The best way to control Claude Code, Codex and more from any computer or phone

main
md 213 lines 7.93 KB
Rendered Raw
1 ---
2 title: 2.0.0-beta.65 - SageMaker HyperPod
3 date: 2026-07-27T06:00:00Z
4 category: release
5 excerpt: v2.0.0-beta.65 provisions SageMaker HyperPod clusters — Slurm or EKS orchestrated — with task governance and typed workloads. Plus Worker bundling fixes for Hono-style routes and Drizzle schemas, one-pass destroys for VPC-attached Lambdas, and a smarter nuke.
6 ---
7
8 beta.65 brings **SageMaker HyperPod** to the AWS provider:
9 persistent ML training fleets, Slurm or EKS orchestrated, with
10 task governance and typed workloads. It also fixes two Worker
11 bundling bugs that broke deployed apps, and makes `destroy` and
12 `nuke` converge in one pass.
13
14 :::caution
15 **Breaking change**
16
17 The `alchemy/SQL` barrel is removed
18 ([#951](https://github.com/alchemy-run/alchemy/pull/951)). It
19 eagerly loaded every backend, so projects without the optional
20 `@effect/sql-pg` peer crashed on import. Import each client from
21 its own subpath instead:
22
23 ```diff
24 -import * as SQL from "alchemy/SQL";
25 +import * as SQL from "alchemy/SQL/D1";
26 +import * as SQL from "alchemy/SQL/Postgres";
27 ```
28
29 :::
30
31 ## SageMaker HyperPod
32
33 HyperPod is AWS's persistent fleet for ML training and inference:
34 accelerated instances with automatic health checks and
35 faulty-node replacement, staying up between jobs. Where ECS and
36 EKS run application containers, HyperPod runs distributed
37 training ([#924](https://github.com/alchemy-run/alchemy/pull/924)).
38
39 [`AWS.SageMaker.Cluster`](/providers/aws/sagemaker/cluster)
40 provisions the fleet. The default orchestrator is Slurm — an S3
41 lifecycle script, an execution role, and instance groups:
42
43 ```typescript
44 const cluster = yield* AWS.SageMaker.Cluster("TrainingCluster", {
45 instanceGroups: {
46 controller: {
47 InstanceType: "ml.t3.medium",
48 InstanceCount: 1,
49 ExecutionRole: role.roleArn,
50 LifeCycleConfig: {
51 SourceS3Uri: script.sourceS3Uri,
52 OnCreate: script.onCreate,
53 },
54 },
55 },
56 });
57 ```
58
59 Instance groups reconcile in place: change a count or add a group
60 and the cluster updates; remove one and it is deleted from the
61 cluster. Jobs are submitted with `sbatch` over SSM — every node
62 is an SSM target.
63
64 Pass `orchestrator: { Eks }` and the fleet's nodes join an EKS
65 cluster you provide. From there, workloads are ordinary
66 Kubernetes objects — and
67 [`AWS.EKS.Job`](/providers/kubernetes/job) and
68 [`Deployment`](/providers/kubernetes/deployment) opt onto HyperPod
69 nodes with the `hyperpod` prop. The instance-group keys carry
70 through the cluster's attributes as types, so a typo'd group name
71 is a compile error:
72
73 ```typescript
74 const train = yield* AWS.EKS.Job(
75 "Train",
76 {
77 cluster: eks,
78 main: import.meta.url,
79 hyperpod: { instanceGroup: hyperpod.instanceGroups.workers },
80 },
81 Effect.gen(function* () {
82 return {
83 run: Effect.gen(function* () {
84 // training logic; bindings land IAM on pod identity
85 }),
86 };
87 }),
88 );
89 ```
90
91 Task governance arbitrates the fleet between teams:
92 [`ClusterSchedulerConfig`](/providers/aws/sagemaker/clusterschedulerconfig)
93 sets priority classes and
94 [`ComputeQuota`](/providers/aws/sagemaker/computequota) reserves
95 per-team capacity (Kueue under the hood). A workload submits
96 through governance by passing the quota resource:
97
98 ```typescript
99 hyperpod: {
100 instanceGroup: hyperpod.instanceGroups.workers,
101 quota, // → hyperpod-ns-research + queue label
102 priorityClass: "training",
103 },
104 ```
105
106 Docs: [HyperPod](/aws/compute/hyperpod). The full working stack
107 (both orchestrators, governance, every workload tier) is
108 [`examples/aws-hyperpod`](https://github.com/alchemy-run/alchemy/tree/main/examples/aws-hyperpod).
109
110 ## Worker bundling fixes
111
112 Two bundling bugs could break a deployed Worker even though the
113 deploy succeeded. Both are fixed.
114
115 Hono/Express-style route registrations were stripped from the
116 bundle ([#952](https://github.com/alchemy-run/alchemy/pull/952)).
117 `purePlugin` annotated top-level calls like `app.get("/",
118 handler)` as pure, and under minification that annotation is a
119 deletion instruction — so every request 404'd. Statement-position
120 calls are now only annotated in packages that explicitly declare
121 `sideEffects: false`, and never in entry modules.
122
123 Cross-chunk evaluation order is now correct
124 ([#940](https://github.com/alchemy-run/alchemy/pull/940)).
125 Rolldown could split top-level initializers (e.g. Drizzle
126 `pgTable` schemas) into a chunk that ran before its dependencies,
127 failing Cloudflare startup validation with `Cannot access 'a'
128 before initialization`. `strictExecutionOrder` is now on by
129 default, preserving ESM semantics across chunk boundaries.
130
131 The same PR opens up per-Worker control of chunking: `build`
132 forwards rolldown output options and `preserveEntrySignatures`:
133
134 ```typescript
135 Cloudflare.Worker("Api", {
136 main,
137 build: {
138 output: {
139 codeSplitting: {
140 groups: [{ name: "auth", test: "packages/(schema|auth)/" }],
141 },
142 },
143 },
144 });
145 ```
146
147 ## Destroy and nuke converge in one pass
148
149 Destroying a stack with a VPC-attached Lambda used to strand
150 subnets and security groups on `DependencyViolation`, forcing
151 repeated `destroy` runs
152 ([#937](https://github.com/alchemy-run/alchemy/pull/937)). The
153 cause: Lambda's Hyperplane ENIs linger in the VPC for many
154 minutes after the function is deleted, and AWS's own reaper can
155 take 20+ minutes to clear them. Subnet and security-group deletes
156 now identify detached Lambda ENIs and delete them directly, so
157 one `destroy` removes everything.
158
159 `alchemy unsafe nuke` gets the same treatment
160 ([#950](https://github.com/alchemy-run/alchemy/pull/950)):
161
162 - `--independent` retries each delete on its own capped backoff
163 instead of a shared barrier, so one hanging delete only ever
164 stalls itself.
165 - Providers declare which resource types their cloud-side
166 teardown consumes (a HyperPod cluster deletes its ENIs by
167 assuming its IAM role inside its VPC), and nuke deletes in
168 topologically-ordered waves so a role is never deleted out from
169 under an in-flight cluster teardown.
170
171 ## Also in this release
172
173 - **`_redirects` and `_headers` now reach Cloudflare**
174 ([#928](https://github.com/alchemy-run/alchemy/pull/928)) —
175 the files were read and hashed but never sent, so redirect and
176 header rules were silently dropped. They now apply in deploys,
177 no-op redeploys, and `alchemy dev`.
178 - **Workflows under `alchemy dev`**: pausing a workflow no longer
179 reports `errored`
180 ([#934](https://github.com/alchemy-run/alchemy/pull/934)), and
181 workflow classes export correctly in vite dev
182 ([#933](https://github.com/alchemy-run/alchemy/pull/933)).
183 - **`Docker.Container` supports `labels` and `stopTimeout`**
184 ([#931](https://github.com/alchemy-run/alchemy/pull/931)) —
185 enabling Traefik-managed routing in a remote Docker context.
186 Thanks [Gerbuuun](https://github.com/Gerbuuun)!
187 - **`alchemy/Test/Vitest` is restored**
188 ([#947](https://github.com/alchemy-run/alchemy/pull/947)) —
189 the internal move to alchemy-test accidentally removed the
190 public vitest harness that shipped through beta.63.
191 - **Branded primitives stay opaque in Outputs**
192 ([#941](https://github.com/alchemy-run/alchemy/pull/941)) — a
193 `string & Brand<"UserId">` Action output is now `Output<UserId>`
194 instead of exploding into an `ObjectExpr` over every string
195 method.
196 - **SQL fragment helpers compose on wrapped clients**
197 ([#951](https://github.com/alchemy-run/alchemy/pull/951)) —
198 `` sql`INSERT INTO users ${sql.insert(row)}` `` and nested
199 `sql.and`/`sql.or` now replay against the resolved client, with
200 new live test suites covering the full `SqlClient` surface on
201 D1 and Postgres.
202 - **The alchemy-test plain reporter** prints a `[done/total]`
203 progress counter per test and a failed-test list at the end of
204 the run.
205
206 ## Where to go next
207
208 - [HyperPod](/aws/compute/hyperpod)
209 - [EKS](/aws/compute/eks)
210 - [Choosing a runtime](/aws/compute/choosing-a-runtime)
211 - [SQL](/sql)
212 - [CHANGELOG](https://github.com/alchemy-run/alchemy/blob/main/CHANGELOG.md#v200-beta65)
213 - [Compare v2.0.0-beta.64 → v2.0.0-beta.65](https://github.com/alchemy-run/alchemy/compare/v2.0.0-beta.64...v2.0.0-beta.65)