Cases45 cases

6 to 10 of 45

1. IL-0581 29CM product display servers lost pods and saturated Netty event loops on the first day of the Igu Week sale

   MUSINSA2025

   Symptoms

   Shortly after 29CM's Igu Week sale began on 3 November 2025, the servers for the display domain, which serve search results and product listings, showed anomalies. Some pods went down and stopped taking traffic, and the remaining pods received more traffic than they could handle and saturated their Netty event loops.

   Probe failuresIntermittent connection failuresHigh latency

   Trigger

   Igu Week, 29CM's largest Black Friday sale and its busiest traffic of the year, raised Redis network traffic to about four times normal.

   Root cause

   Capacity exhaustion

   Component

   In-memory cacheLiveness and readiness probes

   Contributing factor

   Unknown limitMonitoring blind spotHealth check tied to a dependency

   Detection

   Anomalies on the display servers, pods going down and Netty event loops saturating, were the first sign. The Redis cause was found from Redis health check failure logs and then the Redis network metrics.

   Remediation

   • Scaled the Redis nodes up from cache.r7g.large to cache.r7g.2xlarge, about twice the baseline network bandwidth

   Action items

   • Extended the monitoring dashboard to show in real time whether network in and out bandwidth exceeds the baseline, with metrics such as aws.elasticache.network_bandwidth_out_allowance_exceeded and aws.elasticache.traffic_management_active
   • Added Datadog alerts that fire when network bandwidth crosses a threshold
   • Added a Caffeine local cache in front of Redis for data that changes rarely
   • Look into compressing cached data with Snappy to reduce Redis network bandwidth
   • Look into storing cached data as Protocol Buffers instead of JSON; measured on the cached Item Document data, Snappy and protobuf together are expected to cut the size by about 73%
   • Keep improving the architecture for traffic growth, building on the cache restructuring, stronger monitoring and read separation already applied

   Lessons

   CPU and memory metrics alone did not show the problem, so Redis needs network monitoring against the node's bandwidth baseline. Burst credits kept the nodes working for about two hours after traffic passed the baseline, which delayed the failure and made its cause slower to find.

   View case
2. IL-0100 PostHog's feature flags pods were packed onto Kubernetes nodes that were too small for their CPU requests, and CPU saturation above 90% caused three outages in ten days through Redis overload, connection pool exhaustion and slow queries

   PostHog2025

   Symptoms

   On 21 October 2025 at 21:45 UTC a change to timeout handling was deployed to the feature flags service, and two minutes later automated monitoring reported higher error rates. The change was rolled back by 21:49, but error rates stayed high: about 38% of flag evaluation requests in the US datacenter returned errors. At 22:30 Redis metrics showed memory exhaustion on ElastiCache, at 22:35 Postgres connections spiked, and at 22:45 the team found a massive data transfer from Postgres into Redis. The main PostHog application was affected as well, because it shares the Redis instance with the flags service.

   Sustained high node CPUHigh error rateHigh latencyContainer crash loopDatabase overload

   Trigger

   On 21 October a deployment meant to reduce timeout errors (PR #39821); on 24 October the deployment of IP-based rate limiting (PR #40074) that saw all traffic as coming from the load balancer's IP; on 28 October an unrelated routine deployment that rolled the feature flag pods in the US region.

   Root cause

   Capacity exhaustion

   Component

   Node poolPod scheduling and placementIn-memory cacheDatabase

   Contributing factor

   No alert on limitRetry stormIneffective rollbackSingle point of failureNo staged rolloutUnrepresentative staging

   Detection

   On 21 October automated monitoring detected higher error rates two minutes after the deploy. On 24 October no alert existed for 429 errors, and the team learned of the incident from customer reports that began at 18:10 UTC, a 62-minute detection delay. On 28 October on-call received a high error count alert. On 29 October the incident was reported with higher error rates and latency. No alert existed for CPU pressure, so the CPU saturation behind all three incidents was not seen directly.

   Remediation

   • rolled back the timeout handling change within 2 minutes on 21 October, which did not stop the errors
   • throttled connections and Redis writes on 21 October
   • halted the rollout on 28 October, scaled the US flags pods to zero until Redis recovered, then scaled them back gradually
   • rolled back hardware changes on 30 October at 00:03 UTC, which removed most errors but not the latency
   • increased the pod resource requests and adjusted the Kubernetes fleet size so fewer pods ran on each node
   • added retry limits to Envoy between the load balancer and the service
   • moved database connection timeouts and other critical settings to runtime configuration and raised the timeout values
   • added per-node and per-pod CPU alerts at 80% sustained for 5 minutes that page on-call

   Action items

   • Form a strike team from the flags, ingestion and infrastructure teams to review the application and infrastructure for remaining bottlenecks
   • Investigate separating the flags Redis instance from the application Redis instance
   • Before re-enabling ArgoCD sync, align the flags service chart configuration with the values in ArgoCD and define the deployment and Redis strategy
   • Make the feature flags team hard code-owners for flags-related code
   • Build a dashboard of key flag metrics with runbook links and update the feature flag runbooks
   • Add missing alerts on existing service and infrastructure metrics
   • Make readiness checks validate dependencies that degrade under load, for example by pinging the database
   • Rate-limit cache population so it cannot overwhelm Redis
   • Monitor connection pools with automatic throttling near exhaustion and limit concurrent connections
   • Build a production-scale load testing framework
   • Build progressive rollout infrastructure to limit blast radius
   • Document Postgres-to-Redis data flows and write runbooks for data transfer storms

   Lessons

   Nodes that are small relative to pod requests let the scheduler pack many pods on each node, and the resulting CPU saturation showed up three different ways: excessive parallelism, connection pool exhaustion and slow queries. Treating each symptom separately let the problem return; right-sizing the pods and nodes fixed all three. CPU must be alerted on from the start. Unbounded Envoy retries and unbounded cache population amplified the overload, and a Redis instance shared with the main application spread a flags problem to the main application. A rollback did not end the incident, because the data transfer and resource exhaustion it had started kept running.

   View case
3. IL-0150 A latent race condition in DynamoDB's DNS automation left an empty DNS record for dynamodb.us-east-1.amazonaws.com, and the DynamoDB and EC2 launch failures that followed caused EKS, ECS and Fargate container launch failures and cluster scaling delays in us-east-1

   Amazon Web Services2025

   Symptoms

   At 11:48 PM PDT on 19 October 2025, every system that reached DynamoDB in us-east-1 through the public regional endpoint started getting DNS failures and could not connect. This included customer traffic and internal AWS services that depend on DynamoDB. DynamoDB global table replicas in other regions still answered but fell behind in replication to and from us-east-1.

   DNS resolution failuresHigh error rateScale-up failureIntermittent connection failuresHigh latency

   Trigger

   One DNS Enactor, slowed by unusually high retry delays, applied a much older DNS plan to the regional DynamoDB endpoint just as another Enactor that had applied the newest plan ran its clean-up, which deleted that older plan and with it every IP address of the endpoint.

   Root cause

   Software defect

   Component

   Authoritative DNS servers and recordsDatabaseCloud load balancer

   Contributing factor

   Automation amplificationRetry stormNo runbook

   Detection

   DNS failures to DynamoDB began immediately at 11:48 PM PDT and the engineering teams of affected AWS services were engaged at once; by 12:38 AM they had identified DynamoDB's DNS state as the source. The later NLB health check problem was detected by monitoring at 6:52 AM.

   Remediation

   • applied temporary mitigations by 1:15 AM that let some internal services reach DynamoDB and repaired key internal tooling
   • restored all DynamoDB DNS information by 2:25 AM through manual operator intervention
   • throttled incoming work and selectively restarted DWFM hosts from 4:14 AM so droplet leases could be re-established
   • reduced the load on Network Manager to bring network configuration propagation times back to normal by 10:36 AM
   • disabled automatic health check failover for NLB at 9:36 AM, and re-enabled it at 2:09 PM after EC2 recovered
   • relaxed the EC2 request throttles from 11:23 AM until all EC2 APIs and launches were normal at 1:50 PM

   Action items

   • Disable the DynamoDB DNS Planner and DNS Enactor automation worldwide
   • Fix the race condition and add protections against applying incorrect DNS plans before re-enabling the DNS automation
   • Add a velocity control to NLB that limits how much capacity a single load balancer can remove when health check failures cause availability zone failover
   • Build an additional EC2 test suite that exercises the DWFM recovery workflow at scale
   • Improve throttling in EC2 data propagation systems so incoming work is rate limited by the size of the waiting queue

   Lessons

   A single DNS record that automation can empty is enough to take a widely used regional endpoint off the network, and every service built on it fails in turn. A check made once at the start of a long operation can be stale by the time the operation writes its result. Recovery of the dependency is not recovery of the dependents: systems that lost state during the outage, such as EC2's droplet leases and network propagation backlog, can collapse under their own catch-up work, and container services that rely on new instance launches keep failing to launch and scale until that work is done.

   View case
4. IL-0413 LLM inference engines timed out or ran out of resources under load in about 60% of 156 severe incidents at Microsoft

   Microsoft2025

   Symptoms

   In Microsoft's cloud LLM inference service, the inference engines themselves failed in 94 of the 156 high-severity incidents from April to June 2025 (≈60%). Requests timed out or ran into resource exhaustion under load, and callers got HTTP 408, 429 and 5xx errors.

   Intermittent timeoutsHigh error rateHigh latency

   Trigger

   Bursty load, traffic with unusually high sampling parameters and long-running requests pushed the inference engines into resource exhaustion and timeouts.

   Root cause

   Capacity exhaustion

   Component

   Model inference server

   Contributing factor

   Monitoring blind spot

   Detection

   Across the 156 incidents, 115 (≈74%) were auto-detected, meaning health probes or DRI pipelines alerted before customers reported the problem; the others were not.

   Remediation

   • Enabled dynamic pushback at the load balancer with caps on pending tokens and requests
   • Tuned request batching and GPU/CPU replica counts
   • Improved gRPC backoff and error mapping
   • Limited sampling parameters and throttled per client
   • Enabled streaming, tuned request batching and capped pending tokens against engine connection timeouts on long-running requests
   • Increased capacity, routed traffic to healthy endpoints and restarted nodes, the actions applied mostly to inference engine failures

   Action items

   • Optimise serving for long-running requests and high sampling regimes, with connection liveness and GPU capacity-aware routing as the main levers
   • Expand health-probe-driven failover and GPU capacity-aware routing so traffic moves off degraded endpoints on its own
   • Forecast token and throughput growth and automate capacity increase decisions
   • Deploy adaptive alerting and expand engine-level telemetry to close detection gaps

   Lessons

   Hotfixes restore correctness but lengthen time to mitigate, while traffic routing and node rebalancing mitigate faster under overload and timeout patterns. Even with AIOps tooling a meaningful share of incidents still needs manual detection or judgement, so health-probe-driven traffic routing, restarts and capacity increases are the decisions to automate first.

   View case
5. IL-0247 One in a million search requests took 100 times longer on Pinterest's Manas leaf nodes after moving to Kubernetes

   Pinterest2025

   Symptoms

   During performance testing of Pinterest's search system Manas on its new Kubernetes platform, a small but persistent share of search requests timed out. About one request in a million took 100 times longer than usual.

   High latencyIntermittent timeouts

   Root cause

   Software defect

   Component

   Telemetry agentLinux kernel

   Contributing factor

   Unreviewed defaults

   Detection

   Performance testing with traffic sent to a test cluster showed a small but persistent percentage of requests timing out.

   Remediation

   • Disabled cAdvisor's working set size estimation on all PinCompute nodes

   Lessons

   CPU shielding does not fully isolate a process from its neighbours. Showing early that the node image was not the cause narrowed the search and avoided dead ends. Simple blackbox debugging, breaking things until the system stops working, works as a way of bisecting the problem.

   View case

By root cause7