Slurm: endpoint ports can collide with ephemeral readiness connections

Read packages/data-designer-slurm/src/data_designer/slurm/runtime/ports.py, preflight.py, and entrypoint.sh; trace port selection, reservation cleanup, and readiness polling. Start by reproducing the supplied socket-level MRE, then inspect related issue #868 and existing regression tests. Done means startup avoids the probe and listener race, with tests for overlapping ephemeral ranges, occupied ports, and preserved failure diagnostics.

Priority Level

High (Major functionality broken)

Describe the bug

Slurm's allocation-local port selection can overlap Linux's ephemeral TCP port range. A readiness probe issued before the logical endpoint binds can automatically select that endpoint's port as its own source port, connect to itself, and leave the port in TIME_WAIT. The endpoint then fails with EADDRINUSE, even though the earlier bind preflight succeeded and no other listener was needed.

Steps/Code to reproduce bug

Requirements: Linux with unprivileged user/network namespaces, Bash, Python 3, curl, and iproute2 (ip, ss). No Slurm, GPU, model, dataset, or DataDesigner installation is required for this socket-level MRE.

Save the following as repro.sh and run bash repro.sh. It changes sysctls only after entering a private network namespace. A synthetic one-port ephemeral range makes the source-port choice deterministic; it is not a suggested host configuration.

#!/usr/bin/env bash
set -euo pipefail

export MRE_PARENT_NETNS="$(readlink /proc/self/ns/net)"
export MRE_PYTHON="${MRE_PYTHON:-python3}"

unshare --user --map-root-user --net bash <<'SH'
set -euo pipefail
ip link set lo up
"$MRE_PYTHON" - <<'PY'
import errno
import os
import socket
import subprocess
from pathlib import Path

assert str(Path("/proc/self/ns/net").readlink()) != os.environ["MRE_PARENT_NETNS"]
settings = Path("/proc/sys/net/ipv4")
settings.joinpath("ip_local_reserved_ports").write_text("\n")
# Synthetic single-port range forces an otherwise intermittent source-port choice.
settings.joinpath("ip_local_port_range").write_text("10000 10000\n")

# Mirror the preflight bind-and-release before the endpoint starts.
with socket.socket() as reservation:
    reservation.bind(("0.0.0.0", 10000))
print("Preflight bind succeeded; reservation released.", flush=True)

probe = subprocess.run(
    ["curl", "--noproxy", "*", "--fail", "--silent", "--show-error",
     "--max-time", "1", "http://127.0.0.1:10000/health"],
    capture_output=True, text=True,
)
print(probe.stderr.strip())
subprocess.run(["ss", "-tan", "( sport = :10000 )"], check=True)

with socket.socket() as listener:
    listener.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
    try:
        listener.bind(("127.0.0.1", 10000))
    except OSError as error:
        assert error.errno == errno.EADDRINUSE
        print(f"Endpoint bind failed: {error}")
    else:
        raise AssertionError("Expected EADDRINUSE after the readiness self-connection")

with socket.socket() as control:
    control.bind(("127.0.0.1", 2000))
    print("Control bind outside the synthetic ephemeral range succeeded.")
PY
SH

Observed:

• The preflight bind succeeds and releases its socket.
• Curl reports Received HTTP/0.9 when not allowed after connecting to itself.
• ss shows TIME-WAIT from 127.0.0.1:10000 to 127.0.0.1:10000.
• The subsequent endpoint bind raises [Errno 98] Address already in use, including with SO_REUSEADDR enabled.
• The control bind outside the synthetic ephemeral range succeeds.

Expected behavior

Endpoint startup should not allow readiness probes to consume the endpoint's own port. Allocation should account for the host's effective ephemeral/reserved port ranges, and startup should establish listener ownership before connection-based readiness polling can race it.

Agent Diagnostic / Prior Investigation

Affected code was verified at public commit 68c516f724ee1e0e3e3c5a965388a9610e587207. The public feat/slurm-execution branch still uses the same fixed-range allocation:

• Port selection: 10000 + min(SLURM_JOB_GPUS) * 256, without excluding the host ephemeral range.
• Preflight: test-bind reservations are closed before startup.
• Startup ordering: servers become ready before endpoint processes start.
• Readiness polling: curl starts polling without first establishing that the endpoint owns its listening socket.

The unmodified DD proxy also reproduces the bind failure under the same forced-collision setup. The MRE establishes a failure mechanism and deliberately forces the timing/source-port choice; it does not measure its frequency under an ordinary host configuration.

Additional context

Related to #868. Regression coverage should include a probe before the listener binds, overlapping ephemeral/service ranges, and an already occupied port. Occupied-port failures should remain explicit and cleanup should preserve the failed attempt's diagnostics.

Checklist

• I reproduced this issue or provided a minimal example
• I searched the docs/issues myself, or had my agent do so
• If I used an agent, I included its diagnostics above