The best way to control Claude Code, Codex and more from any computer or phone
| 1 | import * as aiplatform from "@distilled.cloud/gcp/aiplatform_v1"; |
| 2 | import { Unowned } from "alchemy/AdoptPolicy"; |
| 3 | import { isResolved } from "alchemy/Diff"; |
| 4 | import { createPhysicalName } from "alchemy/PhysicalName"; |
| 5 | import * as Provider from "alchemy/Provider"; |
| 6 | import { Resource } from "alchemy/Resource"; |
| 7 | import { tagRecord } from "alchemy/Tags"; |
| 8 | import * as Data from "effect/Data"; |
| 9 | import * as Effect from "effect/Effect"; |
| 10 | import * as Schedule from "effect/Schedule"; |
| 11 | import * as Stream from "effect/Stream"; |
| 12 | import { GcpEnvironment } from "../Environment.ts"; |
| 13 | import { createInternalLabels, diffLabels, hasAlchemyLabels, toLabels } from "../Labels.ts"; |
| 14 | import type { Providers } from "../Providers.ts"; |
| 15 | import { |
| 16 | DEFAULT_TABULAR_METADATA_SCHEMA_URI, |
| 17 | hasAlchemyLabelKeys, |
| 18 | normalizeLocation, |
| 19 | parentOf, |
| 20 | parseResourceName, |
| 21 | rfc1035, |
| 22 | resourceNameFromOperation, |
| 23 | userLabels, |
| 24 | waitForOperation, |
| 25 | } from "./internal.ts"; |
| 26 | import { listLocations } from "./names.ts"; |
| 27 | import type { EncryptionSpec } from "./shared.ts"; |
| 28 | |
| 29 | export type DatasetProps = { |
| 30 | /** |
| 31 | * Dataset id (the `{dataset}` segment of |
| 32 | * `projects/{project}/locations/{location}/datasets/{dataset}`). If |
| 33 | * omitted, Vertex AI assigns an id. Immutable — changing it replaces |
| 34 | * the dataset. |
| 35 | */ |
| 36 | datasetId?: string; |
| 37 | /** |
| 38 | * Region (`us-central1`, `us-east1`, …). Immutable — changing it |
| 39 | * replaces the dataset. `US-CENTRAL1` is accepted and normalized to |
| 40 | * `us-central1`. |
| 41 | * @default the stack's GCP region (`GCP.Region`, else the profile region, else `us-central1`) |
| 42 | */ |
| 43 | location?: string; |
| 44 | /** |
| 45 | * User-facing display name (max 128 characters). Defaults to the |
| 46 | * dataset id. |
| 47 | */ |
| 48 | displayName?: string; |
| 49 | /** |
| 50 | * Human-readable description. |
| 51 | */ |
| 52 | description?: string; |
| 53 | /** |
| 54 | * User labels. Alchemy ownership labels are merged in automatically. |
| 55 | */ |
| 56 | labels?: Record<string, string>; |
| 57 | /** |
| 58 | * GCS URI of the dataset metadata schema YAML (OpenAPI 3.0.2 Schema |
| 59 | * Object). Immutable — changing it replaces the dataset. |
| 60 | * @default "gs://google-cloud-aiplatform/schema/dataset/metadata/tabular_1.0.0.yaml" |
| 61 | */ |
| 62 | metadataSchemaUri?: string; |
| 63 | /** |
| 64 | * Additional schema-specific metadata. For an empty tabular dataset, |
| 65 | * `{}` is sufficient. |
| 66 | */ |
| 67 | metadata?: unknown; |
| 68 | /** |
| 69 | * Customer-managed encryption. Immutable — changing it replaces the |
| 70 | * dataset. |
| 71 | */ |
| 72 | encryptionSpec?: EncryptionSpec; |
| 73 | /** |
| 74 | * Public base model last used by a prompt dataset. |
| 75 | */ |
| 76 | modelReference?: string; |
| 77 | /** |
| 78 | * Saved queries created with the dataset (max one on create). Required |
| 79 | * for DatasetVersion — Vertex rejects versioning an empty dataset with |
| 80 | * no SavedQuery (`Unsupported dataset version creation with empty set |
| 81 | * of saved queries`). |
| 82 | */ |
| 83 | savedQueries?: Array<{ |
| 84 | /** User-facing saved query name. */ |
| 85 | displayName?: string; |
| 86 | /** |
| 87 | * Problem type (`IMAGE_CLASSIFICATION_SINGLE_LABEL`, |
| 88 | * `TEXT_CLASSIFICATION_SINGLE_LABEL`, …). |
| 89 | */ |
| 90 | problemType?: string; |
| 91 | /** Additional saved-query metadata. */ |
| 92 | metadata?: unknown; |
| 93 | }>; |
| 94 | }; |
| 95 | |
| 96 | export type Dataset = Resource< |
| 97 | "GCP.AIPlatform.Dataset", |
| 98 | DatasetProps, |
| 99 | { |
| 100 | /** Full resource name `projects/{project}/locations/{location}/datasets/{dataset}`. */ |
| 101 | name: string; |
| 102 | /** Dataset id (last path segment). */ |
| 103 | datasetId: string; |
| 104 | /** Project id. */ |
| 105 | project: string; |
| 106 | /** Location id (`us-central1`, …). */ |
| 107 | location: string; |
| 108 | /** User-facing display name. */ |
| 109 | displayName: string | undefined; |
| 110 | /** Description. */ |
| 111 | description: string | undefined; |
| 112 | /** User labels (Alchemy ownership labels stripped). */ |
| 113 | labels: Record<string, string>; |
| 114 | /** Metadata schema URI. */ |
| 115 | metadataSchemaUri: string | undefined; |
| 116 | /** Schema-specific metadata. */ |
| 117 | metadata: unknown; |
| 118 | /** Number of DataItems (non-structured datasets). */ |
| 119 | dataItemCount: string | undefined; |
| 120 | /** MetadataStore artifact created with the dataset. */ |
| 121 | metadataArtifact: string | undefined; |
| 122 | /** Public base model reference, if any. */ |
| 123 | modelReference: string | undefined; |
| 124 | /** Customer-managed KMS key, if any. */ |
| 125 | kmsKeyName: string | undefined; |
| 126 | /** RFC3339 creation timestamp. */ |
| 127 | createTime: string | undefined; |
| 128 | /** RFC3339 last-update timestamp. */ |
| 129 | updateTime: string | undefined; |
| 130 | }, |
| 131 | never, |
| 132 | Providers |
| 133 | >; |
| 134 | |
| 135 | /** |
| 136 | * A Vertex AI Dataset — a collection of DataItems and Annotations. |
| 137 | * |
| 138 | * Location, dataset id, metadata schema, and CMEK are immutable. Display |
| 139 | * name, description, and labels update in place. Alchemy ownership labels |
| 140 | * are merged into `labels` so `list` / `pnpm nuke:gcp` can find the |
| 141 | * dataset. |
| 142 | * |
| 143 | * ### Creating a Dataset |
| 144 | * **Example:** Generated id, empty tabular dataset |
| 145 | * ```typescript |
| 146 | * const dataset = yield* GCP.AIPlatform.Dataset("Samples", { |
| 147 | * displayName: "sample rows", |
| 148 | * labels: { env: "test" }, |
| 149 | * }); |
| 150 | * ``` |
| 151 | * |
| 152 | * **Example:** Explicit id and image metadata schema |
| 153 | * ```typescript |
| 154 | * const dataset = yield* GCP.AIPlatform.Dataset("Images", { |
| 155 | * datasetId: "app-images", |
| 156 | * location: "us-central1", |
| 157 | * metadataSchemaUri: |
| 158 | * "gs://google-cloud-aiplatform/schema/dataset/metadata/image_1.0.0.yaml", |
| 159 | * metadata: {}, |
| 160 | * }); |
| 161 | * ``` |
| 162 | * |
| 163 | * ### Updating a Dataset |
| 164 | * **Example:** Rename and relabel |
| 165 | * ```typescript |
| 166 | * // Same logical id, changed props: the engine updates it in place. |
| 167 | * const dataset = yield* GCP.AIPlatform.Dataset("Samples", { |
| 168 | * displayName: "prod rows", |
| 169 | * labels: { env: "prod" }, |
| 170 | * }); |
| 171 | * ``` |
| 172 | * |
| 173 | * @resource |
| 174 | * @category AIPlatform |
| 175 | */ |
| 176 | export const Dataset = Resource<Dataset>("GCP.AIPlatform.Dataset"); |
| 177 | |
| 178 | export class DatasetNotResolved extends Data.TaggedError("GCP.AIPlatform.DatasetNotResolved")<{ |
| 179 | resourceName: string; |
| 180 | }> {} |
| 181 | |
| 182 | export class DatasetStillExists extends Data.TaggedError("GCP.AIPlatform.DatasetStillExists")<{ |
| 183 | resourceName: string; |
| 184 | }> {} |
| 185 | |
| 186 | const resourceName = (project: string, location: string, datasetId: string) => |
| 187 | `projects/${project}/locations/${location}/datasets/${datasetId}`; |
| 188 | |
| 189 | const toId = (id: string, datasetId: string | undefined, existing?: string) => |
| 190 | Effect.gen(function* () { |
| 191 | if (datasetId !== undefined) return rfc1035(datasetId); |
| 192 | if (existing !== undefined) return existing; |
| 193 | return rfc1035( |
| 194 | yield* createPhysicalName({ |
| 195 | id, |
| 196 | maxLength: 63, |
| 197 | lowercase: true, |
| 198 | }), |
| 199 | ); |
| 200 | }); |
| 201 | |
| 202 | const schemaUriOf = (uri: string | undefined) => |
| 203 | uri && uri.length > 0 ? uri : DEFAULT_TABULAR_METADATA_SCHEMA_URI; |
| 204 | |
| 205 | const toAttrs = (dataset: aiplatform.GoogleCloudAiplatformV1Dataset, project: string) => { |
| 206 | const name = dataset.name ?? ""; |
| 207 | const parsed = parseResourceName(name, "datasets"); |
| 208 | return { |
| 209 | name, |
| 210 | datasetId: parsed.id, |
| 211 | project: parsed.project || project, |
| 212 | location: parsed.location, |
| 213 | displayName: dataset.displayName, |
| 214 | description: dataset.description, |
| 215 | labels: userLabels(dataset.labels), |
| 216 | metadataSchemaUri: dataset.metadataSchemaUri, |
| 217 | metadata: dataset.metadata, |
| 218 | dataItemCount: dataset.dataItemCount, |
| 219 | metadataArtifact: dataset.metadataArtifact, |
| 220 | modelReference: dataset.modelReference, |
| 221 | kmsKeyName: dataset.encryptionSpec?.kmsKeyName, |
| 222 | createTime: dataset.createTime, |
| 223 | updateTime: dataset.updateTime, |
| 224 | }; |
| 225 | }; |
| 226 | |
| 227 | const getByName = (name: string) => |
| 228 | name.length === 0 |
| 229 | ? Effect.undefined |
| 230 | : aiplatform.getDatasets({ name }).pipe(Effect.catchTag("NotFound", () => Effect.undefined)); |
| 231 | |
| 232 | const listDatasets = (project: string, region: string) => |
| 233 | Effect.forEach(listLocations(region), (location) => |
| 234 | listDatasetsAt(parentOf(project, location)), |
| 235 | ).pipe(Effect.map((pages) => pages.flat())); |
| 236 | |
| 237 | const listDatasetsAt = (parent: string) => |
| 238 | aiplatform.listProjectsLocationsDatasets |
| 239 | .pages({ |
| 240 | parent, |
| 241 | pageSize: 100, |
| 242 | }) |
| 243 | .pipe( |
| 244 | Stream.flatMap((page) => Stream.fromIterable(page.datasets ?? [])), |
| 245 | Stream.filter((dataset) => hasAlchemyLabelKeys(dataset.labels)), |
| 246 | Stream.runCollect, |
| 247 | Effect.map((chunk) => Array.from(chunk)), |
| 248 | Effect.catchTag("NotFound", () => Effect.succeed([])), |
| 249 | ); |
| 250 | |
| 251 | const findOwned = (id: string, project: string, region: string, hinted?: string) => |
| 252 | Effect.gen(function* () { |
| 253 | if (hinted !== undefined && hinted.length > 0) { |
| 254 | const existing = yield* getByName(hinted); |
| 255 | if (existing !== undefined) return existing; |
| 256 | } |
| 257 | const owned = yield* listDatasets(project, region); |
| 258 | for (const dataset of owned) { |
| 259 | if (yield* hasAlchemyLabels(id, tagRecord(dataset.labels))) { |
| 260 | return dataset; |
| 261 | } |
| 262 | } |
| 263 | return undefined as aiplatform.GoogleCloudAiplatformV1Dataset | undefined; |
| 264 | }); |
| 265 | |
| 266 | const waitUntilExists = (name: string) => |
| 267 | getByName(name).pipe( |
| 268 | Effect.flatMap((dataset) => |
| 269 | dataset |
| 270 | ? Effect.succeed(dataset) |
| 271 | : Effect.fail(new DatasetNotResolved({ resourceName: name })), |
| 272 | ), |
| 273 | Effect.retry({ |
| 274 | while: (error) => error._tag === "GCP.AIPlatform.DatasetNotResolved", |
| 275 | times: 8, |
| 276 | schedule: Schedule.spaced("1 second"), |
| 277 | }), |
| 278 | ); |
| 279 | |
| 280 | const waitUntilGone = (name: string) => |
| 281 | getByName(name).pipe( |
| 282 | Effect.flatMap((dataset) => |
| 283 | dataset === undefined |
| 284 | ? Effect.void |
| 285 | : Effect.fail(new DatasetStillExists({ resourceName: name })), |
| 286 | ), |
| 287 | Effect.retry({ |
| 288 | while: (error) => error._tag === "GCP.AIPlatform.DatasetStillExists", |
| 289 | times: 10, |
| 290 | schedule: Schedule.spaced("2 seconds"), |
| 291 | }), |
| 292 | ); |
| 293 | |
| 294 | export const DatasetProvider = () => |
| 295 | Provider.succeed(Dataset, { |
| 296 | stables: ["name", "datasetId", "project", "location", "metadataSchemaUri", "createTime"], |
| 297 | |
| 298 | diff: Effect.fn(function* ({ news, olds, output }) { |
| 299 | const env = yield* GcpEnvironment.current; |
| 300 | if (!isResolved(news)) return undefined; |
| 301 | const previousId = olds?.datasetId ?? output?.datasetId; |
| 302 | const nextId = news.datasetId ?? previousId; |
| 303 | const previousLocation = normalizeLocation(olds?.location ?? output?.location, env.region); |
| 304 | const nextLocation = normalizeLocation( |
| 305 | news.location ?? olds?.location ?? output?.location, |
| 306 | env.region, |
| 307 | ); |
| 308 | const previousSchema = schemaUriOf(olds?.metadataSchemaUri ?? output?.metadataSchemaUri); |
| 309 | const nextSchema = schemaUriOf( |
| 310 | news.metadataSchemaUri ?? olds?.metadataSchemaUri ?? output?.metadataSchemaUri, |
| 311 | ); |
| 312 | const previousKey = olds?.encryptionSpec?.kmsKeyName ?? output?.kmsKeyName ?? ""; |
| 313 | const nextKey = news.encryptionSpec?.kmsKeyName ?? previousKey; |
| 314 | const replace = |
| 315 | (previousId !== undefined && nextId !== undefined && nextId !== previousId) || |
| 316 | previousLocation !== nextLocation || |
| 317 | previousSchema !== nextSchema || |
| 318 | previousKey !== nextKey; |
| 319 | if (!replace) return undefined; |
| 320 | return { |
| 321 | action: "replace" as const, |
| 322 | deleteFirst: |
| 323 | previousLocation === nextLocation && previousId !== undefined && nextId === previousId, |
| 324 | }; |
| 325 | }), |
| 326 | |
| 327 | read: Effect.fn(function* ({ id, olds, output }) { |
| 328 | const env = yield* GcpEnvironment.current; |
| 329 | const existing = yield* findOwned(id, env.project, env.region, output?.name); |
| 330 | if (existing === undefined) { |
| 331 | if (output?.name) return undefined; |
| 332 | // Dataset ids are server-assigned numbers; only a caller-supplied |
| 333 | // id can be looked up by name. |
| 334 | const datasetId = olds?.datasetId; |
| 335 | if (datasetId === undefined) return undefined; |
| 336 | const location = normalizeLocation(olds?.location ?? output?.location, env.region); |
| 337 | const named = yield* getByName(resourceName(env.project, location, datasetId)); |
| 338 | if (named === undefined) return undefined; |
| 339 | const attrs = toAttrs(named, env.project); |
| 340 | return (yield* hasAlchemyLabels(id, tagRecord(named.labels))) ? attrs : Unowned(attrs); |
| 341 | } |
| 342 | const attrs = toAttrs(existing, env.project); |
| 343 | return (yield* hasAlchemyLabels(id, tagRecord(existing.labels))) ? attrs : Unowned(attrs); |
| 344 | }), |
| 345 | |
| 346 | list: () => |
| 347 | Effect.gen(function* () { |
| 348 | const env = yield* GcpEnvironment.current; |
| 349 | const datasets = yield* listDatasets(env.project, env.region); |
| 350 | return datasets.map((dataset) => toAttrs(dataset, env.project)); |
| 351 | }), |
| 352 | |
| 353 | reconcile: Effect.fn(function* ({ id, news, output }) { |
| 354 | const env = yield* GcpEnvironment.current; |
| 355 | const location = normalizeLocation(news.location ?? output?.location, env.region); |
| 356 | const datasetId = yield* toId(id, news.datasetId, output?.datasetId); |
| 357 | const desiredLabels = { |
| 358 | ...toLabels(news.labels), |
| 359 | ...(yield* createInternalLabels(id)), |
| 360 | }; |
| 361 | const displayName = news.displayName ?? datasetId; |
| 362 | const metadataSchemaUri = schemaUriOf(news.metadataSchemaUri); |
| 363 | |
| 364 | let current = yield* findOwned(id, env.project, env.region, output?.name); |
| 365 | if (current === undefined && news.datasetId !== undefined) { |
| 366 | current = yield* getByName(resourceName(env.project, location, news.datasetId)); |
| 367 | } |
| 368 | |
| 369 | if (current === undefined) { |
| 370 | const created = yield* aiplatform |
| 371 | .createProjectsLocationsDatasets({ |
| 372 | parent: parentOf(env.project, location), |
| 373 | body: { |
| 374 | displayName, |
| 375 | description: news.description, |
| 376 | labels: desiredLabels, |
| 377 | metadataSchemaUri, |
| 378 | metadata: news.metadata ?? {}, |
| 379 | encryptionSpec: news.encryptionSpec, |
| 380 | modelReference: news.modelReference, |
| 381 | savedQueries: news.savedQueries, |
| 382 | }, |
| 383 | }) |
| 384 | .pipe(Effect.catchTag("Conflict", () => Effect.undefined)); |
| 385 | if (created !== undefined) { |
| 386 | const done = yield* waitForOperation(created); |
| 387 | const createdName = |
| 388 | resourceNameFromOperation(done) ?? |
| 389 | (yield* findOwned(id, env.project, env.region))?.name; |
| 390 | if (createdName !== undefined && createdName.length > 0) { |
| 391 | current = yield* waitUntilExists(createdName); |
| 392 | } |
| 393 | } |
| 394 | if (current === undefined) { |
| 395 | current = yield* findOwned(id, env.project, env.region); |
| 396 | } |
| 397 | } |
| 398 | |
| 399 | if (current === undefined) { |
| 400 | return yield* new DatasetNotResolved({ |
| 401 | resourceName: output?.name ?? resourceName(env.project, location, datasetId), |
| 402 | }); |
| 403 | } |
| 404 | |
| 405 | const name = current.name ?? ""; |
| 406 | const observedLabels = tagRecord(current.labels); |
| 407 | const { upsert, removed } = diffLabels(observedLabels, desiredLabels); |
| 408 | const labelsChanged = upsert.length > 0 || removed.length > 0; |
| 409 | const displayNameChanged = (current.displayName ?? "") !== displayName; |
| 410 | const descriptionChanged = (current.description ?? "") !== (news.description ?? ""); |
| 411 | |
| 412 | if (labelsChanged || displayNameChanged || descriptionChanged) { |
| 413 | current = yield* aiplatform.patchProjectsLocationsDatasets({ |
| 414 | name, |
| 415 | updateMask: [ |
| 416 | labelsChanged ? "labels" : undefined, |
| 417 | displayNameChanged ? "display_name" : undefined, |
| 418 | descriptionChanged ? "description" : undefined, |
| 419 | ] |
| 420 | .filter((field): field is string => field !== undefined) |
| 421 | .join(","), |
| 422 | body: { |
| 423 | name, |
| 424 | labels: desiredLabels, |
| 425 | displayName, |
| 426 | description: news.description, |
| 427 | }, |
| 428 | }); |
| 429 | } |
| 430 | |
| 431 | return toAttrs(current, env.project); |
| 432 | }), |
| 433 | |
| 434 | delete: Effect.fn(function* ({ output }) { |
| 435 | const existing = yield* getByName(output.name); |
| 436 | if (existing === undefined) return; |
| 437 | const operation = yield* aiplatform |
| 438 | .deleteDatasets({ name: output.name }) |
| 439 | .pipe(Effect.catchTag("NotFound", () => Effect.undefined)); |
| 440 | if (operation !== undefined) { |
| 441 | yield* waitForOperation(operation, { notFoundOk: true }); |
| 442 | } |
| 443 | yield* waitUntilGone(output.name); |
| 444 | }), |
| 445 | }); |