Skip to content

Commit

Permalink
[ML] Trained Model: Fix start deployment with ML autoscaling and 0 ac…
Browse files Browse the repository at this point in the history 
…tive nodes (#201256)

## Summary

During my testing, I used the current user with all required privileges
but failed to notice that, after switching to the internal`
kibana_system` user, it lacked the manage_autoscaling privilege required
for the `GET /_autoscaling/policy` API.

As a result, the `isMlAutoscalingEnabled` flag, which we rely on in the
Start Deployment modal, was always set to false. This caused a bug in
scenarios with zero active ML nodes, where falling back to deriving
available processors from ML limits was not possible.


You can check the created deployment, it correctly identifies ML
autoscaling:

<img width="670" alt="image"
src="https://github.com/user-attachments/assets/ff1f835e-2b90-4b73-bea8-a49da8846fbd">


Also fixes restoring vCPU levels from the API deployment params.

### Checklist

Check the PR satisfies following conditions. 

- [x] [Unit or functional
tests](https://www.elastic.co/guide/en/kibana/master/development-tests.html)
were updated or added to match the most common scenarios
  • Loading branch information
@darnautov
darnautov authored Nov 26, 2024
1 parent 97a43f7 commit 9827a07
Show file tree
Hide file tree
Showing 4 changed files with 315 additions and 11 deletions.
291 changes: 291 additions & 0 deletions x-pack/plugins/ml/public/application/model_management/deployment_params_mapper.test.ts
View file
Original file line number Diff line number Diff line change
Expand Up @@ -627,6 +627,297 @@ describe('DeploymentParamsMapper', () => {
},
});
});

describe('mapApiToUiDeploymentParams', () => {
it('should map API params to UI correctly', () => {
// Optimized for search
expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 16,
number_of_allocations: 2,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'medium',
});

// Lower value
expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 16,
number_of_allocations: 1,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'medium',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 8,
number_of_allocations: 2,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'medium',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 2,
number_of_allocations: 1,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'low',
});

// Exact match
expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 16,
number_of_allocations: 8,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'high',
});

// Higher value
expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 16,
number_of_allocations: 12,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'high',
});

// Lower value
expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 16,
number_of_allocations: 5,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'high',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 16,
number_of_allocations: 6,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: false,
vCPUUsage: 'high',
});

// Optimized for ingest
expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 1,
number_of_allocations: 1,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForIngest',
adaptiveResources: false,
vCPUUsage: 'low',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 1,
number_of_allocations: 2,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForIngest',
adaptiveResources: false,
vCPUUsage: 'low',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 1,
number_of_allocations: 6,
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForIngest',
adaptiveResources: false,
vCPUUsage: 'medium',
});
});

it('should map API params to UI correctly with adaptive resources', () => {
expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 8,
adaptive_allocations: {
enabled: true,
min_number_of_allocations: 2,
max_number_of_allocations: 2,
},
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: true,
vCPUUsage: 'medium',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 2,
adaptive_allocations: {
enabled: true,
min_number_of_allocations: 2,
max_number_of_allocations: 2,
},
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: true,
vCPUUsage: 'medium',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 1,
adaptive_allocations: {
enabled: true,
min_number_of_allocations: 1,
max_number_of_allocations: 1,
},
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForIngest',
adaptiveResources: true,
vCPUUsage: 'low',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 2,
adaptive_allocations: {
enabled: true,
min_number_of_allocations: 0,
max_number_of_allocations: 1,
},
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: true,
vCPUUsage: 'low',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 1,
adaptive_allocations: {
enabled: true,
min_number_of_allocations: 0,
max_number_of_allocations: 64,
},
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForIngest',
adaptiveResources: true,
vCPUUsage: 'high',
});

expect(
mapper.mapApiToUiDeploymentParams({
model_id: modelId,
deployment_id: 'test-deployment',
priority: 'normal',
threads_per_allocation: 16,
adaptive_allocations: {
enabled: true,
min_number_of_allocations: 0,
max_number_of_allocations: 12,
},
} as unknown as MlTrainedModelAssignmentTaskParametersAdaptive)
).toEqual({
deploymentId: 'test-deployment',
optimized: 'optimizedForSearch',
adaptiveResources: true,
vCPUUsage: 'high',
});
});
});
});
});
});
22 changes: 15 additions & 7 deletions x-pack/plugins/ml/public/application/model_management/deployment_params_mapper.ts
View file
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ type VCPUBreakpoints = Record<
max: number;
/**
* Static value is used for the number of vCPUs when the adaptive resources are disabled.
* Not allowed in certain environments.
* Not allowed in certain environments, Obs and Security serverless projects.
*/
static?: number;
}
Expand Down Expand Up @@ -89,6 +89,7 @@ export class DeploymentParamsMapper {
) {
/**
* Initial value can be different for serverless and ESS with autoscaling.
* Also not available with 0 ML active nodes.
*/
const maxSingleMlNodeProcessors = this.mlServerLimits.max_single_ml_node_processors;

Expand Down Expand Up @@ -236,18 +237,25 @@ export class DeploymentParamsMapper {
? input.adaptive_allocations!.max_number_of_allocations!
: input.number_of_allocations);

// The deployment can be created via API with a number of allocations that do not exactly match our vCPU ranges.
// In this case, we should find the closest vCPU range that does not exceed the max or static value of the range.
const [vCPUUsage] = Object.entries(this.vCpuBreakpoints)
.reverse()
.find(([key, val]) => vCPUs >= val.min) as [
DeploymentParamsUI['vCPUUsage'],
{ min: number; max: number }
];
.filter(([, range]) => vCPUs <= (adaptiveResources ? range.max : range.static!))
.reduce(
(prev, curr) => {
const prevValue = adaptiveResources ? prev[1].max : prev[1].static!;
const currValue = adaptiveResources ? curr[1].max : curr[1].static!;
return Math.abs(vCPUs - prevValue) <= Math.abs(vCPUs - currValue) ? prev : curr;
},
// in case allocation params exceed the max value of the high range
['high', this.vCpuBreakpoints.high]
);

return {
deploymentId: input.deployment_id,
optimized,
adaptiveResources,
vCPUUsage,
vCPUUsage: vCPUUsage as DeploymentParamsUI['vCPUUsage'],
};
}
}
Expand Down
2 changes: 1 addition & 1 deletion x-pack/plugins/ml/server/lib/node_utils.ts
View file
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ export async function getMlNodeCount(client: IScopedClusterClient): Promise<MlNo
return { count, lazyNodeCount };
}

export async function getLazyMlNodeCount(client: IScopedClusterClient) {
export async function getLazyMlNodeCount(client: IScopedClusterClient): Promise<number> {
const body = await client.asInternalUser.cluster.getSettings(
{
include_defaults: true,
Expand Down
11 changes: 8 additions & 3 deletions x-pack/plugins/ml/server/routes/system.ts
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@ import { mlLog } from '../lib/log';
import { capabilitiesProvider } from '../lib/capabilities';
import { spacesUtilsProvider } from '../lib/spaces_utils';
import type { RouteInitialization, SystemRouteDeps } from '../types';
import { getMlNodeCount } from '../lib/node_utils';
import { getLazyMlNodeCount, getMlNodeCount } from '../lib/node_utils';

/**
* System routes
Expand Down Expand Up @@ -187,10 +187,15 @@ export function systemRoutes(

let isMlAutoscalingEnabled = false;
try {
await client.asInternalUser.autoscaling.getAutoscalingPolicy({ name: 'ml' });
// kibana_system user does not have the manage_autoscaling cluster privilege.
// perform this check as a current user.
await client.asCurrentUser.autoscaling.getAutoscalingPolicy({ name: 'ml' });
isMlAutoscalingEnabled = true;
} catch (e) {
// If doesn't exist, then keep the false
// If ml autoscaling policy doesn't exist or the user does not have privileges to fetch it,
// check the number of lazy ml nodes to determine if autoscaling is enabled.
const lazyMlNodeCount = await getLazyMlNodeCount(client);
isMlAutoscalingEnabled = lazyMlNodeCount > 0;
}

return response.ok({
Expand Down

0 comments on commit 9827a07

Please sign in to comment.