Cases9 cases

4 to 8 of 9

1. IL-0012 CPU limits and an unpatched kernel CFS quota bug throttled multi-threaded services into stalls, failed health checks and timeouts

   Omio2020

   Symptoms

   For about six months, applications on Omio's Google Kubernetes Engine clusters stalled at random: they got stuck, stopped answering health check requests or lost network connections, and nothing the team looked at explained why.

   CPU throttlingProbe failuresHigh error rateHigh latencyIntermittent timeoutsIntermittent connection failures

   Trigger

   Multi-threaded workloads running with CPU limits on nodes whose kernel did not yet have the fix for a known CFS quota bug that causes unnecessary throttling.

   Root cause

   Software defect

   Component

   Container CPU limitLinux kernelkubelet

   Contributing factor

   Unknown limitMonitoring blind spotUnpatched known bug

   Detection

   Months of random stalls and failed health checks led the team to read the cgroup cpu.stat throttling counters inside the pods, which showed far more throttling than the limits should allow.

   Remediation

   • removed CPU limits until the kernel with the CFS bug fix was deployed across all clusters

   Action items

   • Deploy the kernel with the CFS quota bug fix across all clusters
   • Monitor utilisation, saturation and error metrics, API latency and error rates to confirm the change behaves as expected

   Lessons

   A CPU limit is a time quota per 100 ms period, not a share of cores, so a multi-threaded container can be stopped for most of each period long before it uses the cores it was given, and a kernel bug can make that worse. Throttling is invisible in standard Kubernetes metrics and has to be read from the cgroup counters. Removing limits trades away protection against a container that takes a whole node, which is only acceptable when runtimes are tuned.

   View case
2. IL-0193 A GCP persistent disk degradation took down a Cassandra StatefulSet whose OrderedReady policy blocked recovery, and Metrictank could not reconnect to the recreated pods, causing a 23-hour Graphite query outage

   Grafana Labs2019

   Symptoms

   An alert fired because some Metrictank instances could not drain their write queue, which meant they could not write to Cassandra. Google had not yet published an incident, but a team member's Discord server, also hosted on GCP, was down, which pointed at GCP; Google confirmed an SSD persistent disk incident about 90 minutes after it began.

   High error rateNode NotReadyPods PendingService outage

   Trigger

   A Google Cloud incident in which SSD persistent disks performed so poorly that many I/O operations failed for an extended period.

   Root cause

   Cloud provider outage

   Component

   Persistent volumes and stateful setsApplication databaseNode root disk

   Contributing factor

   Overlapping failuresUnpatched known bugUnverified assumption

   Detection

   An alert that Metrictank write instances could not flush their write queue to Cassandra. Error rates alone rose too slowly to notice quickly, because most queries were served from memory.

   Remediation

   • Added compute nodes hoping new nodes could attach the disks, which did not work
   • Deleted and recreated the Cassandra StatefulSet with an updated pod management policy, so all pods but one started
   • Raised Kafka topic retention to avoid data loss during the long downtime
   • Temporarily added nodes so Metrictank restarts were not slowed by CPU starvation
   • Restarted all Metrictank instances, which took more than 20 hours, so they connected to the new Cassandra IPs

   Action items

   • Re-instantiate the gocql connection in Metrictank when it is detected dead, so it reconnects even when all Cassandra IPs change
   • Fix the Metrictank write-queue deadlock by updating an internal data structure with atomics instead of a lock

   Lessons

   An OrderedReady StatefulSet cannot recover past a single pod whose volume will not attach; when every pod is down, the ordering that protects normal start-up blocks the whole cluster. A client that learns peer addresses only from live peers needs a way to go back to the service name when all peers are replaced at once. A long-tolerated bug with a config workaround can surface at the worst time; this one turned out to have a simple fix. Writing incoming data to a Kafka write-ahead log first meant all data could still be ingested after recovery, and Kafka on non-SSD disks escaped the disk incident. With both Metrictank bugs fixed, the team expects a similar event to take less than 3 hours to recover from instead of 23.

   View case
3. IL-0057 Higher p99 latency after moving a reverse proxy to EKS, from outgoing connections delayed by a SNAT conntrack race

   Prezi2019

   Symptoms

   Prometheus alerts fired on tail latency: from 7/24, the p99 response time of a few services went up after Prezi moved its Nginx-based reverse proxy, smartgateway, from dedicated Elastic Beanstalk machines onto shared Kubernetes worker nodes. The bump affected only the tail, so the earlier monitoring of average response times would most certainly have missed it.

   High latencyIntermittent timeouts

   Trigger

   Moving the reverse proxy onto shared EKS worker nodes, where its outgoing connections went through source NAT on the node and some SYN packets were dropped by a race in Netfilter's conntrack.

   Root cause

   DNS or network fault

   Component

   conntrack tableAWS VPC CNILinux kernelNode operating system image

   Contributing factor

   Unreviewed defaultsUnpatched known bug

   Detection

   Prometheus alerts on p99 latency fired after the reverse proxy migration.

   Remediation

   • added AWS X-Ray tracing and an $upstream_connect_time trace attribute to locate the delay
   • ran more pods of the service on more nodes, which reduced but did not end the delays
   • set AWS_VPC_K8S_CNI_EXTERNALSNAT=true on the AWS VPC CNI to turn off SNAT on the nodes

   Lessons

   Alerting on tail latency instead of averages caught a problem that would otherwise have gone unnoticed, and tracing turned a vague slowdown into a measured connect delay. Delays of exactly 1 or 3 seconds on new outgoing connections point to dropped SYN packets, and on EKS the SNAT done by the VPC CNI is where to look.

   View case
4. IL-0030 Most nodes of a cluster went NotReady together for 15 minutes while the API server sat behind an AWS ELB

   Zalando2019

   Symptoms

   Fewer than 20% of the nodes in one Zalando cluster were available. Most nodes reported NotReady at the same moment, and only a handful stayed Ready.

   Node NotReady

   Root cause

   Software defect

   Component

   Kubernetes API serverCloud load balancer

   Contributing factor

   Unpatched known bug

   Remediation

   • none; the nodes recovered on their own after about 15 minutes

   View case
5. IL-0073 Intermittent Kafka connectivity overloaded Docker daemons, set off a node rescheduling cascade and poisoned the Consul gossip mesh

   Target2019sev1

   Symptoms

   An upgrade of the network subsystem under Target's datacenter OpenStack, meant to cause only a brief disruption, ran into problems and left network connectivity disrupted for several hours. Kafka, which runs on that OpenStack, became intermittently available.

   Node NotReadyLoad imbalanceHigh latencyCredentials unavailable

   Trigger

   A network subsystem upgrade in the datacenter OpenStack infrastructure disrupted connectivity for several hours and made Kafka intermittently available.

   Root cause

   Dependency failure

   Component

   Container runtimeService discoveryTelemetry agentCredential broker for pods

   Contributing factor

   Excessive blast radiusAutomation amplificationSingle point of failureUnpatched known bug

   Remediation

   • enabled gossip encryption on Consul so messages from the poisoned mesh were rejected
   • stabilized the deployment engine and redeployed the development environment with encryption, terminating instances without it
   • upgraded Consul in place to 1.2.3 and then to 1.3.0

   Action items

   • Reduce how much of an outage a single Docker daemon can cause, together with running smaller and more numerous clusters

   Lessons

   Smaller and more numerous clusters limited the damage: the smaller development clusters and production were not hit the same way as the one large cluster, and dev and prod on one cluster would have been far worse. A Docker daemon shared by many workloads is a brittle failure point. Per-workload logging and metrics sidecars were judged the right design; a shared cluster-level agent would almost certainly have made the problem worse and harder to fix.

   View case

By root cause5