Services that run continuous CPU profiling and call crypto/rand often crash with fatal error: invalid runtime symbol table, naming a different function each time (one production trace was in crypto/tls.(*halfConn).decrypt). Services without profiling don't crash. The reporter saw it on go1.25.8 and go1.26.0 on linux/amd64 with kernel 6.12, where Go uses the vDSO getrandom. The code is the same on master. The cause: runtimeΒ·vgetrandom1 in sys_linux_amd64.s stores m.vdsoSP before m.vdsoPC. sigprof starts unwinding from (vdsoPC, vdsoSP) once vdsoSP != 0, so a SIGPROF between the two stores sees a new SP with a stale PC (usually 0). It then reads a stale stack slot as the return PC, and pcvalue throws. nanotime1/walltime on amd64, and vgetrandom1 on arm64, already store PC first. A reproducer crashed 4 out of 4 runs within 2β14 s on 8 CPUs. With the reordering it ran 600 s clean. Gopherbot links CL 864861 ("runtime: fix order of m.vdso* writes"). No milestone yet.
π₯ What Changed
A quiet Wednesday, so three items. The big one: Go services running continuous CPU profiling can crash with "invalid runtime symbol table" when they call crypto/rand on Linux 6.11+ (vDSO getrandom) (#83145). The cause is a two-instruction store ordering bug in the amd64 assembly, and the fix CL is up. Also: Cilium 1.20.5 / 1.19.11 / 1.18.17 fix a fail-open egress policy: toServices plus toPorts allowed any destination when the Service had no backends. For Worth Your Time, Kubernetes SIG Network schedules removing kube-proxy's IPVS mode in 1.40 and explains why.
Also seen: Go #83156 a C-created thread calling into Go gets preempted ~20 Β΅s after taking an idle P in exitsyscall (stale sysmon schedtick; not a regression, reproduces back to 1.21). Go #83163 http.ServeContent range-sum int64 overflow gives a negative Content-Length and an empty 206 (needs ~24 TiB content; reporter says not a security issue). Go #82948 (CVE-2026-97047) still open on the Go1.28 milestone; reporter credited 11-17. ZITADEL v4.20.4 makes PKCE the default for Google/GitHub/GitLab/Azure AD IdPs. Eclipse Mosquitto CVE-2026-110486 (RCE via websockets listeners) plus three more Mosquitto CVEs on oss-security. KAFKA-21568 (Blocker) RemoteIndexCache drops eviction cleanup tasks when restoring an oversized disk cache (tiered storage). No new releases since yesterday morning from Kubernetes, NATS, etcd, grpc-go, Kafka, franz-go, OTel-Go, SpiceDB, Keycloak, Istio, Envoy, Temporal or Vault.
Holds: Go #83145 fix (CL 864861); Go #83126 backports #83136/#83135 β 1.27.5/1.26.12 together with #82948 and #82374; Go #83036 phiopt fix (#83076 open); Go #83068 fix (#83078 open); KAFKA-21491 (open); Keycloak CVE-2026-120984 upstream fix; Kafka 3.9.3 (code freeze 11-20); Postgres 19 GA ~11-19.
The patch releases include the fix from cilium/cilium#49430 ("policy: Treat unresolved toServices as selecting nothing"). An egress rule that combines toServices with toPorts turned into an L3 wildcard allow on those ports when the referenced Service didn't exist, or was selector-less with no backends. The CNP watcher translates toServices into generated ToEndpoints/ToCIDRSet, but with nothing to translate both stay empty. The policy entry then has an empty L3 and a non-empty L4, which Cilium treats like a toPorts-only rule. According to the PR, every release from v1.17 on is affected, since validation stopped rejecting that combination. The same releases also fix an L7 redirect collision (HTTP and TLS rules on one endpoint/port now get separate proxy listeners), a NodePort LB map entry leak, unredacted Hubble URL queries, and a DNS-proxy bootstrap race. There's no GHSA for the egress fix at scan time; it's listed as a bug fix.
toServices egress rules, upgrade to the matching patch release, and treat it as a reminder that zero-endpoint selectors are an edge case to test in any policy engine.
π Worth Your Time
SIG Network deprecated ipvs mode in 1.35 and now plans to remove it in Kubernetes 1.40. The post is a compact history of service-proxy design. iptables had O(n) per-connection rule matching and, worse, full-ruleset rewrites on every Service/Endpoint change (in big clusters kube-proxy pinned a CPU). IPVS fixed both with kernel hash tables and incremental updates, but it never matched every Service edge case. It had no batch-update API, so large change bursts were slower. It still fell back to iptables/ipset for non-LB features, and it inherits iptables' RHEL end-of-life problem. The post also explains why IPVS schedulers (even Maglev mh) don't help: kube-proxy instances don't share scheduler state across nodes, so they do no better than random selection. Advice: on 1.36+ with kernel 5.13+, switch to mode: nftables (GA since 1.33, scaling fixes only from 1.36; benchmarks show first-packet latency on par with ipvs). Otherwise go to iptables, which has improved a lot since 2017 (partial syncs via KEP-3453, assume-current-state updates).
mode: ipvs, and a strong system-design talking point on service load balancing.