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 * as Provider from "alchemy/Provider"; |
| 5 | import { Resource } from "alchemy/Resource"; |
| 6 | import { tagRecord } from "alchemy/Tags"; |
| 7 | import * as Data from "effect/Data"; |
| 8 | import * as Effect from "effect/Effect"; |
| 9 | import * as Schedule from "effect/Schedule"; |
| 10 | import { GcpEnvironment } from "../Environment.ts"; |
| 11 | import { createInternalLabels, hasAlchemyLabels, toLabels } from "../Labels.ts"; |
| 12 | import type { Providers } from "../Providers.ts"; |
| 13 | import { listLocations } from "./names.ts"; |
| 14 | import { waitForOperation } from "./operations.ts"; |
| 15 | import { |
| 16 | AiPlatformStillExists, |
| 17 | MAX_PIPELINE_JOB_ID_LENGTH, |
| 18 | collectPages, |
| 19 | jsonEqual, |
| 20 | locationParent, |
| 21 | normalizeLocation, |
| 22 | parseResourceName, |
| 23 | toPhysicalId, |
| 24 | userLabels, |
| 25 | type EncryptionSpec, |
| 26 | } from "./shared.ts"; |
| 27 | |
| 28 | const COLLECTION = "pipelineJobs"; |
| 29 | |
| 30 | export type PipelineRuntimeConfig = { |
| 31 | /** Cloud Storage root for pipeline artifacts. */ |
| 32 | gcsOutputDirectory?: string; |
| 33 | /** Runtime parameters (schema 2.1.0+). */ |
| 34 | parameterValues?: Record<string, unknown>; |
| 35 | /** Failure policy (`PIPELINE_FAILURE_POLICY_FAIL_SLOW` / `_FAIL_FAST`). */ |
| 36 | failurePolicy?: string; |
| 37 | /** Input artifacts keyed by name. */ |
| 38 | inputArtifacts?: Record<string, { artifactId?: string }>; |
| 39 | }; |
| 40 | |
| 41 | export type PipelineJobProps = { |
| 42 | /** |
| 43 | * Job id. If omitted, a unique RFC1035 name is generated. Less than |
| 44 | * 128 characters, matching `/[a-z0-9-]+/`. Immutable — changing it |
| 45 | * replaces the job (a new run is started). |
| 46 | */ |
| 47 | pipelineJobId?: string; |
| 48 | /** |
| 49 | * Vertex AI location. Immutable — changing it replaces the job. |
| 50 | * @default the stack's GCP region (`GCP.Region`, else the profile region, else `us-central1`) |
| 51 | */ |
| 52 | location?: string; |
| 53 | /** |
| 54 | * Display name (max 128 UTF-8 characters). Defaults to the job id. |
| 55 | */ |
| 56 | displayName?: string; |
| 57 | /** |
| 58 | * User labels. Alchemy ownership labels are merged in automatically. |
| 59 | */ |
| 60 | labels?: Record<string, string>; |
| 61 | /** |
| 62 | * Compiled Kubeflow pipeline spec as a JSON object. |
| 63 | */ |
| 64 | pipelineSpec?: Record<string, unknown>; |
| 65 | /** |
| 66 | * Template URI (Vertex Template Registry or Gallery). Used when |
| 67 | * `pipelineSpec` is omitted. |
| 68 | */ |
| 69 | templateUri?: string; |
| 70 | /** |
| 71 | * Runtime config (output directory, parameters, failure policy). |
| 72 | */ |
| 73 | runtimeConfig?: PipelineRuntimeConfig; |
| 74 | /** |
| 75 | * Service account the pipeline runs as. |
| 76 | */ |
| 77 | serviceAccount?: string; |
| 78 | /** |
| 79 | * VPC network to peer. |
| 80 | */ |
| 81 | network?: string; |
| 82 | /** |
| 83 | * Reserved IP ranges under the VPC. |
| 84 | */ |
| 85 | reservedIpRanges?: string[]; |
| 86 | /** |
| 87 | * Customer-managed encryption key. |
| 88 | */ |
| 89 | encryptionSpec?: EncryptionSpec; |
| 90 | /** |
| 91 | * Run component-level validations before create. |
| 92 | */ |
| 93 | preflightValidations?: boolean; |
| 94 | }; |
| 95 | |
| 96 | export type PipelineJob = Resource< |
| 97 | "GCP.AIPlatform.PipelineJob", |
| 98 | PipelineJobProps, |
| 99 | { |
| 100 | /** Full resource name. */ |
| 101 | name: string; |
| 102 | /** Job id (last path segment). */ |
| 103 | pipelineJobId: string; |
| 104 | /** Project id. */ |
| 105 | project: string; |
| 106 | /** Location id. */ |
| 107 | location: string; |
| 108 | /** Display name. */ |
| 109 | displayName: string | undefined; |
| 110 | /** User labels (Alchemy ownership labels stripped). */ |
| 111 | labels: Record<string, string>; |
| 112 | /** Pipeline state (`PIPELINE_STATE_RUNNING`, …). */ |
| 113 | state: string | undefined; |
| 114 | /** Template URI. */ |
| 115 | templateUri: string | undefined; |
| 116 | /** Service account. */ |
| 117 | serviceAccount: string | undefined; |
| 118 | /** RFC3339 creation timestamp. */ |
| 119 | createTime: string | undefined; |
| 120 | /** RFC3339 start timestamp. */ |
| 121 | startTime: string | undefined; |
| 122 | /** RFC3339 end timestamp. */ |
| 123 | endTime: string | undefined; |
| 124 | /** RFC3339 last-update timestamp. */ |
| 125 | updateTime: string | undefined; |
| 126 | }, |
| 127 | never, |
| 128 | Providers |
| 129 | >; |
| 130 | |
| 131 | /** |
| 132 | * A Vertex AI PipelineJob — a compiled Kubeflow pipeline that runs |
| 133 | * immediately on create. |
| 134 | * |
| 135 | * There is no update API. Changing identity (`pipelineJobId`, `location`) |
| 136 | * or the pipeline spec / template URI replaces the job. |
| 137 | * |
| 138 | * ### Creating a Pipeline Job |
| 139 | * **Example:** Run a compiled spec |
| 140 | * ```typescript |
| 141 | * const job = yield* GCP.AIPlatform.PipelineJob("Train", { |
| 142 | * pipelineSpec: compiled, |
| 143 | * runtimeConfig: { gcsOutputDirectory: "gs://bucket/pipeline-out" }, |
| 144 | * labels: { env: "prod" }, |
| 145 | * }); |
| 146 | * ``` |
| 147 | * |
| 148 | * @resource |
| 149 | * @category AIPlatform |
| 150 | */ |
| 151 | export const PipelineJob = Resource<PipelineJob>("GCP.AIPlatform.PipelineJob"); |
| 152 | |
| 153 | export class PipelineJobNotResolved extends Data.TaggedError( |
| 154 | "GCP.AIPlatform.PipelineJobNotResolved", |
| 155 | )<{ |
| 156 | resourceName: string; |
| 157 | }> {} |
| 158 | |
| 159 | const resourceName = (project: string, location: string, id: string) => |
| 160 | `${locationParent(project, location)}/${COLLECTION}/${id}`; |
| 161 | |
| 162 | const toAttrs = (job: aiplatform.GoogleCloudAiplatformV1PipelineJob, project: string) => { |
| 163 | const name = job.name ?? ""; |
| 164 | const parsed = parseResourceName(name, COLLECTION); |
| 165 | return { |
| 166 | name, |
| 167 | pipelineJobId: parsed.id, |
| 168 | project: parsed.project || project, |
| 169 | location: parsed.location, |
| 170 | displayName: job.displayName, |
| 171 | labels: userLabels(job.labels), |
| 172 | state: job.state, |
| 173 | templateUri: job.templateUri, |
| 174 | serviceAccount: job.serviceAccount, |
| 175 | createTime: job.createTime, |
| 176 | startTime: job.startTime, |
| 177 | endTime: job.endTime, |
| 178 | updateTime: job.updateTime, |
| 179 | }; |
| 180 | }; |
| 181 | |
| 182 | const getByName = (name: string) => |
| 183 | aiplatform |
| 184 | .getProjectsLocationsPipelineJobs({ name }) |
| 185 | .pipe(Effect.catchTag("NotFound", () => Effect.undefined)); |
| 186 | |
| 187 | const TERMINAL_STATES = new Set([ |
| 188 | "PIPELINE_STATE_SUCCEEDED", |
| 189 | "PIPELINE_STATE_FAILED", |
| 190 | "PIPELINE_STATE_CANCELLED", |
| 191 | ]); |
| 192 | |
| 193 | class PipelineJobNotTerminal extends Data.TaggedError("GCP.AIPlatform.PipelineJobNotTerminal")<{ |
| 194 | resourceName: string; |
| 195 | state: string | undefined; |
| 196 | }> {} |
| 197 | |
| 198 | /** |
| 199 | * A pending or running job cannot be deleted: cancel it and wait until it |
| 200 | * settles (cancellation takes up to a few minutes). |
| 201 | */ |
| 202 | const cancelUntilTerminal = (name: string) => |
| 203 | Effect.gen(function* () { |
| 204 | const job = yield* getByName(name); |
| 205 | if (job === undefined || TERMINAL_STATES.has(job.state ?? "")) return; |
| 206 | yield* aiplatform.cancelProjectsLocationsPipelineJobs({ name, body: {} }).pipe( |
| 207 | Effect.catchTags({ |
| 208 | NotFound: () => Effect.void, |
| 209 | // The job can reach a terminal state between the read above and the |
| 210 | // cancel ("… is in state JOB_STATE_FAILED and cannot be canceled"); |
| 211 | // the wait below re-reads the state and still fails if it is not |
| 212 | // terminal. |
| 213 | BadRequest: () => Effect.void, |
| 214 | }), |
| 215 | ); |
| 216 | yield* getByName(name).pipe( |
| 217 | Effect.flatMap((current) => |
| 218 | current === undefined || TERMINAL_STATES.has(current.state ?? "") |
| 219 | ? Effect.void |
| 220 | : Effect.fail(new PipelineJobNotTerminal({ resourceName: name, state: current.state })), |
| 221 | ), |
| 222 | Effect.retry({ |
| 223 | while: (error) => error._tag === "GCP.AIPlatform.PipelineJobNotTerminal", |
| 224 | times: 36, |
| 225 | schedule: Schedule.spaced("5 seconds"), |
| 226 | }), |
| 227 | ); |
| 228 | }); |
| 229 | |
| 230 | const waitUntilGone = (name: string) => |
| 231 | getByName(name).pipe( |
| 232 | Effect.filterOrFail( |
| 233 | (job) => job === undefined, |
| 234 | () => new AiPlatformStillExists({ resourceName: name }), |
| 235 | ), |
| 236 | Effect.asVoid, |
| 237 | Effect.retry({ |
| 238 | while: (error) => error._tag === "GCP.AIPlatform.StillExists", |
| 239 | times: 10, |
| 240 | schedule: Schedule.spaced("2 seconds"), |
| 241 | }), |
| 242 | ); |
| 243 | |
| 244 | export const PipelineJobProvider = () => |
| 245 | Provider.succeed(PipelineJob, { |
| 246 | stables: ["name", "pipelineJobId", "project", "location", "createTime"], |
| 247 | |
| 248 | diff: Effect.fn(function* ({ news, olds, output }) { |
| 249 | const env = yield* GcpEnvironment.current; |
| 250 | if (!isResolved(news)) return undefined; |
| 251 | const previousId = olds?.pipelineJobId ?? output?.pipelineJobId; |
| 252 | const nextId = news.pipelineJobId ?? previousId; |
| 253 | const previousLocation = normalizeLocation(olds?.location ?? output?.location, env.region); |
| 254 | const nextLocation = normalizeLocation(news.location ?? output?.location, env.region); |
| 255 | const specChanged = |
| 256 | olds !== undefined && |
| 257 | (!jsonEqual(news.pipelineSpec, olds.pipelineSpec) || |
| 258 | (news.templateUri ?? "") !== (olds.templateUri ?? "")); |
| 259 | const replace = |
| 260 | (previousId !== undefined && nextId !== undefined && nextId !== previousId) || |
| 261 | previousLocation !== nextLocation || |
| 262 | specChanged; |
| 263 | if (!replace) return undefined; |
| 264 | return { |
| 265 | action: "replace" as const, |
| 266 | deleteFirst: |
| 267 | previousLocation === nextLocation && previousId !== undefined && nextId === previousId, |
| 268 | }; |
| 269 | }), |
| 270 | |
| 271 | read: Effect.fn(function* ({ id, olds, output }) { |
| 272 | const env = yield* GcpEnvironment.current; |
| 273 | const jobId = yield* toPhysicalId( |
| 274 | id, |
| 275 | olds?.pipelineJobId, |
| 276 | output?.pipelineJobId, |
| 277 | MAX_PIPELINE_JOB_ID_LENGTH, |
| 278 | ); |
| 279 | const location = normalizeLocation(olds?.location ?? output?.location, env.region); |
| 280 | const name = output?.name ?? resourceName(env.project, location, jobId); |
| 281 | const existing = yield* getByName(name); |
| 282 | if (existing === undefined) return undefined; |
| 283 | const attrs = toAttrs(existing, env.project); |
| 284 | return (yield* hasAlchemyLabels(id, tagRecord(existing.labels))) ? attrs : Unowned(attrs); |
| 285 | }), |
| 286 | |
| 287 | list: () => |
| 288 | Effect.gen(function* () { |
| 289 | const env = yield* GcpEnvironment.current; |
| 290 | const pages = (yield* Effect.forEach(listLocations(env.region), (location) => |
| 291 | collectPages( |
| 292 | aiplatform.listProjectsLocationsPipelineJobs.pages({ |
| 293 | parent: locationParent(env.project, location), |
| 294 | pageSize: 100, |
| 295 | }), |
| 296 | ).pipe(Effect.catchTag("NotFound", () => Effect.succeed([]))), |
| 297 | )).flat(); |
| 298 | return pages.flatMap((page) => |
| 299 | (page.pipelineJobs ?? []) |
| 300 | .filter((job) => |
| 301 | Object.keys(job.labels ?? {}).some((key) => key.startsWith("alchemy-")), |
| 302 | ) |
| 303 | .map((job) => toAttrs(job, env.project)), |
| 304 | ); |
| 305 | }), |
| 306 | |
| 307 | reconcile: Effect.fn(function* ({ id, news, output }) { |
| 308 | const env = yield* GcpEnvironment.current; |
| 309 | const jobId = yield* toPhysicalId( |
| 310 | id, |
| 311 | news.pipelineJobId, |
| 312 | output?.pipelineJobId, |
| 313 | MAX_PIPELINE_JOB_ID_LENGTH, |
| 314 | ); |
| 315 | const location = normalizeLocation(news.location ?? output?.location, env.region); |
| 316 | const name = resourceName(env.project, location, jobId); |
| 317 | const desiredLabels = { |
| 318 | ...toLabels(news.labels), |
| 319 | ...(yield* createInternalLabels(id)), |
| 320 | }; |
| 321 | |
| 322 | let current = yield* getByName(output?.name ?? name); |
| 323 | |
| 324 | if (current === undefined) { |
| 325 | const created = yield* aiplatform |
| 326 | .createProjectsLocationsPipelineJobs({ |
| 327 | parent: locationParent(env.project, location), |
| 328 | pipelineJobId: jobId, |
| 329 | body: { |
| 330 | displayName: news.displayName ?? jobId, |
| 331 | labels: desiredLabels, |
| 332 | pipelineSpec: news.pipelineSpec, |
| 333 | templateUri: news.templateUri, |
| 334 | runtimeConfig: news.runtimeConfig, |
| 335 | serviceAccount: news.serviceAccount, |
| 336 | network: news.network, |
| 337 | reservedIpRanges: news.reservedIpRanges, |
| 338 | encryptionSpec: news.encryptionSpec, |
| 339 | preflightValidations: news.preflightValidations, |
| 340 | }, |
| 341 | }) |
| 342 | .pipe(Effect.catchTag("Conflict", () => getByName(name))); |
| 343 | current = created ?? undefined; |
| 344 | } |
| 345 | |
| 346 | if (current === undefined) { |
| 347 | return yield* new PipelineJobNotResolved({ resourceName: name }); |
| 348 | } |
| 349 | return toAttrs(current, env.project); |
| 350 | }), |
| 351 | |
| 352 | delete: Effect.fn(function* ({ output }) { |
| 353 | yield* cancelUntilTerminal(output.name); |
| 354 | const operation = yield* aiplatform |
| 355 | .deleteProjectsLocationsPipelineJobs({ name: output.name }) |
| 356 | .pipe(Effect.catchTag("NotFound", () => Effect.undefined)); |
| 357 | if (operation !== undefined) { |
| 358 | yield* waitForOperation(operation, { notFoundOk: true }); |
| 359 | } |
| 360 | yield* waitUntilGone(output.name); |
| 361 | }), |
| 362 | }); |