Cases4 cases

1 to 4 of 4

1. IL-0316 Railway's control plane slowed and deployments stopped after a PgBouncer version jump used abnormal CPU and memory

   Railway2025

   Symptoms

   At 11:48 UTC on 22 September 2025, minutes after Railway had resolved a dashboard outage, its control plane showed abnormal resource consumption and elevated API latency. A large backlog of deployments queued during the first outage had started all at once, and the control plane began returning elevated database errors.

   High latencyIntermittent timeoutsDeploys failing

   Trigger

   After a dashboard outage caused by a schema change was fixed, a large backlog of queued deployments started at the same time and put the control plane and its database, behind a newly upgraded PgBouncer, under heavy load.

   Root cause

   Breaking change in an upgrade

   Component

   DatabaseDatabase proxy and connection pooler

   Contributing factor

   Unpinned versionNo staged rolloutMonitoring blind spotUnreviewed defaults

   Detection

   Users reported at 11:05 UTC that the project canvas failed to load; the control plane's abnormal resource consumption and elevated API latency were seen at 11:48 UTC and opened as a new incident.

   Remediation

   • Deployed a fix for the API failure caused by the schema change within 15 minutes
   • Suspended deployments for non-Pro users so Pro users could deploy and the queue could drain
   • Suspended all deployments, Pro included, for up to 20 minutes during peak instability
   • Scaled up the control plane database's CPU, memory, connection limits and IOPS
   • Rolled PgBouncer back to the previous stable version, which brought all metrics back to nominal
   • Re-enabled deployments for all non-Pro users
   • Fixed the edge case in which private networks and their endpoints were missing from the dashboard

   Action items

   • Re-engineer parts of the control plane to be more resilient, moving large parts of its storage layer away from Postgres
   • Add CI validation that keeps disruptive schema changes out of production and automated schema compatibility checks before backend deployments
   • Monitor and alert on PgBouncer node CPU, memory and connection pool utilization
   • Upgrade PgBouncer in phases, instance by instance, instead of everywhere at once

   Lessons

   Railway believes that without the breaking schema change reaching production the cascade would not have happened, that monitoring PgBouncer's own CPU, memory and pool metrics would have shown the performance regression earlier and shortened recovery, and that upgrading PgBouncer instances one at a time would have exposed the version-specific problem before it reached every instance.

   View case
2. IL-0498 Clerk's Postgres database had load spikes every 15 minutes and requests failed and slowed intermittently for four days

   Clerk2025

   Symptoms

   At 13:08 UTC on 15 September 2025, Clerk's internal monitoring alerted on significant request failures and increased latency. Over the following days the request failures and latency spikes kept coming back while the Postgres database stayed unhealthy, with repeated spikes in database load.

   High error rateHigh latencyDatabase overloadConnection pool exhaustion

   Trigger

   At 05:30 UTC on 14 September 2025 the cloud provider automatically upgraded Clerk's Postgres database to a new minor version, which changed how Postgres handled connection locks.

   Root cause

   Breaking change in an upgrade

   Component

   DatabaseDatabase proxy and connection pooler

   Contributing factor

   Unpinned versionRetry stormMonitoring blind spotOverlapping failuresHarmful manual action

   Detection

   Internal monitoring alerted on significant request failures and increased latency at 13:08 UTC on 15 September. Shortly after the upgrade on 14 September it had already shown an increased but still acceptable database load.

   Remediation

   • Optimised several expensive queries, refined critical indexes and re-indexed the database
   • Blocked abusive traffic more aggressively for a time
   • Performed a manual minor version upgrade and increased database capacity
   • Added jitter to the connection pooling configuration so connections no longer recycled in step

   Action items

   • Evaluate other database providers that give more control over upgrades and downgrades
   • Isolate more services from each other, for example Sign Up and Sign In infrastructure from the rest of the system
   • Keep optimising the most costly database queries
   • Evaluate sharding and multi-tenant infrastructure to isolate one application from others

   Lessons

   Clerk's review found that connection metrics sampled every 60 seconds could not show a recycling pattern that lasted seconds and repeated every 15 minutes, that an attack and background jobs with similar periods pulled attention away from the database layer, that the Backend API looked healthy only because it manages far fewer connections, and that the release freeze removed the staggering of connection creation that normal deploys provide.

   View case
3. IL-0312 Gadget's main Postgres did not start again after a restart to raise max_connections, and the platform went down

   Gadget2024

   Symptoms

   At 11:25 AM EST on 13 December 2024, Gadget restarted its main PostgreSQL server in place to raise max_connections. Postgres shut down but did not start again, and every request to Gadget applications failed, including frontend requests, GraphQL API requests, webhook processing and syncs.

   Service outageHigh latency

   Trigger

   An emergency in-place restart of the main Postgres server at 11:25 AM EST, decided at 11:00 AM, to raise max_connections, which Postgres cannot change without a full restart.

   Root cause

   Capacity exhaustion

   Component

   DatabaseLinux kernelDatabase proxy and connection pooler

   Contributing factor

   Unverified assumptionUnrepresentative stagingAutomation amplificationExcessive blast radius

   Detection

   The team carrying out the planned restart saw at 11:26 AM EST that Postgres had shut down and did not start again.

   Remediation

   • Restarted the database server with a hardware reset, which cleared the fragmented memory so Postgres could start
   • Provisioned almost 4x the compute after the autoscalers had scaled it down during the outage

   Action items

   • Use only the replica-failover approach for disruptive database operations such as restarts
   • Run a more aggressive capacity planning process before Black Friday / Cyber Monday, preparing for a 6x increase in load and proving it with load tests well in advance
   • Announce planned downtime ahead of time on the status page and community Discord, and make the operating procedures tell planned maintenance apart from service disruptions
   • Finish the database sharding project so apps can use different databases and configuration changes can go through canary and blue/green setups
   • Finish building a zero-downtime database failover system
   • Set much higher scaling floors on the compute autoscalers

   Lessons

   Gadget names three deeper factors. Choosing the in-place restart put the database into a state the team had not seen before and was not confident in. Capacity planning for Black Friday / Cyber Monday assumed a doubling of traffic instead of the almost 5x that came, so settings such as max_connections stopped fitting and had to be changed in an emergency instead of in a maintenance window practised in staging. And there was no way to test database configuration changes in a production-like environment: integration tests run against a synthetic Postgres database that never goes through the months of production load behind the memory fragmentation or the memory usage issues.

   View case
4. 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 cause2