deploy= isn't wired up on the database helpers, so your dict never reaches the graph. That's why the plan shows a removal instead of a no-op: the compiled node has no deploy.sleepApplication at all, so diffing it against the environment (where the dashboard already set it to true) comes out as true -> null.

postgres() is a thin wrapper over database(), and database() cherry-picks options by name:

def database(name: str, engine: str, **options: Any) -> Service:
    image_name = options["image"]
    output = options.get("output") or "DATABASE_URL"
    node: dict[str, Any] = { ...address / type / kind / engine / name / image / output / source... }
    if options.get("defaultMountPath"):
        node["defaultMountPath"] = options["defaultMountPath"]
    if options.get("region"):
        node["deploy"] = {"multiRegionConfig": {options["region"]: {"numReplicas": 1}}}
    return Service(node)

Only image, output, defaultMountPath and region are ever read. Anything else is swallowed by **options and dropped without a warning. service() goes through a different path (_service_node -> _normalize_deploy) which does merge deploy, which is why the same dict works there and not here.

Your call is doubly unlucky: region="sfo" assigns node["deploy"] outright, so even if deploy were merged in, the region branch would overwrite it. Both kwargs land on the same key.

Compiling the node shows it (railway-sdk, railwayapp/railway-py-sdk @ e2df8d4):

\>>> postgres("Postgres", region="sfo", deploy={"sleepApplication": True}).to_graph()["deploy"]
{'multiRegionConfig': {'sfo': {'numReplicas': 1}}}

\>>> "deploy" in postgres("Postgres", deploy={"sleepApplication": True}).to_graph()
False

\>>> service("web", source=github("org/app"), deploy={"sleepApplication": True}).to_graph()["deploy"]
{'sleepApplication': True}

Until the helper takes it, set the field on the node yourself. with_fields() is a public method on the returned Service:

Postgres = postgres("Postgres").with_fields(
    deploy={
        "multiRegionConfig": {"sfo": {"numReplicas": 1}},
        "sleepApplication": True,
    }
)

with_fields shallow-merges onto the node, so it replaces deploy wholesale. That's why the region has to be written back by hand, and why you want to drop region="sfo" from the postgres() call rather than leave both in and depend on which one runs last. Result:

{'multiRegionConfig': {'sfo': {'numReplicas': 1}}, 'sleepApplication': True}

Same shape service() would have produced, so the plan should stop trying to unset the field.

Probably worth an issue on railwayapp/railway-py-sdk too. The helpers accept **config and silently discard whatever they don't recognise, so any typo or unsupported key fails exactly like this: no error, and a plan that quietly reverts something you set in the dashboard.