Flink

A Flink JobManager container (REST + RPC), with an optional companion TaskManager — Docker only.

Default image: flink:1.20.5Exposed ports: 8081 (REST), 6123 (RPC) Wait strategy: Wait.forHttp("/overview").forPort(8081), 120s startup timeout Memory: withMemoryLimit(1024) on both the JobManager and, when requested, the TaskManager

Member	Returns
FlinkContainer.start(image?)	Promise<FlinkContainer> — boots the JobManager
.withTaskManager()	this — requests a companion TaskManager on a shared Network, started once the JobManager is ready. Docker only — see below
.restUrl	The REST base URL for the running JobManager

Example: JobManager only (both backends)

import { FlinkContainer } from "rightsize/modules";

await using flink = await FlinkContainer.start();
console.log(await (await fetch(`${flink.restUrl}/overview`)).json());

Example: a full session cluster (Docker only)

import { FlinkContainer } from "rightsize/modules";

await using flink = await new FlinkContainer().withTaskManager().start();
// only reaches this point on the docker backend
console.log(await (await fetch(`${flink.restUrl}/taskmanagers`)).json()); // one registered TM

Backend notes: the withTaskManager() limitation, stated precisely

A real Flink session cluster needs a TaskManager to run anything, and a TaskManager registers with its JobManager over a persistent bidirectional RPC connection (Pekko/Akka). Two independent facts rule this topology out on the microsandbox backend:

1. The flink image ships neither nc nor busybox — confirmed directly against the pinned tag (command -v nc and command -v busybox both exit 127 inside the image). microsandbox's network-link emulation depends on an in-guest nc -l listener (see Networking); without it, the emulation's own preflight probe fails before any tunnel could even be attempted.
2. Independent of fact 1, the tunnel's client-speaks-first, one-connection-at-a-time contract wouldn't carry a persistent bidirectional RPC connection anyway — but whether Flink's registration RPC would actually survive being pumped through the tunnel was never tested, because fact 1 already makes the attempt moot.

withTaskManager() therefore throws UnsupportedByBackendError the instant start() resolves the microsandbox backend — before any network setup, port allocation, or the JobManager container is even created — naming both the missing binary and RIGHTSIZE_BACKEND=docker as the remedy. A bare JobManager (no withTaskManager() call) boots and serves /overview normally on microsandbox; only the two-container topology is affected. On Docker, the same builder produces a real two-container session cluster with the TaskManager registering successfully.