Cases33 cases

11 to 15 of 33

1. IL-0526 OpenStreetMap.org went offline, then read-only, for 56 hours after HE.net routing equipment cut both fibre uplinks

   OpenStreetMap Foundation2024

   Symptoms

   At 04:01 UTC on 15 December 2024, StatusCake monitoring sent SMS alerts that OpenStreetMap.org was unreachable. The website and the mapping API had gone completely offline at 03:53 UTC.

   Network unreachableService outage

   Trigger

   HE.net's routing equipment in Amsterdam failed at 03:53 UTC on 15 December 2024.

   Root cause

   Dependency failure

   Component

   ISP and transit networksDatabase replication

   Contributing factor

   Single point of failureNo runbook

   Detection

   StatusCake monitoring detected the outage at 04:01 UTC and sent SMS alerts, and the incident was reported to the operations team at once.

   Remediation

   • Activated a read-only backup instance of OpenStreetMap.org in Dublin
   • Brought a new Equinix Internet link into service in Amsterdam, bypassing HE.net and restoring full functionality
   • HE.net replaced its routing hardware and restored its service, and routing through HE.net resumed

   Action items

   • Move to a multi-ISP architecture so another provider can take over when one fails
   • Research backup ISPs for the Amsterdam servers, including BGP and keeping HE.net as a fallback
   • Add deeper route and latency monitoring and an automated escalation protocol
   • Hold failover drills for switching between ISPs and keep replication up to date
   • Agree SLAs with HE.net and Equinix Internet with defined escalation procedures and response times
   • Prepare the 4G out-of-band system as a manual fallback uplink that can sync Postgres replication data, and upgrade its hardware
   • Document outage communication procedures and delegate communication during outages
   • Investigate decoupling the OAuth service so authentication keeps working while the database is read-only

   Lessons

   Relying on HE.net alone left the Amsterdam servers with a single point of failure, so OpenStreetMap will keep multiple upstream providers. The 4G out-of-band link proved invaluable and should be able to carry data synchronisation. Resumed replication would have restored editing faster than the read-only backup. Frequent contact with the ISP gave no useful estimates, so the SLA needs to be better understood. Alerting on upstream connectivity needs refining, and a read-only OpenStreetMap.org also blocked logins to services that were not affected, because it could not save new tokens.

   View case
2. IL-0392 MariaDB replication of section x1 between eqiad and codfw broke on a missing heartbeat row during a cookbook run

   Wikimedia Foundation2024

   Symptoms

   On 18 September 2024, while a cookbook was reconfiguring Wikimedia's core MariaDB databases from one-way eqiad -> codfw replication to circular replication ahead of a datacenter switchover, replication in the codfw -> eqiad direction broke on the x1 and es6 sections. Engineers watching the run noticed at 21:28 UTC that the x1 primary was not replicating.

   Replication lag

   Trigger

   Running the cookbook sre.switchdc.databases.prepare, which set up codfw -> eqiad replication and started pt-heartbeat, on the row-based x1 and es sections.

   Root cause

   Configuration change

   Component

   DatabaseDatabase replication

   Contributing factor

   Monitoring blind spotGap in automated testsHarmful manual action

   Detection

   Engineers watching the first run of the automation saw on orchestrator that the x1 primary was not replicating, because monitoring of circular replication was not yet enabled. The break in the eqiad -> codfw direction was reported by Icinga.

   Remediation

   • Skipped replication of the heartbeat tables with Replication_wild_ignore_table='heartbeat.%'
   • Inserted the missing heartbeat row with binary logging off on the es hosts
   • Inserted the missing heartbeat row with binary logging off on the x1 codfw hosts and removed the ignore rule

   Action items

   • Avoid the same issue again, probably by inserting the missing row when it is detected
   • Add a post-maintenance health check that replication works in both directions
   • Add options to run the maintenance in another order or on each section individually
   • Update the Phabricator task for each section rather than only at the end
   • Fix the monitoring gaps for circular replication
   • Made pt-heartbeat-wikimedia use STATEMENT format for its session on the x1 and row-based read-write es hosts, fixed in a later deploy

   Lessons

   Circular replication is not a normal working mode, so the first fix only moved the break to the other side; TRUNCATE would have worked better than DELETE because it always replicates as a statement, but it was not considered because it is more dangerous. Running sql_skip_slave_counter was avoided, since it could have made things worse or lost data. The check of the replication channel should also run once all tasks are complete, not only right after setup. The combination of row-based replication, pt-heartbeat and differing heartbeat rows was not described in earlier maintenance or documentation.

   View case
3. IL-0374 DigitalOcean creates and management actions failed globally when its service discovery datastore lost quorum

   DigitalOcean2024

   Symptoms

   From 16:30 UTC on 5 August 2024, DigitalOcean's internal services could not discover the other services they depended on. Creates, event processing and the management of DigitalOcean products failed in all regions through the Cloud Control Panel and the API, while existing customer resources such as Droplets stayed online.

   Service outageCluster or node provisioning stall

   Trigger

   During the rollout of a new datastore cluster in a new datacenter, a replication configuration was deployed against the datastore behind internal service discovery that gave the new datacenter, which held zero keys, 100% ownership of all keys.

   Root cause

   Configuration change

   Component

   Coordination and key-value storeService discoveryDatabase replication

   Contributing factor

   Excessive blast radius

   Detection

   Service discovery unavailability was first reported internally at 16:35 UTC, five minutes after the rollout of the new datastore cluster began.

   Remediation

   • Reverted the replication configuration and ran it against the region that had become the datastore owner
   • Applied the configuration by hand on the nodes of lagging regions and ran local data repairs
   • Cleaned up failed events and processed pending events that had not timed out

   Action items

   • Perform this kind of datastore extension inside a declared maintenance window, announced on the status page
   • Update the process documentation for this type of deployment, stage by stage
   • Automate the manual steps used in this deployment to reduce human error
   • Evaluate the internal datastore topology and whether to regionalise or layer it

   View case
4. IL-0488 Kit's RDS MySQL write latency rose 100x and running threads jumped from 9 to 900, backing up Sidekiq job queues

   Kit2022

   Symptoms

   At 5:20 am EST on December 24, 2022, Kit's infrastructure team got its first page because the number of MySQL threads running on its RDS MySQL primary in us-east-2 had jumped 100x, from 9 to 900, at a time of day when Kit's traffic is very low. Sidekiq queues were backing up.

   Database overloadThroughput degradation

   Trigger

   An EBS interruption on the standby of the Multi-AZ RDS instance, reported by AWS as an LSE at 02:12:47 UTC on 2022-12-24, started a replaceSecondaryHostAndVolumesTask that replaced the secondary host and volumes.

   Root cause

   Cloud provider outage

   Component

   DatabaseCloud block storageDatabase replication

   Contributing factor

   Ignored early signal

   Detection

   A page fired at 5:20 am EST because MySQL threads running jumped from 9 to 900. An AWS RSS feed alert in Slack about degraded EC2 and EBS in use2-az1 had arrived the evening before and was not acted on.

   Remediation

   • Scaled the RDS storage from 12.5 TB to 13.75 TB, which put the instance through "optimizing storage" and brought write latency back to 1-2 ms

   Lessons

   Multi-AZ RDS keeps data consistent and durable, but synchronous replication means each write also waits for the standby, so a degraded EBS volume on the standby while it is being replaced slows writes on the primary. Large databases recover more slowly from such events than small ones. A single-AZ instance would not have been hit by a standby problem, but would have suffered hours of downtime if its own zone failed, and that trade-off has to be weighed when deploying production RDS instances.

   View case
5. IL-0308 Most RevenueCat backend systems failed after cutover to a new Aurora PostgreSQL 14 cluster as queries became very slow

   RevenueCat2022

   Symptoms

   At 9:34 UTC on November 23, 2022, RevenueCat started moving traffic from its Aurora PostgreSQL 10.x cluster to a new Aurora PostgreSQL 14.x cluster. By 9:39 UTC all traffic was on the new cluster, latency was bad, and the RevenueCat dashboard and the endpoints used by the RevenueCat SDK stopped working.

   Database overloadHigh latencyService outageConnection pool exhaustion

   Trigger

   At 9:34 UTC production traffic was switched from the Aurora PostgreSQL 10.x cluster to the new Aurora PostgreSQL 14.x cluster, before ANALYZE had run on its largest tables.

   Root cause

   Breaking change in an upgrade

   Component

   DatabasePostgreSQL vacuumDatabase replicationDatabase proxy and connection poolerQuery planner and table statistics

   Contributing factor

   Unrepresentative stagingUnverified assumptionAutomation amplificationMonitoring blind spot

   Detection

   At 9:39 UTC, with all traffic on the new cluster, latency looked bad and the infrastructure team began to investigate.

   Remediation

   • Added a statement timeout to the database user so queries could not hold a connection too long
   • Shed 90% of traffic by returning 503s from the load balancer
   • Dropped indexes that the planner was using wrongly and that were not needed
   • Stopped accepting Apple webhooks for a time
   • Rolled out a release without New Relic to stop its EXPLAIN queries
   • Ran ANALYZE by hand on the biggest tables
   • Raised the values of the affected primary key sequences with a prepared script
   • Stopped the integration event dispatcher by hand
   • Replayed the webhooks that were dispatched but not recorded
   • Restored integration event dispatching, ETLs and other non-web server jobs

   Action items

   • Consider change freezes during business-critical periods
   • Explain better how the SDK behaves during backend outages
   • Tell customers about maintenance and outages in the developer dashboard and by direct email to technical contacts
   • Replay a subset of write traffic in migration dry runs to mimic the cutover
   • Move costly processes to asynchronous work to keep the API fast and stable
   • Re-evaluate table indexes, including unused ones the database may choose for queries
   • Show the load of each endpoint and system in the database dashboards
   • Involve support and customer communications staff during outages

   Lessons

   The new cluster had not been analyzed, and the tests missed it because the test cluster had been up long enough to gather statistics. Dry runs replayed no write traffic, on the belief that writes are simpler than reads, and the team assumed sequence state was replicated with the data. Rolling back at the start would have been better than fixing forward. The migration, timed to finish before Thanksgiving and Black Friday, should have been delayed until the week after, because the US had entered a sensitive period and part of the US support team was already on holiday.

   View case

By root cause11