Apache Celeborn is a shuffle and spilled-data service for large data-processing workloads. Its management API is easy to overlook because the product is usually discussed as middleware inside a Spark or Flink cluster. The API can still be reachable from a network if its bind host and authentication settings are left at permissive defaults.

ZoomEye returned three matches for title="Celeborn" and zero for app="Apache Celeborn" on 2026-11-22 at 00:45 UTC. The query used sub_type=all, page 1, and pagesize=1. Three is a small sample. It is enough to show a title signal, not enough to estimate the deployed population.

Inspect the title query in ZoomEye. The application gap may reflect a weak product fingerprint, a page title that does not match the application classification, or both. A zero in one field is not an all-clear.

The REST API is part of the cluster boundary

Celeborn's documentation says the REST API is available on both master and worker processes. Master and worker HTTP hosts default to 0.0.0.0; the documented default ports are 9098 for the master and 9096 for the worker. Swagger UI is mounted under /swagger.

The security documentation says SASL authentication and TLS are supported but celeborn.auth.enabled defaults to false. The REST documentation also notes that /healthz bypasses HTTP authentication by default because a kubelet cannot present credentials. Those defaults do not prove that every deployment is open. They do show why an operator must distinguish a health probe from the authenticated management API.

The master controls shuffle services and cluster metadata. A worker serves data associated with active jobs. Exposing either endpoint can give an attacker reconnaissance information even when the API refuses a write. If authentication is disabled, the management surface can become much more than a liveness check.

Separate the three network paths

Test the master REST port, the worker REST port, and the health endpoint independently. Record the bind address, TLS state, authentication response, and whether Swagger is exposed. A health response is not evidence that the management API is safe.

If the deployment runs in Kubernetes, prefer a service and network policy that allow only the components that require access. Put authentication and TLS at the service boundary as well as the application. Do not rely on an internal Kubernetes network name as a security control if the service type or ingress reaches a broader network.

Review the data stored behind the service. Celeborn is not a database, but it handles intermediate records from analytic jobs. The confidentiality impact of an exposed worker depends on the data in flight, the job isolation model, and whether the worker accepts requests from an untrusted peer.

Use the small sample carefully

Three title matches should be treated as individual leads. Validate the page title, response headers, API paths, and owner before calling any of them a Celeborn cluster. A title may appear in documentation or a dashboard that references the product.

For each confirmed cluster, check celeborn.master.http.host, celeborn.worker.http.host, the ports, celeborn.auth.enabled, TLS configuration, and any bypass path list. Compare the running configuration with the intended deployment manifest rather than assuming that a saved file is active.

Run a read-only request from a permitted network location, such as the health endpoint or an authenticated status call, and confirm that unauthorized access fails. Preserve the exact response and timestamp. The article's ZoomEye count is only the external observation.

What the measurement supports

The query supports a small inventory lead and a warning about relying on the app fingerprint. It does not support a claim that three Celeborn clusters are vulnerable. The next step is ownership and configuration evidence.

The operator who finds a match should ask who owns the workload, which cluster it belongs to, and whether the REST endpoint is reachable from outside the job network. If the answer is uncertain, close the external path first, then continue the review from a controlled management network.

References