the model within the specified hub (hubContentArn).",
357 )
358 modelVersion: Optional[str] = Field(
359 default=None,
360 alias="model_version",
361 description="The version of the model to deploy, in semantic versioning format (e.g., 1.0.0).",
362 )
363
364
[docs]
365class SageMakerEndpoint(BaseModel):
366 model_config = ConfigDict(extra="forbid", populate_by_name=True)
367
368 name: Optional[str] = Field(
369 default="",
370 description="Name of a SageMaker endpoint to be created for this JumpStartModel. The default value of empty string, when used, will skip endpoint creation.",
371 )
372
373
[docs]
374class Validations(BaseModel):
375 model_config = ConfigDict(extra='forbid')
376
377 acceleratorPartitionValidation: Optional[bool] = Field(
378 default=True,
379 alias="accelerator_partition_validation",
380 description="Enable MIG validation for GPU partitioning. Default is true."
381 )
382
383
[docs]
384class Server(BaseModel):
385 model_config = ConfigDict(extra="forbid", populate_by_name=True)
386
387 executionRole: Optional[str] = Field(
388 default=None,
389 alias="execution_role",
390 description="The Amazon Resource Name (ARN) of an IAM role that will be used to deploy and manage the inference server",
391 )
392 instanceType: str = Field(
393 alias="instance_type",
394 description="The EC2 instance type to use for the inference server. Must be one of the supported types.",
395 )
396
397 acceleratorPartitionType: Optional[str] = Field(
398 default=None,
399 alias="accelerator_partition_type",
400 description="MIG profile to use for GPU partitioning"
401 )
402
403 validations: Optional[Validations] = Field(
404 default=None,
405 description="Validations configuration for the server"
406 )
407
408
[docs]
409class IntelligentRoutingSpec(BaseModel):
410 model_config = ConfigDict(extra="forbid", populate_by_name=True)
411
412 """Configuration for intelligent routing"""
413
414
415 autoScalingSpec: Optional[AutoScalingSpec] = Field(
416 default=None, alias="auto_scaling_spec"
417 )
418 enabled: Optional[bool] = Field(
419 default=False, description="Once set, the enabled field cannot be modified"
420 )
421 routingStrategy: Optional[
422 Literal["prefixaware", "kvaware", "session", "roundrobin"]
423 ] = Field(default="prefixaware", alias="routing_strategy")
424
425
[docs]
426class L2CacheSpec(BaseModel):
427 model_config = ConfigDict(extra="forbid", populate_by_name=True)
428
429 l2CacheBackend: Optional[str] = Field(
430 default=None, alias="l2_cache_backend"
431 )
432 l2CacheLocalUrl: Optional[str] = Field(
433 default=None, alias="l2_cache_local_url"
434 )
435
436
[docs]
437class KvCacheSpec(BaseModel):
438 model_config = ConfigDict(extra="forbid", populate_by_name=True)
439
440 cacheConfigFile: Optional[str] = Field(
441 default=None, alias="cache_config_file"
442 )
443 enableL1Cache: Optional[bool] = Field(
444 default=True, alias="enable_l1_cache"
445 )
446 enableL2Cache: Optional[bool] = Field(
447 default=False, alias="enable_l2_cache"
448 )
449 l2CacheSpec: Optional[L2CacheSpec] = Field(
450 default=None, alias="l2_cache_spec"
451 )
452
453
[docs]
454class LoadBalancer(BaseModel):
455 model_config = ConfigDict(extra="forbid", populate_by_name=True)
456
457 healthCheckPath: Optional[str] = Field(
458 default="/ping", alias="health_check_path"
459 )
460 routingAlgorithm: Optional[Literal["least_outstanding_requests", "round_robin"]] = (
461 Field(default="least_outstanding_requests", alias="routing_algorithm")
462 )
463
464
[docs]
465class CustomCertificateConfig(BaseModel):
466 model_config = ConfigDict(extra="forbid", populate_by_name=True)
467
468 acmArn: str = Field(alias="acm_arn", description="ACM certificate ARN")
469 domainName: str = Field(alias="domain_name")
470
471
[docs]
472class TlsConfig(BaseModel):
473 model_config = ConfigDict(extra="forbid", populate_by_name=True)
474
475 customCertificateConfig: Optional[CustomCertificateConfig] = Field(
476 default=None, alias="custom_certificate_config"
477 )
478 tlsCertificateOutputS3Uri: Optional[str] = Field(
479 default=None, alias="tls_certificate_output_s3_uri"
480 )
481
482
499
500
[docs]
501class CaptureOptions(BaseModel):
502 model_config = ConfigDict(extra="forbid", populate_by_name=True)
503
504 """CaptureOption defines what data to capture (input, output, or both)."""
505
506
507 captureMode: Literal["Input", "Output"] = Field(
508 alias="capture_mode", description="Capture mode: Input or Output"
509 )
510
511
[docs]
512class BufferConfig(BaseModel):
513 model_config = ConfigDict(extra="forbid", populate_by_name=True)
514
515 """Configuration for buffering and flushing captured data"""
516
517
518 batchSize: Optional[int] = Field(
519 default=10,
520 alias="batch_size",
521 description="Number of records to batch before writing to S3",
522 )
523 flushIntervalSeconds: Optional[int] = Field(
524 default=60,
525 alias="flush_interval_seconds",
526 description="Flush interval in seconds",
527 )
528
529
[docs]
530class PayloadConfig(BaseModel):
531 model_config = ConfigDict(extra="forbid", populate_by_name=True)
532
533 """Configuration for payload size limits"""
534
535
536 maxPayloadSizeKB: Optional[int] = Field(
537 default=0,
538 alias="max_payload_size_kb",
539 description="Maximum payload size in KB to capture. 0 means no limit.",
540 )
541
542
[docs]
543class DataCaptureModelPod(BaseModel):
544 model_config = ConfigDict(extra="forbid", populate_by_name=True)
545
546 """Configuration for Model Pod level data capture (Tier 3)"""
547
548
549 bufferConfig: Optional[BufferConfig] = Field(
550 default=None, alias="buffer_config",
551 )
552 captureContentTypeHeader: Optional[CaptureContentTypeHeader] = Field(
553 default=None, alias="capture_content_type_header",
554 )
555 captureOptions: Optional[List[CaptureOptions]] = Field(
556 default=None, alias="capture_options",
557 )
558 enabled: bool = Field(description="Enable or disable model pod data capture")
559 initialSamplingPercentage: Optional[int] = Field(
560 default=None, alias="initial_sampling_percentage",
561 )
562 kmsKeyId: Optional[str] = Field(default=None, alias="kms_key_id")
563 payloadConfig: Optional[PayloadConfig] = Field(
564 default=None, alias="payload_config",
565 )
566
567
[docs]
568class DataCaptureSagemakerEndpoint(BaseModel):
569 model_config = ConfigDict(extra="forbid", populate_by_name=True)
570
571 """Configuration for SageMaker Endpoint level data capture (Tier 1)"""
572
573
574 captureContentTypeHeader: Optional[CaptureContentTypeHeader] = Field(
575 default=None, alias="capture_content_type_header",
576 )
577 captureOptions: Optional[List[CaptureOptions]] = Field(
578 default=None, alias="capture_options",
579 )
580 enabled: bool = Field(description="Enable or disable SageMaker endpoint data capture")
581 initialSamplingPercentage: Optional[int] = Field(
582 default=None, alias="initial_sampling_percentage",
583 )
584 kmsKeyId: Optional[str] = Field(default=None, alias="kms_key_id")
585
586
[docs]
587class DataCaptureLoadBalancer(BaseModel):
588 """Configuration for LoadBalancer level data capture (Tier 2)"""
589
590 model_config = ConfigDict(extra="forbid", populate_by_name=True)
591
592 enabled: bool = Field(description="Enable or disable load balancer access logs")
593
594
[docs]
595class DataCapture(BaseModel):
596 model_config = ConfigDict(extra="forbid", populate_by_name=True)
597
598 """Configuration for data capture across multiple tiers (SageMaker, LoadBalancer, Model Pod)"""
599
600
601 loadBalancer: Optional[DataCaptureLoadBalancer] = Field(
602 default=None, alias="load_balancer",
603 )
604 modelPod: Optional[DataCaptureModelPod] = Field(
605 default=None, alias="model_pod",
606 )
607 s3Uri: Optional[str] = Field(default=None, alias="s3_uri")
608 sagemakerEndpoint: Optional[DataCaptureSagemakerEndpoint] = Field(
609 default=None, alias="sagemaker_endpoint",
610 )
611
612
[docs]
613class DnsConfig(BaseModel):
614 model_config = ConfigDict(extra="forbid", populate_by_name=True)
615
616 """DNS automation configuration for Route53."""
617
618
619 hostedZoneId: str = Field(
620 alias="hosted_zone_id",
621 description="Route53 Hosted Zone ID where the DNS record will be created.",
622 )
623
624
625class _HPJumpStartEndpoint(BaseModel):
626 """Config defines the desired state of JumpStartModel."""
627
628 model_config = ConfigDict(extra="ignore", populate_by_name=True)
629
630 autoScalingSpec: Optional[AutoScalingSpec] = Field(
631 default=None, alias="auto_scaling_spec"
632 )
633 dataCapture: Optional[DataCapture] = Field(
634 default=None,
635 alias="data_capture",
636 description="Configuration for data capture across multiple tiers (SageMaker, LoadBalancer, Model Pod)",
637 )
638 dnsConfig: Optional[DnsConfig] = Field(
639 default=None,
640 alias="dns_config",
641 description="DNS automation configuration for Route53. Requires tlsConfig.customCertificateConfig to be set.",
642 )
643 environmentVariables: Optional[List[EnvironmentVariables]] = Field(
644 default=None,
645 alias="environment_variables",
646 description="Additional environment variables to be passed to the inference server. Limited to 100 key-value pairs.",
647 )
648 maxDeployTimeInSeconds: Optional[int] = Field(
649 default=3600,
650 alias="max_deploy_time_in_seconds",
651 description="Maximum allowed time in seconds for the deployment to complete before timing out. Defaults to 1 hour (3600 seconds)",
652 )
653 intelligentRoutingSpec: Optional[IntelligentRoutingSpec] = Field(
654 default=None,
655 alias="intelligent_routing_spec",
656 description="Configuration for intelligent routing",
657 )
658 kvCacheSpec: Optional[KvCacheSpec] = Field(
659 default=None,
660 alias="kv_cache_spec",
661 description="Configuration for KV Cache specification",
662 )
663 loadBalancer: Optional[LoadBalancer] = Field(
664 default=None,
665 alias="load_balancer",
666 description="Configuration for Application Load Balancer",
667 )
668 metrics: Optional[Metrics] = Field(
669 default=None, description="Configuration for metrics collection and exposure"
670 )
671 model: Model
672 replicas: Optional[int] = Field(
673 default=1,
674 description="The desired number of inference server replicas. Default 1.",
675 )
676 sageMakerEndpoint: Optional[SageMakerEndpoint] = Field(
677 default=None, alias="sage_maker_endpoint"
678 )
679 server: Server
680 tlsConfig: Optional[TlsConfig] = Field(default=None, alias="tls_config")
681
682
[docs]
683class Conditions(BaseModel):
684 model_config = ConfigDict(extra="forbid", populate_by_name=True)
685
686 """DeploymentCondition describes the state of a deployment at a certain point."""
687
688
689 lastTransitionTime: Optional[str] = Field(
690 default=None,
691 alias="last_transition_time",
692 description="Last time the condition transitioned from one status to another.",
693 )
694 lastUpdateTime: Optional[str] = Field(
695 default=None,
696 alias="last_update_time",
697 description="The last time this condition was updated.",
698 )
699 message: Optional[str] = Field(
700 default=None,
701 description="A human readable message indicating details about the transition.",
702 )
703 reason: Optional[str] = Field(
704 default=None, description="The reason for the condition's last transition."
705 )
706 status: str = Field(
707 description="Status of the condition, one of True, False, Unknown."
708 )
709 type: str = Field(description="Type of deployment condition.")
710 observedGeneration: Optional[int] = Field(
711 default=None,
712 alias="observed_generation",
713 description="observedGeneration represents the .metadata.generation that the condition was set based upon. For instance, if .metadata.generation is currently 12, but the .status.conditions[x].observedGeneration is 9, the condition is out of date with respect to the current state of the instance.",
714 )
715
716
[docs]
717class Status(BaseModel):
718 model_config = ConfigDict(extra="forbid", populate_by_name=True)
719
720 """Status of the Deployment Object"""
721
722
723 availableReplicas: Optional[int] = Field(
724 default=None,
725 alias="available_replicas",
726 description="Total number of available pods (ready for at least minReadySeconds) targeted by this deployment.",
727 )
728 collisionCount: Optional[int] = Field(
729 default=None,
730 alias="collision_count",
731 description="Count of hash collisions for the Deployment. The Deployment controller uses this field as a collision avoidance mechanism when it needs to create the name for the newest ReplicaSet.",
732 )
733 conditions: Optional[List[Conditions]] = Field(
734 default=None,
735 description="Represents the latest available observations of a deployment's current state.",
736 )
737 observedGeneration: Optional[int] = Field(
738 default=None,
739 alias="observed_generation",
740 description="The generation observed by the deployment controller.",
741 )
742 readyReplicas: Optional[int] = Field(
743 default=None,
744 alias="ready_replicas",
745 description="readyReplicas is the number of pods targeted by this Deployment with a Ready Condition.",
746 )
747 replicas: Optional[int] = Field(
748 default=None,
749 description="Total number of non-terminated pods targeted by this deployment (their labels match the selector).",
750 )
751 terminatingReplicas: Optional[int] = Field(
752 default=None,
753 alias="terminating_replicas",
754 description="Total number of terminating pods targeted by this deployment.",
755 )
756 unavailableReplicas: Optional[int] = Field(
757 default=None,
758 alias="unavailable_replicas",
759 description="Total number of unavailable pods targeted by this deployment. This is the total number of pods that are still required for the deployment to have 100% available capacity. They may either be pods that are running but not yet available or pods that still have not been created.",
760 )
761 updatedReplicas: Optional[int] = Field(
762 default=None,
763 alias="updated_replicas",
764 description="Total number of non-terminated pods targeted by this deployment that have the desired template spec.",
765 )
766
767
[docs]
768class DeploymentStatus(BaseModel):
769 model_config = ConfigDict(extra="forbid", populate_by_name=True)
770
771 """Details of the native kubernetes deployment that hosts the model"""
772
773
774 deploymentObjectOverallState: Optional[str] = Field(
775 default=None,
776 alias="deployment_object_overall_state",
777 description="Overall State of the Deployment Object",
778 )
779 lastUpdated: str = Field(alias="last_updated", description="Last Update Time")
780 message: Optional[str] = Field(
781 default=None,
782 description="Message populated in the root CRD while updating the status of underlying Deployment",
783 )
784 name: str = Field(description="Name of the Deployment Object")
785 reason: Optional[str] = Field(
786 default=None,
787 description="Reason populated in the root CRD while updating the status of underlying Deployment",
788 )
789 status: Optional[Status] = Field(
790 default=None, description="Status of the Deployment Object"
791 )
792
793
[docs]
794class Sagemaker(BaseModel):
795 model_config = ConfigDict(extra="forbid", populate_by_name=True)
796
797 """Status of the SageMaker endpoint"""
798
799
800 configArn: Optional[str] = Field(
801 default=None,
802 alias="config_arn",
803 description="The Amazon Resource Name (ARN) of the endpoint configuration.",
804 )
805 endpointArn: Optional[str] = Field(
806 default=None,
807 alias="endpoint_arn",
808 description="The Amazon Resource Name (ARN) of the SageMaker endpoint",
809 )
810 modelArn: Optional[str] = Field(
811 default=None,
812 alias="model_arn",
813 description="The ARN of the model created in SageMaker.",
814 )
815 state: str = Field(description="The current state of the SageMaker endpoint")
816
817
[docs]
818class Endpoints(BaseModel):
819 """EndpointStatus contains the status of SageMaker endpoints"""
820
821 model_config = ConfigDict(extra="forbid", populate_by_name=True)
822
823 sagemaker: Optional[Sagemaker] = Field(
824 default=None, description="Status of the SageMaker endpoint"
825 )
826
827
[docs]
828class ModelMetricsStatus(BaseModel):
829 """Status of model container metrics collection"""
830
831 model_config = ConfigDict(extra="forbid", populate_by_name=True)
832
833 path: Optional[str] = Field(
834 default=None, description="The path where metrics are available"
835 )
836 port: Optional[int] = Field(
837 default=None, description="The port on which metrics are exposed"
838 )
839
840
[docs]
841class MetricsStatus(BaseModel):
842 model_config = ConfigDict(extra="forbid", populate_by_name=True)
843
844 """Status of metrics collection"""
845
846
847 enabled: bool = Field(description="Whether metrics collection is enabled")
848 errorMessage: Optional[str] = Field(
849 default=None,
850 alias="error_message",
851 description="Error message if metrics collection is in error state",
852 )
853 metricsScrapeIntervalSeconds: Optional[int] = Field(
854 default=None,
855 alias="metrics_scrape_interval_seconds",
856 description="Scrape interval in seconds for metrics collection from sidecar and model container.",
857 )
858 modelMetrics: Optional[ModelMetricsStatus] = Field(
859 default=None,
860 alias="model_metrics",
861 description="Status of model container metrics collection",
862 )
863 state: Optional[str] = Field(
864 default=None, description="Current state of metrics collection"
865 )
866
867
[docs]
868class TlsCertificate(BaseModel):
869 model_config = ConfigDict(extra="forbid", populate_by_name=True)
870
871 """CertificateStatus represents the status of TLS certificates"""
872
873
874 certificateARN: Optional[str] = Field(
875 default=None,
876 alias="certificate_arn",
877 description="The Amazon Resource Name (ARN) of the ACM certificate",
878 )
879 certificateDomainNames: Optional[List[str]] = Field(
880 default=None,
881 alias="certificate_domain_names",
882 description="The certificate domain names that is attached to the certificate",
883 )
884 certificateHealth: Optional[Literal["Valid", "Expiring", "Expired"]] = Field(
885 default=None,
886 alias="certificate_health",
887 description="Certificate health status",
888 )
889 certificateName: Optional[str] = Field(
890 default=None,
891 alias="certificate_name",
892 description="The certificate name of cert manager",
893 )
894 importedCertificates: Optional[List[str]] = Field(
895 default=None,
896 alias="imported_certificates",
897 description="Used for tracking the imported certificates to ACM",
898 )
899 issuerName: Optional[str] = Field(
900 default=None, alias="issuer_name", description="The issuer name of cert manager"
901 )
902 lastCertExpiryTime: Optional[str] = Field(
903 default=None,
904 alias="last_cert_expiry_time",
905 description="The last certificate expiry time",
906 )
907 tlsCertificateOutputS3Bucket: Optional[str] = Field(
908 default=None,
909 alias="tls_certificate_output_s3_bucket",
910 description="S3 bucket that stores the certificate that needs to be trusted",
911 )
912 tlsCertificateS3Keys: Optional[List[str]] = Field(
913 default=None,
914 alias="tls_certificate_s3_keys",
915 description="The output tls certificate S3 key that points to the .pem file",
916 )
917
918
[docs]
919class DataCaptureModelPodStatus(BaseModel):
920 model_config = ConfigDict(extra="forbid", populate_by_name=True)
921
922 """Health status of the model pod data capture tier"""
923
924
925 lastTransitionTime: Optional[str] = Field(
926 default=None,
927 alias="last_transition_time",
928 description="Time of the last health state transition",
929 )
930 message: Optional[str] = Field(
931 default=None,
932 description="Human-readable message describing the health state",
933 )
934 reason: Optional[str] = Field(
935 default=None,
936 description="Reason for unhealthy status (e.g., OOMKilled, S3UploadFailure, MultipleContainerRestarts)",
937 )
938 status: Literal["Healthy", "Unhealthy"] = Field(
939 description="Current health status"
940 )
941
942
[docs]
943class DataCaptureStatus(BaseModel):
944 model_config = ConfigDict(extra="forbid", populate_by_name=True)
945
946 """Health status of the data capture pipeline"""
947
948
949 modelPod: Optional[DataCaptureModelPodStatus] = Field(
950 default=None,
951 alias="model_pod",
952 description="Health status of the model pod data capture tier",
953 )
954
955
[docs]
956class DnsStatus(BaseModel):
957 model_config = ConfigDict(extra="forbid", populate_by_name=True)
958
959 """Status of the operator-managed Route53 DNS record"""
960
961
962 dnsHealth: Optional[Literal["Active", "Pending", "Error"]] = Field(
963 default=None, alias="dns_health",
964 description="DNS resolution status: Active, Pending, or Error.",
965 )
966 hostedZoneId: Optional[str] = Field(
967 default=None, alias="hosted_zone_id",
968 description="Route53 hosted zone ID.",
969 )
970 lastTransitionTime: Optional[str] = Field(
971 default=None, alias="last_transition_time",
972 description="When the status last transitioned, used for propagation timeout.",
973 )
974 managedByOperator: bool = Field(
975 alias="managed_by_operator",
976 description="Whether the operator manages this DNS record.",
977 )
978 message: Optional[str] = Field(
979 default=None, description="Human-readable status or error message."
980 )
981 previousHostedZoneId: Optional[str] = Field(
982 default=None, alias="previous_hosted_zone_id",
983 )
984 previousRecordName: Optional[str] = Field(
985 default=None, alias="previous_record_name",
986 )
987 recordName: Optional[str] = Field(
988 default=None, alias="record_name", description="Route53 record name."
989 )
990
991
[docs]
992class JumpStartModelStatus(BaseModel):
993 model_config = ConfigDict(extra="forbid", populate_by_name=True)
994
995 """ModelDeploymentStatus defines the observed state of ModelDeployment"""
996
997
998 conditions: Optional[List[Conditions]] = Field(
999 default=None,
1000 description="Detailed conditions representing the state of the deployment",
1001 )
1002 dataCaptureStatus: Optional[DataCaptureStatus] = Field(
1003 default=None,
1004 alias="data_capture_status",
1005 description="Health status of the data capture pipeline",
1006 )
1007 deploymentStatus: Optional[DeploymentStatus] = Field(
1008 default=None,
1009 alias="deployment_status",
1010 description="Details of the native kubernetes deployment that hosts the model",
1011 )
1012 dnsStatus: Optional[DnsStatus] = Field(
1013 default=None,
1014 alias="dns_status",
1015 description="Status of the operator-managed Route53 DNS record",
1016 )
1017 endpoints: Optional[Endpoints] = Field(
1018 default=None,
1019 description="EndpointStatus contains the status of SageMaker endpoints",
1020 )
1021 metricsStatus: Optional[MetricsStatus] = Field(
1022 default=None, alias="metrics_status", description="Status of metrics collection"
1023 )
1024 observedGeneration: Optional[int] = Field(
1025 default=None,
1026 alias="observed_generation",
1027 description="Latest generation reconciled by controller",
1028 )
1029 replicas: Optional[int] = Field(
1030 default=None, description="The observed number of inference server replicas."
1031 )
1032 selector: Optional[str] = Field(
1033 default=None, description="LabelSelector for the deployment."
1034 )
1035 state: Optional[
1036 Literal[
1037 "DeploymentPending",
1038 "DeploymentInProgress",
1039 "DeploymentFailed",
1040 "DeploymentComplete",
1041 "DeletionPending",
1042 "DeletionInProgress",
1043 "DeletionFailed",
1044 "DeletionComplete",
1045 ]
1046 ] = Field(default=None, description="Current phase of the model deployment")
1047 tlsCertificate: Optional[TlsCertificate] = Field(
1048 default=None,
1049 alias="tls_certificate",
1050 description="CertificateStatus represents the status of TLS certificates",
1051 )
