feat(shared): add a modular Model Serving client to the workspace client

knip.json

@@ -17,6 +17,7 @@
         "@databricks/sdk-genie",
         "@databricks/sdk-jobs",
         "@databricks/sdk-postgres",
+        "@databricks/sdk-modelserving",
         "@databricks/sdk-uc-tables",
         "@databricks/sdk-vectorsearch",
         "@databricks/sdk-options",

packages/appkit/package.json

@@ -77,6 +77,7 @@
     "@databricks/sdk-files": "0.52.0",
     "@databricks/sdk-genie": "0.54.0",
     "@databricks/sdk-jobs": "0.57.0",
+    "@databricks/sdk-modelserving": "0.53.0",
     "@databricks/sdk-uc-tables": "0.52.0",
     "@databricks/sdk-vectorsearch": "0.51.0",
     "@databricks/sdk-options": "0.51.0",

packages/appkit/src/testing/mock-workspace-client.ts

@@ -73,6 +73,7 @@ const FACADE_SERVICES = [
   "tables",
   "postgres",
   "statementExecution",
+  "modelServing",
   "servingEndpoints",
   "currentUser",
 ] as const;

packages/shared/package.json

@@ -54,6 +54,7 @@
     "@databricks/sdk-files": "0.52.0",
     "@databricks/sdk-genie": "0.54.0",
     "@databricks/sdk-jobs": "0.57.0",
+    "@databricks/sdk-modelserving": "0.53.0",
     "@databricks/sdk-uc-tables": "0.52.0",
     "@databricks/sdk-vectorsearch": "0.51.0",
     "@databricks/sdk-options": "0.51.0",

packages/shared/src/workspace-client/client.ts

@@ -22,6 +22,8 @@ import {
   type GenieClient,
   buildJobsClient,
   type JobsClient,
+  buildModelServingClient,
+  type ModelServingClient,
   buildFilesClient,
   type FilesClient,
   buildTablesClient,
@@ -46,6 +48,7 @@ export class AppKitWorkspaceClient implements WorkspaceClient {
   #currentUser?: ScimClient;
   #genie?: GenieClient;
   #jobs?: JobsClient;
+  #modelServing?: ModelServingClient;
   #files?: FilesClient;
   #vectorSearch?: VectorSearchClient;
   #tables?: TablesClient;
@@ -119,6 +122,15 @@ export class AppKitWorkspaceClient implements WorkspaceClient {
     return this.#statementExecution;
   }
 
+  // Modular serving endpoints client. A new accessor because `servingEndpoints`
+  // is the legacy client's public type and can't change.
+  get modelServing(): ModelServingClient {
+    if (!this.#modelServing) {
+      this.#modelServing = buildModelServingClient(this.#opts);
+    }
+    return this.#modelServing;
+  }
+
   get servingEndpoints() {
     return this.#getLegacy().servingEndpoints;
   }

packages/shared/src/workspace-client/modular.ts

@@ -55,6 +55,7 @@ import { type ProfileOptions, resolve } from "@databricks/sdk-core/profiles";
 import type { FilesClient } from "@databricks/sdk-files/v2";
 import type { GenieClient } from "@databricks/sdk-genie/v1";
 import type { JobsClient } from "@databricks/sdk-jobs/v2";
+import type { ModelServingClient } from "@databricks/sdk-modelserving/v1";
 import type { ClientOptions } from "@databricks/sdk-options/client";
 import type { PostgresClient } from "@databricks/sdk-postgres/v1";
 import { ScimClient } from "@databricks/sdk-scim/v1";
@@ -616,6 +617,19 @@ export function buildVectorSearchClient(
   );
 }
 
+/** Build a modular Model Serving (serving endpoints) client from wrapper options. */
+export function buildModelServingClient(
+  opts: WorkspaceClientOptions,
+): ModelServingClient {
+  return lazyClient(
+    () =>
+      import("@databricks/sdk-modelserving/v1").then(
+        (m) => m.ModelServingClient,
+      ),
+    mapToClientOptions(opts),
+  );
+}
+
 /** Build a modular Unity Catalog Tables client from wrapper options. */
 export function buildTablesClient(opts: WorkspaceClientOptions): TablesClient {
   return lazyClient(
@@ -646,6 +660,7 @@ export type {
 } from "@databricks/sdk-files/v2";
 export type { GenieClient } from "@databricks/sdk-genie/v1";
 export type { JobsClient } from "@databricks/sdk-jobs/v2";
+export type { ModelServingClient } from "@databricks/sdk-modelserving/v1";
 export type { PostgresClient } from "@databricks/sdk-postgres/v1";
 export type { ScimClient } from "@databricks/sdk-scim/v1";
 export type { StatementExecutionClient } from "@databricks/sdk-statementexecution/v1";
@@ -692,6 +707,13 @@ export type {
   SubmitRunRequest,
   SubmitRunResponse,
 } from "@databricks/sdk-jobs/v2";
+export type {
+  GetInferenceEndpointRequest,
+  InferenceEndpoint,
+  InferenceEndpointDetailed,
+  ListInferenceEndpointsRequest,
+  ListInferenceEndpointsResponse,
+} from "@databricks/sdk-modelserving/v1";
 export type {
   EndpointHealth,
   EndpointInfo,

packages/shared/src/workspace-client/tests/modular.test.ts

@@ -55,6 +55,12 @@ vi.mock("@databricks/sdk-genie/v1", () => ({
     };
   }),
 }));
+vi.mock("@databricks/sdk-modelserving/v1", () => ({
+  ModelServingClient: vi.fn().mockImplementation((opts) => {
+    ctorOpts.push(opts);
+    return { opts, getInferenceEndpoint: async () => ({}) };
+  }),
+}));
 vi.mock("@databricks/sdk-postgres/v1", () => ({
   PostgresClient: vi.fn().mockImplementation((opts) => {
     ctorOpts.push(opts);
@@ -125,6 +131,7 @@ vi.mock("node:child_process", () => ({
 import { ApiError } from "../errors";
 import {
   buildGenieClient,
+  buildModelServingClient,
   buildWarehousesClient,
   buildPostgresClient,
   buildWorkspaceAuth,
@@ -352,6 +359,31 @@ describe("modular mapToClientOptions (via buildWarehousesClient)", () => {
     expect(Reflect.get(client, "then")).toBeUndefined();
   });
 
+  test("modelServing normalizes the host and passes SP M2M credentials", async () => {
+    process.env.DATABRICKS_HOST = "envhost.cloud.databricks.com";
+    process.env.DATABRICKS_CLIENT_ID = "sp-client-id";
+    process.env.DATABRICKS_CLIENT_SECRET = "sp-secret";
+    await buildModelServingClient({}).getInferenceEndpoint({} as never);
+    expect(ctorOpts[0].host).toBe("https://envhost.cloud.databricks.com");
+    expect((ctorOpts[0].credentials as { name: () => string }).name()).toBe(
+      "oauth-m2m",
+    );
+  });
+
+  test("modelServing (asUser) uses the OBO token as PAT, not the SP env creds", async () => {
+    process.env.DATABRICKS_CLIENT_ID = "sp-client-id";
+    process.env.DATABRICKS_CLIENT_SECRET = "sp-secret";
+    await buildModelServingClient({
+      token: "user-token",
+      host: "https://x",
+    }).getInferenceEndpoint({} as never);
+    expect(ctorOpts[0].credentials).toMatchObject({
+      kind: "pat",
+      token: "user-token",
+    });
+    expect(m2mOpts).toEqual([]);
+  });
+
   test("postgres normalizes the host and passes SP M2M credentials", async () => {
     process.env.DATABRICKS_HOST = "envhost.cloud.databricks.com";
     process.env.DATABRICKS_CLIENT_ID = "sp-client-id";

packages/shared/src/workspace-client/types.ts

@@ -21,6 +21,7 @@ import type {
   ScimClient,
   GenieClient,
   JobsClient,
+  ModelServingClient,
   FilesClient,
   TablesClient,
   VectorSearchClient,
@@ -71,6 +72,9 @@ export interface WorkspaceClient extends WorkspaceAuth {
   /** Statement Execution (modular SDK). */
   readonly statementExecution: StatementExecutionClient;
 
+  /** Serving endpoints (modular SDK). `servingEndpoints` below stays legacy-typed. */
+  readonly modelServing: ModelServingClient;
+
   /** Serving Endpoints. */
   readonly servingEndpoints: LegacyWorkspaceClient["servingEndpoints"];