Cases18 cases

13 to 17 of 18

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-0013 The conntrack table filled up on downsized GKE nodes and a high-traffic aggregation proxy saw climbing connection resets

   loveholidays2020

   Symptoms

   Shortly after a round of changes, loveholidays saw connectivity problems and raised error rates on its aggregation service. The service is a proxy between two other services of more than 100 instances each, handles over 10,000 requests per second continuously and can spike to four times that.

   Connection resetsHigh error rateIntermittent connection failures

   Trigger

   Three changes together: a change in service architecture, halving the RAM of the nodes that run the aggregation service, and sharding a downstream service, which doubled the number of connections to track.

   Root cause

   Capacity exhaustion

   Component

   conntrack tablekube-proxyNode poolHAProxy

   Contributing factor

   Unknown limitNo capacity planningNo alert on limit

   Detection

   Rising error rates and climbing connection reset errors in HAProxy, then node_exporter conntrack metrics compared with their limit.

   Remediation

   • moved the aggregation service to nodes with more RAM
   • limited the number of aggregation service pods per node

   Action items

   • Alert when conntrack table use goes above 75%, with (node_nf_conntrack_entries / on (pod) node_nf_conntrack_entries_limit / on (pod) group_right kube_pod_info) > 0.75

   Lessons

   The conntrack limit is tied to node memory, so resizing a node pool silently resizes the connection table, and on GKE the ratio did not match the commonly cited formula. A high-traffic proxy on small nodes can exhaust it, and the symptoms are connection resets and timeouts, so conntrack use needs its own metric and alert.

   View case
3. IL-0015 A GKE cluster's /16 pod address range ran out at 256 nodes, leaving new pods Pending and node scale-ups failing

   loveholidays2020

   Symptoms

   A colleague wrote in Slack that his application was taking a long time to deploy. A deployment that normally finishes in minutes had been running for about an hour: about half of the new pods were running and serving traffic, and the other half were Pending.

   Pods PendingIP address exhaustionScale-up failure

   Trigger

   The cluster reaching 256 nodes, the most its /16 pod secondary range could hold at 256 addresses per node; every further node creation failed.

   Root cause

   Capacity exhaustion

   Component

   VPC subnet or pod address rangeCluster autoscalerNode pool

   Contributing factor

   Unreviewed defaultsUnknown limit

   Detection

   A colleague reported in Slack that a deployment was slow; the pending pods' events and the cluster autoscaler status ConfigMap showed the exhausted IP space.

   Remediation

   • lowered the maximum pods per node on selected node pools, cutting IP consumption by about 30%
   • moved some node pools to larger instance types

   Action items

   • Lower the maximum pods per node on node pools that do not run many pods
   • Use larger instance types in some node pools

   Lessons

   In a VPC-native GKE cluster the number of nodes, not the number of pods, uses up the pod range: by default each node reserves 256 addresses, so even a /16 is full at 256 nodes. Maximum pods per node and node size are the practical levers; larger nodes pack better but take more pods down when they fail.

   View case
4. IL-0037 Image registry rate-limited a whole region after a crash-looping DaemonSet with imagePullPolicy Always pulled through three NAT instances

   Datadog2019

   Symptoms

   Users reported that their jobs were not starting, with some sign that image pulls were involved. Many pods across the cluster were in error states related to image pulls, while applications that were already running were not affected.

   Image pull failuresService outage

   Trigger

   A change to the permissions on a storage bucket that one application needed to read. The application could no longer start, and its DaemonSet, running on roughly a thousand nodes of the cluster, went into a crash loop with imagePullPolicy: Always, so every restart pulled the image again although it was already on the node.

   Root cause

   Capacity exhaustion

   Component

   Container registry and its credentialsNAT gatewayApplication databaseAdmission webhook

   Contributing factor

   Automation amplificationSingle point of failureUnknown limitHarmful manual actionExcessive blast radius

   Detection

   Users reported that their jobs were not starting; monitoring had not caught it first.

   Remediation

   • replaced the affected NAT instances so registry traffic came from new source IPs
   • reconfigured the static IPs and firewall rules so applications could reach the Cloud SQL database again

   Action items

   • Add an admission webhook that rejects floating image tags, starting with latest
   • Apply the image tag webhook to workload objects only, not to pods, after it blocked pods created by in-cluster controllers

   Lessons

   The speakers' takeaway is to be very careful with DaemonSets in large clusters: they say they had broken everything DaemonSets depend on, including DNS, Vault and the image repository, and that it is very easy to break things when managing them by the thousand. imagePullPolicy: Always made every crash-loop restart pull the image again. Sending all registry traffic through three NAT instances put the whole region behind three source IPs, which the registry rate-limited, and changing those IPs cut the connections to a critical database reached over the internet through them. An admission webhook on pods rejects the pods that controllers in the cluster create, where no user sees the error.

   View case
5. IL-0075 Kubernetes API servers pegged on CPU when the cluster grew to about 290 nodes, traced to Datadog agents polling the API

   Civis Analytics2019

   Symptoms

   After Civis released its new script infrastructure on Kubernetes internally in September 2018, developers reported that kubectl commands were slow, and users intermittently could not start notebooks and services.

   Control plane lagControl plane unresponsiveHigh latencyHigh error rateDNS resolution failures

   Trigger

   Migrating clients' Python, R and container scripts from the in-house Bocce system to Kubernetes (the project called Croquet) pushed the cluster to about 290 EC2 instances under a large script load.

   Root cause

   Capacity exhaustion

   Component

   Kubernetes API serverTelemetry agentetcd

   Contributing factor

   Unknown limitUnverified assumptionMonitoring blind spotNo capacity planning

   Detection

   Developer reports of slow kubectl commands and user reports of notebooks and services failing to start, then high CPU on the master nodes.

   Remediation

   • moved internal users back to the old Bocce infrastructure
   • resized the three master nodes from m4.large to m4.4xlarge
   • changed the Datadog setup so per-node agents no longer polled the API server for service metadata
   • raised max-requests-inflight for non-mutating requests from 400 to 1200
   • grew the etcd EBS volumes from 20 to 100 GB (100 to 300 IOPS)
   • switched client scripts from the ClusterFirst DNS policy to Default

   Action items

   • Upgrade etcd from 2.2.1 to 3.x once Kops supports a safe migration
   • Solve the in-cluster DNS timeouts so client scripts can talk to other workloads in the cluster

   Lessons

   Kubernetes can reach thousands of nodes only after tuning, and third-party DaemonSets are a common cause of control plane load: an agent on every node that polls or watches the API server multiplies with the cluster size. Datadog, kube2iam and the fluentd metadata filter all showed this. A small configuration change often fixes it, but finding it took load testing and the API server access logs.

   View case

By root cause5