SSH Access
Table of Contents
Generated sshd_config
The operator generates a complete sshd_config and mounts it read-only over
/etc/ssh/sshd_config, so the image’s own file and any setting it carried do
not apply. Login pods always get it; worker pods get it when spec.ssh.enabled
is set, which is how pam_slurm_adopt confines a session to
the job that owns the node.
extraSshdConfig is appended to the end of that file. It exists on a LoginSet,
and under spec.ssh on a NodeSet:
loginsets:
slinky:
extraSshdConfig: |
ClientAliveInterval 60
ClientAliveCountMax 3
Client Keepalives
Nothing in the generated file sets ClientAliveInterval, so it keeps OpenSSH’s
default of 0 and sshd never probes its clients. That leaves a long-lived
session exposed at both ends:
Idle sessions are dropped by whatever sits in front of the pod. The chart exposes login pods through a
LoadBalancerservice, and cloud load balancers close idle flows on their own schedule — 350 seconds on AWS NLB, 600 on GCP, 4 minutes by default on Azure.TCPKeepAlivedoes not help, since Linux waits 7200 seconds before its first probe.A client that goes away without closing leaves
sshdwaiting indefinitely. On a worker pod, that session also holds the Slurm job it was adopted into.
The settings above address both: the probes travel inside the encrypted
connection, so the load balancer sees traffic and keeps the flow open, and an
unresponsive client is disconnected after ClientAliveCountMax missed probes.
Keep the interval below the idle timeout of the load balancer in front of the
login service — 60 seconds clears the shortest of the timeouts above — and note
that a shorter interval reclaims dead sessions sooner but disconnects clients on
briefly flaky links.
Overriding Operator Settings
sshd_config uses the first value obtained for each keyword, and
extraSshdConfig is appended last. A keyword the operator already emits
therefore cannot be changed there; the added line is parsed, found to be a
repeat, and ignored. Those keywords are Include, UsePAM and Subsystem on
both pod types, plus X11Forwarding on login pods and AuthenticationMethods
on worker pods. Anything else, ClientAliveInterval included, is undefined
until you set it.
To change one of them, mount a drop-in such as
/etc/ssh/sshd_config.d/10-local.conf through the pod template. The Include
directive is processed before the rest of the file, so a value set there wins
over both the operator’s settings and extraSshdConfig.