physicalName silently truncates — expose pulumi:autonaming enforce:true or error on overflow

[jhw]

Problem

SST's physicalName() in platform/src/components/naming.ts silently truncates resource names when they exceed AWS per-type limits (e.g. Lambda's 64-char cap). The prefixName helper tries app+stage+name, then stage+name, then name-only, chopping the leftmost component to fit. No warning or error is emitted.

This is compounded by Pulumi's AWS provider auto-naming, which also silently truncates when SST doesn't set an explicit name: (the common case for sst.aws.Function).

The truncated name is frozen in Pulumi state at creation time. On subsequent app-name changes, the physical name never updates — leading to orphaned resources still routed through live API Gateway integrations, running stale code against production data.

We lost a full day to this. See our post-mortem for the gory details: 21 Lambdas stuck with names from 8 different truncated app-name variants, stale code running against live DynamoDB, midnight Slack alerts from code that no longer existed in the source tree.

Related closed issues: #4100, #4410

What Pulumi offers

Pulumi 3.x added pulumi:autonaming configuration:

config:
  pulumi:autonaming:
    pattern: ${name}-${hex(8)}
    enforce: true

enforce: true makes the provider error instead of silently truncating. This is exactly the safety net needed.

The gap

SST generates the Pulumi project config dynamically. Only aws:*-prefixed keys are reachable through sst.config.ts via providers.aws.*. The pulumi:autonaming key is a top-level Pulumi config namespace and isn't exposed.

Even if enforce: true were reachable, it only covers the Pulumi auto-naming layer. SST's own physicalName() runs first and would still silently truncate.

Request

Either (or both):

1. Expose pulumi:autonaming config through sst.config.ts, e.g.:

return {
  name: "myapp",
  home: "aws",
  providers: {
    pulumi: { autonaming: { enforce: true } },
  },
};

2. Change physicalName() / prefixName() to throw instead of silently truncating when the computed name exceeds the limit. If a resource name doesn't fit, the developer needs to know immediately — not discover it months later via orphaned resources.

Our workaround

We added manual guards (assertLambdaNameFits, assertCronNameFits, assertBucketNameFits) that compute the expected physical name using the same formula as SST and throw at deploy time if it overflows. Every new sst.aws.Function(...) call site must be preceded by a guard. It works, but it's a code smell — the infrastructure layer should enforce its own naming constraints.