Cases3 cases

1 to 3 of 3

1. IL-0270 Discord's ScyllaDB authentication cluster overloaded during an OS upgrade, and logins failed for about 50 minutes

   Discord2023

   Symptoms

   On 6 November 2023 at 10:52 AM PST, when an upgrade script drained all connections from zone [c] of the scylla-authentication ScyllaDB cluster, Discord's availability, measured as the share of HTTP requests that return a 200-level response, fell at once to about 25%. Users could not connect, log in, read or send messages.

   Service outageDatabase overloadHigh latency

   Trigger

   During an OS upgrade that takes one zone of a ScyllaDB cluster offline at a time, the upgrade script drained all connections from zone [c] of scylla-authentication at 10:52 AM, while one node in the just-upgraded zone [b] had a degraded NVMe Local SSD.

   Root cause

   Capacity exhaustion

   Component

   Database

   Contributing factor

   No headroomOverlapping failuresDependence on a warm cacheUnverified assumptionIgnored early signalRetry storm

   Detection

   An incident response call had already started when the first alerts fired at 10:56 AM, four minutes after the start. The time spent by downstream service graph showed the authentication service spending far more time on user requests than normal, and the authentication alerts fired moments later.

   Remediation

   • Turned off sending typing events and marking messages as read to shed load from the authentication cluster
   • Waited for the zone [c] RAID rebuild instead of restarting nodes, and the two zone [c] nodes came back online at 11:41 AM
   • Restarted the two zone [b] nodes one at a time

   Action items

   • Double the size of the scylla-authentication cluster so each node takes half as many requests
   • Rebuild scylla-authentication-b-2 with new, non-degraded disks
   • Build a way to detect disk degradation that is too subtle to affect overall latency but still threatens reliability
   • Move upgrade operations into a more automatic control plane for the Scylla clusters that analyses cluster state, such as cache hit rate
   • Increase the redundancy of the scylla-authentication cluster; the options being explored, not yet decided, are a higher replication factor or a hot-standby cluster in another region

   Lessons

   Taking a whole zone out of a three-zone cluster leaves no margin, so one additional slow node in another zone was enough to overload it. For some Scylla clusters the row cache hit rate of restarted nodes must reach a safe level before the next zone goes down, and a short zonal test beforehand gave false confidence. The small but critical authentication cluster needs to be run differently from Discord's much larger Scylla clusters, with more redundancy.

   View case
2. IL-0496 Ably's Cassandra hit 100% CPU in every region and queries timed out globally, and the load returned on two more evenings

   Ably2020

   Symptoms

   Production alarms told Ably that Cassandra queries were timing out for a large fraction of queries globally. Multiple Cassandra instances in each region were pinned at 100% CPU, every production cluster was affected, and the disruption reached Ably's primary realtime messaging service. The first episode started at 21:22.

   Database overloadIntermittent timeouts

   Trigger

   A customer application with a large fleet of set-top box devices was programmed to re-register with Ably at the same time every day. A bug elsewhere in Ably's system meant that the rate limiting applied to every external API request was not applied to this request, and any query that took more than a few seconds was retried.

   Root cause

   Capacity exhaustion

   Component

   Database

   Contributing factor

   Retry stormUnverified assumption

   Detection

   Production alarms reported that Cassandra queries were timing out for a large fraction of queries globally.

   Remediation

   • Applied rate limiting to the request that had been reaching Cassandra without it
   • Redesigned the Cassandra layout and sharding strategy for that data so it no longer used counter columns

   Lessons

   Ably says theory and practice are different things at scale. Counter columns, which in theory should behave like a CRDT with the performance of Cassandra's regular storage, put two orders of magnitude more load on the instances, and further research showed they are not even eventually consistent. Besides the theoretical concerns, a design has to account for load, capacity, what happens when capacity is exceeded and how peer systems react when a service does not respond as expected. Ably's conclusion is not to use Cassandra counter columns.

   View case
3. IL-0333 A downsized RDS PostgreSQL database maxed out its CPU at US peak and failing ELB health checks took buildkite.com down

   Buildkite2016

   Symptoms

   On 22 August 2016 at 17:21 UTC buildkite.com went offline: nobody could log in, view build logs or read the documentation, although builds kept running and GitHub and Bitbucket pull request statuses kept updating. The Buildkite team in Melbourne was asleep and woke at 21:00 UTC, almost 4 hours later, to emails, tweets and Slack messages saying Buildkite was down.

   Service outageDatabase overloadProbe failuresCloud API rate limit exhausted

   Trigger

   Over a weekend maintenance window, Buildkite downgraded its m4.10xlarge Multi-AZ RDS PostgreSQL database to an r3.2xlarge Multi-AZ instance to cut its AWS bill, and moved to a centralised pgbouncer to pool connections. The first US weekday peak after the change pushed the new instance's CPU to its maximum at about 16:00 UTC.

   Root cause

   Capacity exhaustion

   Component

   DatabaseCloud load balancerLiveness and readiness probesNode launch template

   Contributing factor

   Unverified assumptionHealth check tied to a dependencyAutomation amplificationCircular dependencyPages not deliveredOverlapping failures

   Detection

   New Relic, Pingdom, Datadog, Calibre and CloudWatch raised PagerDuty alerts, but the on-call phone call never came and only silent push notifications arrived. The team noticed at 21:00 UTC when they woke to emails, tweets and Slack messages from users.

   Remediation

   • Moved servers from the Agent API to the dashboard ELB to bring buildkite.com back online
   • Switched the Auto Scaling Groups to EC2 health checks and locked in desired counts
   • Upgraded the database to an instance type that could handle the load, using the aws-cli
   • Built a new AMI that did not depend on buildkite.com/_secret/version and rolled out servers that could self-deploy
   • Removed the centralised pgbouncer and went back to pgbouncer on each host

   Action items

   • Plan load testing after significant infrastructure changes without affecting other users of the system
   • Make the ELB health check return OK when the app answers, without checking database connectivity
   • Consider a separate health check endpoint for Pingdom that checks the database and other dependencies
   • Store the latest deployed revision as a file on S3 for newly launched servers to self-deploy from
   • Test by hand that PagerDuty calls reach every on-call team member
   • Investigate regular company-wide tests of everyone's paging setup
   • Ask PagerDuty about calling members when they go on call or requiring an immediate phone rule

   Lessons

   Buildkite listed six lessons. Watch AWS credits, because billing reports show $0 while credits are spent, so the real cost is hidden and big changes end up rushed. Load test after significant infrastructure changes. Think about what a health check should do and what happens when every server behind an ELB fails it. Do not make new servers look up the deployed revision from the service they are replacing. Make sure every on-call member has the right PagerDuty and phone settings before going on call. When servers churn in AWS, stop health checks and autoscaling, for example with the Auto Scaling feature to suspend and resume processes, so the EC2 request limit is not used up.

   View case

By root cause1