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, hasAlchemyLabels, toLabels } from "../Labels.ts"; |
| 14 | import type { Providers } from "../Providers.ts"; |
| 15 | import { |
| 16 | hasAlchemyLabelKeys, |
| 17 | isJobTerminal, |
| 18 | normalizeLocation, |
| 19 | parentOf, |
| 20 | parseResourceName, |
| 21 | rfc1035, |
| 22 | userLabels, |
| 23 | waitForOperation, |
| 24 | } from "./internal.ts"; |
| 25 | import { listLocations } from "./names.ts"; |
| 26 | import type { EncryptionSpec } from "./shared.ts"; |
| 27 | |
| 28 | export type ActiveLearningConfig = aiplatform.GoogleCloudAiplatformV1ActiveLearningConfig; |
| 29 | |
| 30 | export type DataLabelingJobProps = { |
| 31 | /** |
| 32 | * Region. Immutable — changing it replaces the job. |
| 33 | * @default the stack's GCP region (`GCP.Region`, else the profile region, else `us-central1`) |
| 34 | */ |
| 35 | location?: string; |
| 36 | /** |
| 37 | * User-facing display name. Defaults to a generated id. Immutable after |
| 38 | * create — DataLabelingJob has no update API. |
| 39 | */ |
| 40 | displayName?: string; |
| 41 | /** |
| 42 | * Dataset resource names to label. Currently a single dataset. |
| 43 | * Format `projects/{project}/locations/{location}/datasets/{dataset}`. |
| 44 | */ |
| 45 | datasets: string[]; |
| 46 | /** |
| 47 | * Number of labelers per DataItem. |
| 48 | */ |
| 49 | labelerCount: number; |
| 50 | /** |
| 51 | * GCS URI of the instruction PDF shared with labelers. |
| 52 | */ |
| 53 | instructionUri: string; |
| 54 | /** |
| 55 | * GCS URI of the YAML schema describing the job config. Schema files |
| 56 | * live under |
| 57 | * `https://storage.googleapis.com/google-cloud-aiplatform/schema/datalabelingjob/inputs/`. |
| 58 | */ |
| 59 | inputsSchemaUri: string; |
| 60 | /** |
| 61 | * Input config parameters for the job (schema-specific). |
| 62 | */ |
| 63 | inputs: unknown; |
| 64 | /** |
| 65 | * Specialist pool resource names associated with this job. |
| 66 | */ |
| 67 | specialistPools?: string[]; |
| 68 | /** |
| 69 | * Labels applied to annotations created by this job. |
| 70 | */ |
| 71 | annotationLabels?: Record<string, string>; |
| 72 | /** |
| 73 | * Active learning pipeline configuration. |
| 74 | */ |
| 75 | activeLearningConfig?: ActiveLearningConfig; |
| 76 | /** |
| 77 | * User labels. Alchemy ownership labels are merged in automatically. |
| 78 | */ |
| 79 | labels?: Record<string, string>; |
| 80 | /** |
| 81 | * Customer-managed encryption. |
| 82 | */ |
| 83 | encryptionSpec?: EncryptionSpec; |
| 84 | }; |
| 85 | |
| 86 | export type DataLabelingJob = Resource< |
| 87 | "GCP.AIPlatform.DataLabelingJob", |
| 88 | DataLabelingJobProps, |
| 89 | { |
| 90 | /** Full resource name `.../dataLabelingJobs/{job}`. */ |
| 91 | name: string; |
| 92 | /** Job id (last path segment). */ |
| 93 | dataLabelingJobId: string; |
| 94 | /** Project id. */ |
| 95 | project: string; |
| 96 | /** Location id. */ |
| 97 | location: string; |
| 98 | /** Display name. */ |
| 99 | displayName: string | undefined; |
| 100 | /** Dataset resource names. */ |
| 101 | datasets: string[]; |
| 102 | /** Labelers per DataItem. */ |
| 103 | labelerCount: number | undefined; |
| 104 | /** Instruction PDF URI. */ |
| 105 | instructionUri: string | undefined; |
| 106 | /** Inputs schema URI. */ |
| 107 | inputsSchemaUri: string | undefined; |
| 108 | /** Detailed job state. */ |
| 109 | state: string | undefined; |
| 110 | /** Labeling progress in `[0, 100]`. */ |
| 111 | labelingProgress: number | undefined; |
| 112 | /** User labels (Alchemy ownership labels stripped). */ |
| 113 | labels: Record<string, string>; |
| 114 | /** Customer-managed KMS key, if any. */ |
| 115 | kmsKeyName: string | undefined; |
| 116 | /** RFC3339 creation timestamp. */ |
| 117 | createTime: string | undefined; |
| 118 | /** RFC3339 last-update timestamp. */ |
| 119 | updateTime: string | undefined; |
| 120 | /** Failure/cancel status, if any. */ |
| 121 | error: { code?: number; message?: string } | undefined; |
| 122 | }, |
| 123 | never, |
| 124 | Providers |
| 125 | >; |
| 126 | |
| 127 | /** |
| 128 | * A Vertex AI DataLabelingJob — human labeling of Dataset DataItems. |
| 129 | * |
| 130 | * Creating a job starts labeling immediately. There is no update API, so |
| 131 | * reconcile is observe-ensure. Delete cancels a running job, then deletes |
| 132 | * it. Alchemy ownership labels are merged into `labels` so `list` / nuke |
| 133 | * can find the job. |
| 134 | * |
| 135 | * ### Creating a Data Labeling Job |
| 136 | * **Example:** Single-label image classification |
| 137 | * ```typescript |
| 138 | * const job = yield* GCP.AIPlatform.DataLabelingJob("Label", { |
| 139 | * datasets: [dataset.name], |
| 140 | * displayName: "label-images", |
| 141 | * labelerCount: 1, |
| 142 | * instructionUri: "gs://bucket/instructions.pdf", |
| 143 | * inputsSchemaUri: |
| 144 | * "gs://google-cloud-aiplatform/schema/datalabelingjob/inputs/image_classification_1.0.0.yaml", |
| 145 | * inputs: { annotationSpecs: ["cat", "dog"] }, |
| 146 | * }); |
| 147 | * ``` |
| 148 | * |
| 149 | * @resource |
| 150 | * @category AIPlatform |
| 151 | */ |
| 152 | export const DataLabelingJob = Resource<DataLabelingJob>("GCP.AIPlatform.DataLabelingJob"); |
| 153 | |
| 154 | export class DataLabelingJobNotResolved extends Data.TaggedError( |
| 155 | "GCP.AIPlatform.DataLabelingJobNotResolved", |
| 156 | )<{ |
| 157 | resourceName: string; |
| 158 | }> {} |
| 159 | |
| 160 | const toId = (id: string, existing?: string) => |
| 161 | Effect.gen(function* () { |
| 162 | if (existing !== undefined) return existing; |
| 163 | return rfc1035( |
| 164 | yield* createPhysicalName({ |
| 165 | id, |
| 166 | maxLength: 63, |
| 167 | lowercase: true, |
| 168 | }), |
| 169 | ); |
| 170 | }); |
| 171 | |
| 172 | const toAttrs = (job: aiplatform.GoogleCloudAiplatformV1DataLabelingJob, project: string) => { |
| 173 | const name = job.name ?? ""; |
| 174 | const parsed = parseResourceName(name, "dataLabelingJobs"); |
| 175 | return { |
| 176 | name, |
| 177 | dataLabelingJobId: parsed.id, |
| 178 | project: parsed.project || project, |
| 179 | location: parsed.location, |
| 180 | displayName: job.displayName, |
| 181 | datasets: job.datasets ?? [], |
| 182 | labelerCount: job.labelerCount, |
| 183 | instructionUri: job.instructionUri, |
| 184 | inputsSchemaUri: job.inputsSchemaUri, |
| 185 | state: job.state, |
| 186 | labelingProgress: job.labelingProgress, |
| 187 | labels: userLabels(job.labels), |
| 188 | kmsKeyName: job.encryptionSpec?.kmsKeyName, |
| 189 | createTime: job.createTime, |
| 190 | updateTime: job.updateTime, |
| 191 | error: job.error ? { code: job.error.code, message: job.error.message } : undefined, |
| 192 | }; |
| 193 | }; |
| 194 | |
| 195 | const getByName = (name: string) => |
| 196 | aiplatform |
| 197 | .getProjectsLocationsDataLabelingJobs({ name }) |
| 198 | .pipe(Effect.catchTag("NotFound", () => Effect.undefined)); |
| 199 | |
| 200 | const listJobs = (project: string, region: string) => { |
| 201 | const collect = (parent: string) => |
| 202 | aiplatform.listProjectsLocationsDataLabelingJobs.pages({ parent, pageSize: 100 }).pipe( |
| 203 | Stream.flatMap((page) => Stream.fromIterable(page.dataLabelingJobs ?? [])), |
| 204 | Stream.filter((job) => hasAlchemyLabelKeys(job.labels)), |
| 205 | Stream.runCollect, |
| 206 | Effect.map((chunk) => Array.from(chunk)), |
| 207 | ); |
| 208 | const fallback = Effect.forEach(listLocations(region), (location) => |
| 209 | collect(`projects/${project}/locations/${location}`), |
| 210 | ).pipe(Effect.map((pages) => pages.flat())); |
| 211 | // Vertex AI has no `locations/-` wildcard; scan known locations. |
| 212 | return fallback.pipe(Effect.catchTag("NotFound", () => Effect.succeed([]))); |
| 213 | }; |
| 214 | |
| 215 | const findOwned = (id: string, project: string, region: string, hinted?: string) => |
| 216 | Effect.gen(function* () { |
| 217 | if (hinted !== undefined && hinted.length > 0) { |
| 218 | const existing = yield* getByName(hinted); |
| 219 | if (existing !== undefined) return existing; |
| 220 | } |
| 221 | const jobs = yield* listJobs(project, region); |
| 222 | for (const job of jobs) { |
| 223 | if (yield* hasAlchemyLabels(id, tagRecord(job.labels))) return job; |
| 224 | } |
| 225 | return undefined as aiplatform.GoogleCloudAiplatformV1DataLabelingJob | undefined; |
| 226 | }); |
| 227 | |
| 228 | const cancelAndDelete = (name: string) => |
| 229 | Effect.gen(function* () { |
| 230 | const existing = yield* getByName(name); |
| 231 | if (existing === undefined) return; |
| 232 | if (!isJobTerminal(existing.state)) { |
| 233 | yield* aiplatform |
| 234 | .cancelProjectsLocationsDataLabelingJobs({ name }) |
| 235 | .pipe(Effect.catchTag("NotFound", () => Effect.void)); |
| 236 | yield* getByName(name).pipe( |
| 237 | Effect.filterOrFail( |
| 238 | (job) => job === undefined || isJobTerminal(job.state), |
| 239 | () => new DataLabelingJobNotResolved({ resourceName: name }), |
| 240 | ), |
| 241 | Effect.retry({ |
| 242 | while: (error) => error._tag === "GCP.AIPlatform.DataLabelingJobNotResolved", |
| 243 | times: 8, |
| 244 | schedule: Schedule.spaced("4 seconds"), |
| 245 | }), |
| 246 | Effect.catchTag("GCP.AIPlatform.DataLabelingJobNotResolved", () => Effect.void), |
| 247 | ); |
| 248 | } |
| 249 | const operation = yield* aiplatform.deleteProjectsLocationsDataLabelingJobs({ name }).pipe( |
| 250 | Effect.retry({ |
| 251 | while: (error) => error._tag === "Conflict", |
| 252 | times: 8, |
| 253 | schedule: Schedule.spaced("3 seconds"), |
| 254 | }), |
| 255 | Effect.catchTags({ |
| 256 | NotFound: () => Effect.undefined, |
| 257 | Conflict: () => Effect.undefined, |
| 258 | }), |
| 259 | ); |
| 260 | if (operation !== undefined) { |
| 261 | yield* waitForOperation(operation, { notFoundOk: true }); |
| 262 | } |
| 263 | }); |
| 264 | |
| 265 | export const DataLabelingJobProvider = () => |
| 266 | Provider.succeed(DataLabelingJob, { |
| 267 | stables: ["name", "dataLabelingJobId", "project", "location", "createTime"], |
| 268 | |
| 269 | diff: Effect.fn(function* ({ news, olds, output }) { |
| 270 | const env = yield* GcpEnvironment.current; |
| 271 | if (!isResolved(news)) return undefined; |
| 272 | const previousLocation = normalizeLocation(olds?.location ?? output?.location, env.region); |
| 273 | const nextLocation = normalizeLocation( |
| 274 | news.location ?? olds?.location ?? output?.location, |
| 275 | env.region, |
| 276 | ); |
| 277 | if (previousLocation !== nextLocation) { |
| 278 | return { action: "replace" as const, deleteFirst: false }; |
| 279 | } |
| 280 | return undefined; |
| 281 | }), |
| 282 | |
| 283 | read: Effect.fn(function* ({ id, output }) { |
| 284 | const env = yield* GcpEnvironment.current; |
| 285 | const existing = yield* findOwned(id, env.project, env.region, output?.name); |
| 286 | if (existing === undefined) return undefined; |
| 287 | const attrs = toAttrs(existing, env.project); |
| 288 | return (yield* hasAlchemyLabels(id, tagRecord(existing.labels))) ? attrs : Unowned(attrs); |
| 289 | }), |
| 290 | |
| 291 | list: () => |
| 292 | Effect.gen(function* () { |
| 293 | const env = yield* GcpEnvironment.current; |
| 294 | const jobs = yield* listJobs(env.project, env.region); |
| 295 | return jobs.map((job) => toAttrs(job, env.project)); |
| 296 | }), |
| 297 | |
| 298 | reconcile: Effect.fn(function* ({ id, news, output }) { |
| 299 | const env = yield* GcpEnvironment.current; |
| 300 | const location = normalizeLocation(news.location ?? output?.location, env.region); |
| 301 | const displayName = news.displayName ?? (yield* toId(id, output?.dataLabelingJobId)); |
| 302 | const desiredLabels = { |
| 303 | ...toLabels(news.labels), |
| 304 | ...(yield* createInternalLabels(id)), |
| 305 | }; |
| 306 | |
| 307 | let current = yield* findOwned(id, env.project, env.region, output?.name); |
| 308 | |
| 309 | if (current === undefined) { |
| 310 | const created = yield* aiplatform |
| 311 | .createProjectsLocationsDataLabelingJobs({ |
| 312 | parent: parentOf(env.project, location), |
| 313 | body: { |
| 314 | displayName, |
| 315 | datasets: news.datasets, |
| 316 | labelerCount: news.labelerCount, |
| 317 | instructionUri: news.instructionUri, |
| 318 | inputsSchemaUri: news.inputsSchemaUri, |
| 319 | inputs: news.inputs, |
| 320 | specialistPools: news.specialistPools, |
| 321 | annotationLabels: news.annotationLabels, |
| 322 | activeLearningConfig: news.activeLearningConfig, |
| 323 | labels: desiredLabels, |
| 324 | encryptionSpec: news.encryptionSpec, |
| 325 | }, |
| 326 | }) |
| 327 | .pipe(Effect.catchTag("Conflict", () => findOwned(id, env.project, env.region))); |
| 328 | current = created ?? undefined; |
| 329 | } |
| 330 | |
| 331 | if (current === undefined) { |
| 332 | return yield* new DataLabelingJobNotResolved({ |
| 333 | resourceName: output?.name ?? parentOf(env.project, location), |
| 334 | }); |
| 335 | } |
| 336 | |
| 337 | return toAttrs(current, env.project); |
| 338 | }), |
| 339 | |
| 340 | delete: Effect.fn(function* ({ output }) { |
| 341 | yield* cancelAndDelete(output.name); |
| 342 | }), |
| 343 | }); |