Cases9 cases

2 to 6 of 9

1. IL-0128 Network interfaces that dropped for a few seconds made NCCL time out and stalled or slowed LLM training on a cluster of more than 10,000 GPUs

   ByteDance2024

   Symptoms

   ByteDance's large language model training jobs on its MegaScale system, which runs on more than 10,000 GPUs, occasionally stalled or slowed down. The cause was network interfaces that flapped frequently.

   Link flappingNCCL timeoutThroughput degradation

   Trigger

   Network interfaces went down for several seconds and came back up, dropping every packet in transit.

   Root cause

   DNS or network fault

   Component

   NCCLNIC, cable and switch port link

   Contributing factor

   Unreviewed defaults

   Detection

   MegaScale's monitoring system uses second-level monitoring to track link flapping and other NIC issues. When the RDMA traffic metrics carried in the training daemons' heartbeats drop sharply or fluctuate abnormally, the driver raises an alert for manual investigation, and if the traffic stops entirely it starts the fault recovery procedure automatically.

   Remediation

   • set NCCL's timeout threshold explicitly to a larger value than the default
   • tuned NCCL's retransmit timer and retry count for fast recovery under link flapping
   • enabled the adap_retrans feature on the NICs
   • applied lower-level quality control to NIC signal strength, AOC cable quality and switch-side signal strength

   Lessons

   A default NCCL timeout can be shorter than the few seconds an interface needs to come back from a flap, so a brief link loss becomes a failed collective and a stalled job. A larger timeout and faster retransmission give the link time to return, but the flapping itself came from physical link quality between NIC, AOC cable and switch, and quality control at that level reduced it.

   View case
2. IL-0083 Intermittent 5xx errors in a multi-tenant Kubernetes cluster came from internal DNS: the node-local dnsmasq hit its 150 concurrent query limit, did not cache cluster.local answers, and ndots search lookups flooded CoreDNS

   Adevinta2023

   Symptoms

   On a Friday afternoon a tenant of SCHIP, Adevinta's multi-tenant Kubernetes runtime, reported a spike of HTTP 5xx errors in some of its services. The drops in quality of service came back periodically through the following days, each one short, and were reported mostly around midnight.

   High error rateDNS resolution failuresHigh latency

   Root cause

   DNS or network fault

   Component

   CoreDNSClient-side DNS resolver and cacheTelemetry agent

   Contributing factor

   Unreviewed defaultsUnknown limitIgnored early signalUnverified assumption

   Detection

   A tenant reported a spike of HTTP 5xx errors in some of its services; the DNS Latency SLI had already been degraded for some time without being acted on.

   Remediation

   • raised the dnsmasq concurrent query limit from 150 to 1500
   • added a CoreDNS header block that sets the ra and aa flags and clears rd, so dnsmasq caches cluster.local answers
   • configured dnsmasq on each node to answer NXDOMAIN for unresolvable search-domain names instead of forwarding them to CoreDNS
   • changed the fluent-bit log router hostname to a fully qualified name
   • asked the affected tenants to use fully qualified names for internal calls

   Lessons

   The DNS Latency SLI had shown the problem long before anyone looked at it. An SLO only protects you when its SLIs are trusted and a drop is investigated. The investigation took a month partly because several on-call groups worked on it separately, without one place for observations and hypotheses, and sometimes chased hypotheses already discarded. Disprove one hypothesis before starting the next, and collect data the metrics system lacks: DNStop showed the garbage lookups that no dashboard did. With ndots:5, short service names multiply DNS traffic; a node-local cache only helps if it can cache the cluster DNS answers at all.

   View case
3. IL-0178 On EKS, HPA scale-in returned 502 errors because containers stopped before their pod IPs left the Service endpoints, and after blue/green switches conntrack kept sending established connections to deleted pods behind an instance-mode ALB

   Lotte E-commerce2022

   Symptoms

   LOTTE ON, Lotte E-commerce's online shopping mall of about 40 microservices on Amazon EKS, returned errors for a period of time whenever the Horizontal Pod Autoscaler (HPA) scaled pods in after load dropped. Requests reached containers that had already been deleted and failed with 502 errors.

   High error rateIntermittent timeoutsStale endpoint routing

   Trigger

   HPA scale-in deleting pods, and the deletion of Blue pods after a blue/green switch of the Service label selector.

   Root cause

   DNS or network fault

   Component

   Horizontal pod autoscalerPod terminationkube-proxyconntrack tableCloud load balancerIngress controller

   Contributing factor

   Unreviewed defaults

   Remediation

   • delayed container shutdown on pod deletion (a wait in the SIGTERM handler or a preStop sleep, capped by terminationGracePeriodSeconds) so containers stop only after the pod has left the Service endpoints
   • switched some workloads' Ingress-created ALBs from instance mode to IP mode so traffic goes straight to pod IPs without iptables and conntrack

   Action items

   • Move all workloads to ALB IP mode gradually

   Lessons

   Removing a pod from Service endpoints and stopping its containers happen at the same time in Kubernetes, so a pod can stop while traffic is still routed to it unless shutdown is delayed with a preStop hook or in the application. With an instance-mode ALB, traffic passes through node iptables and conntrack, and established connections can keep going to a pod IP after its iptables rules are gone. IP mode avoids that path, within the limit of 1000 targets per target group.

   View case
4. IL-0199 A non-essential container exceeding its memory limit made GitHub's production pods unavailable, and their restarts failed because a DNS maintenance change had left the clusters unable to reach the image registry

   GitHub2020

   Symptoms

   At 08:18 UTC on 13 July 2020, production Kubernetes pods began to be marked unavailable. The loss of pods spread through the clusters and reduced capacity until GitHub.com services went down.

   OOMKilledImage pull failuresService outage

   Trigger

   A container in the production pods exceeded its memory limit and was terminated; a routine DNS maintenance operation completed earlier had left the clusters unable to reach the image registry.

   Root cause

   DNS or network fault

   Component

   Container memory limitContainer registry and its credentials

   Contributing factor

   Overlapping failuresHarmful manual actionSingle point of failure

   Remediation

   • restarted the process with the cached DNS records so the clusters could reach the registry
   • fetched the container images, redeployed and recovered the services

   Action items

   • Improve monitoring so pod restarts cannot fail again in the same pattern
   • Minimise the dependency on the image registry
   • Expand validation during DNS changes
   • Re-evaluate all existing Kubernetes deployment policies
   • Improve progressive deployments so the impact of a deployment can be evaluated incrementally

   Lessons

   A container that production traffic does not need can still take pods out of service, because a pod is available only when all of its containers are healthy. ImagePullPolicy Always turns every container restart into a registry dependency, so an unrelated DNS change that breaks registry access turns a one-minute self-healing event into a long outage, and a mass redeploy spreads it.

   View case
5. IL-0056 Calls from EKS pods to services behind an NLB in another VPC failed intermittently because the AWS VPC CNI did source NAT on the node

   MindTickle2020

   Symptoms

   On the Saturday the production cluster was moved from kops to EKS, the platform became very slow and too unstable to keep running, and the team reverted.

   Intermittent connection failuresHigh latency

   Trigger

   Production was moved from a kops cluster using Weave networking to a new EKS cluster using the AWS VPC CNI, without external SNAT enabled.

   Root cause

   DNS or network fault

   Component

   AWS VPC CNI

   Contributing factor

   Unreviewed defaultsUnverified assumption

   Detection

   The team lead saw the platform behaving very slowly after the switch to EKS and phoned the author.

   Remediation

   • reverted production to the kops cluster
   • set AWS_VPC_K8S_CNI_EXTERNALSNAT=true in the AWS VPC CNI

   Action items

   • Monitor and alert on the AWS VPC CNI

   Lessons

   In a networking problem every component in the stack is a suspect; ruling out EKS early cost a lot of time. A DaemonSet that pinged every component from every other one exposed the anomaly that led to the fix. The team had only skimmed the AWS CNI documentation before the migration; knowing how the plugin works internally would have led to better judgements and hypotheses while debugging.

   View case

By root cause1