clover's git

the-knife

1	---
2	title: Deployments
3	description: The fail-closed Compute deployment lifecycle — health-checked promotion, rollback recovery, old-generation cleanup, and env-only redeploys.
4	---
5
6	When the artifact or deployment configuration changes,
7	`Prisma.Compute` creates a new generation and walks it through a
8	fail-closed sequence: upload the artifact, start the candidate, verify
9	its preview health endpoint, promote it, confirm the app reports it as
10	latest, verify the stable endpoint — and only then touch the previous
11	deployment.
12
13	## Health checks and promotion
14
15	`healthCheck.path` gates promotion. If the candidate never becomes
16	healthy, alchemy attempts to destroy it without changing the stable
17	deployment. When a previous deployment exists, it keeps serving:
18
19	```typescript
20	const app = yield* Prisma.Compute("api", {
21	project,
22	path: "./app",
23	port: 3000,
24	healthCheck: { path: "/api/health" },
25	});
26	```
27
28	When promotion succeeds but the stable endpoint fails verification,
29	alchemy attempts to restore the previous deployment. If recovery
30	converges, it cleans up the failed candidate. If recovery cannot
31	converge—or no safe rollback target exists—alchemy fails the deploy,
32	preserves the generations needed for recovery, and retries recovery
33	before making further cloud changes. Terminally failed deployments
34	are never restarted or selected as rollback targets.
35
36	## Old-generation cleanup
37
38	By default the previous deployment is kept (stopped) after a
39	successful promotion. `destroyOldDeployment: true` deletes it once
40	the replacement is fully promoted and healthy:
41
42	```typescript
43	const app = yield* Prisma.Compute("api", {
44	project,
45	path: "./app",
46	destroyOldDeployment: true,
47	});
48	```
49
50	If cleanup fails, alchemy records the pending action in state and
51	retries it before creating another generation.
52
53	## Env-only redeploys
54
55	`skipCodeUpload: true` creates the next deployment from the currently
56	promoted artifact. It requires an existing promoted deployment and is
57	useful when only environment variables changed:
58
59	```typescript
60	const app = yield* Prisma.Compute("api", {
61	project,
62	skipCodeUpload: true,
63	env: { FEATURE_FLAG: "on" },
64	});
65	```
66
67	## Redeploying for changed inputs
68
69	The low-level `Prisma.Deployment` resource owns no environment
70	variables, so it cannot see an env-value change on its own. Prisma
71	also snapshots environment variables into a deployment when the
72	deployment is created, so changing only a value updates the
73	platform's variable record but never reaches the running app.
74
75	Declare the values the deployment must be recreated for with
76	`triggers`. When the resolved value changes, the next deploy
77	replaces the deployment (create-before-delete) even though the
78	artifact is unchanged:
79
80	```typescript
81	const deployment = yield* Prisma.Deployment("web", {
82	app,
83	artifactPath: "./dist/app.tar.gz",
84	triggers: { DATABASE_URL: Redacted.make(databaseUrl) },
85	start: true,
86	promote: true,
87	});
88	```
89
90	The values are folded into the deployment fingerprint as a salted
91	hash and persisted `Redacted`; the plaintext never lands in state.
92	`Prisma.Compute` needs no equivalent — it owns `env` and already
93	folds those values into its own fingerprint.
94
95	## No-op convergence
96
97	Deploys are hash-driven: when the built artifact and configuration
98	are unchanged, reconcile reuses the same healthy deployment instead
99	of creating a new generation — repeat deploys converge to exactly
100	one active deployment.
101
102	## Where next
103
104	- [Apps](/prisma/compute/apps) — builds, Effect-native services,
105	and local dev.
106
107	Reference:
108
109	- [Compute API reference](/providers/prisma/compute)
110	- [Deployment API reference](/providers/prisma/deployment)