You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
bug(network): policy DNS asks one nameserver once with a 2 s timeout, so lookups fail intermittently under load #5770
Under load, a workload in a sandbox intermittently fails to resolve a host its policy allows: failed to lookup address information: No address associated with hostname (EAI_NODATA). Run alone, the same lookup always succeeds.
Seen with v0.1.8 and the Podman driver; the resolver code below is unchanged on main (6c3d773).
Why
Policy DNS asks the upstream once: crates/openshell-supervisor-network/src/policy_dns/resolver.rs sends one UDP query to the first nameserver of the supervisor's /etc/resolv.conf, waits DEFAULT_EXCHANGE_TIMEOUT (2 s, line 21) and does not retry, nor try another nameserver. A single lost reply becomes a failed lookup for the workload; since AAAA queries are always answered empty, glibc reports it as "No address associated with hostname".
Container-runtime resolvers drop queries under concurrency. Asking Docker Desktop/OrbStack's resolver exactly as the supervisor does (one UDP A query, 2 s, no retry, 2000 queries):
at once
answered
lost
SERVFAIL
20
2000
0
0
100
1995
5
0
400
236
0
1764
With many sandboxes, each with its own supervisor asking the same host resolver, this is reached in practice.
Related: #5197 (the first nameserver is fixed at supervisor start).
Workaround
A local caching resolver (Unbound) as the supervisors' nameserver: 4000 of 4000 answered at 800 at once.
Suggested fix
Retry a lost exchange (with backoff, within the workload's own timeout) and fall back to the other nameservers in /etc/resolv.conf; or document that the supervisor's nameserver should be a local caching resolver.
Activity