Cases7 cases

4 to 7 of 7

1. IL-0295 Pusher's Channels publish API on the mt1 cluster was largely unavailable for 11 minutes while a Redis master failed over

   Pusher2022

   Symptoms

   At 10:42 UTC on 11 April 2022, the API for publishing messages to Pusher Channels on the mt1 cluster became largely unavailable, and from about 10:44 requests to the MT1 API returned more errors.

   Service outageHigh latencySplit brain

   Trigger

   An engineer restarted the master node of an mt1 Redis group to apply a larger replication buffer, after restarting both replicas of that group.

   Root cause

   Software defect

   Component

   In-memory cacheDatabase failover manager

   Contributing factor

   Unverified assumptionUnrepresentative staging

   Detection

   Pusher saw error rates rise for requests to the MT1 API from about 10:44 UTC.

   Action items

   • Fix the Redis shutdown script to check that failover has completed and the node is no longer a master before terminating

   View case
2. IL-0213 Roblox was down for 73 hours because key-value writes on its single Consul cluster slowed from under 300 ms to 2 seconds

   Roblox2021

   Symptoms

   On the afternoon of 28 October 2021 (13:37 PST), Vault performance at Roblox degraded and a single Consul server showed high CPU load. Consul cluster metrics showed that writes to its key-value (KV) store had a 50th percentile latency of 2 seconds instead of the usual under 300 ms. Players were not yet affected.

   High latencySustained high node CPUService outageMonitoring outage

   Trigger

   On 27 October at 14:00, one day before the outage, Roblox enabled the Consul streaming feature on the backend service responsible for traffic routing and, to prepare for end-of-year traffic, added 50% more nodes to traffic routing.

   Root cause

   Software defect

   Component

   Coordination and key-value storeService discoveryIn-memory cache

   Contributing factor

   Single point of failureOverlapping failuresMonitoring blind spotHarmful manual actionChange with a delayed effect

   Detection

   Roblox engineers began investigating when Vault performance degraded and a single Consul server had high CPU load, before players were affected. Consul cluster metrics then showed elevated KV write latency.

   Remediation

   • Replaced one Consul cluster node on the suspicion of degraded hardware
   • Replaced all Consul nodes with 128-core machines with faster NVMe SSDs
   • Shut down the Consul cluster and reset its state from a snapshot taken at the start of the outage, with access blocked by iptables
   • Blocked external traffic, disabled all non-essential Consul use, scaled services down and lowered health check frequency from 60 seconds to 10 minutes
   • Moved the Consul cluster back to 64-core servers
   • Disabled the streaming feature for all Consul systems, which brought KV write latency back to 300ms
   • Kept the slow Consul servers from staying elected as leader
   • Redeployed the caching system and fixed its wrong scheduling data and failed deployments
   • Restarted the stopped services at the right capacity levels
   • Let players back in with DNS steering in roughly 10% steps while checking database load, cache performance and stability

   Action items

   • HashiCorp builds laboratory benchmarks and scale tests that reproduce the streaming contention
   • HashiCorp improves the design of streaming so it does not contend under extreme load
   • Compact the BoltDB database with existing BoltDB tooling
   • Remove the circular dependency between the telemetry systems and Consul
   • Extend telemetry with targeted alerts on Consul and BoltDB performance and with visibility into traffic between services and Consul
   • Run the backend in an additional, geographically distinct data center and in multiple availability zones
   • Split critical services into their own dedicated Consul clusters
   • Move data that core services keep directly in the Consul KV store to a more suitable storage system
   • Delete obsolete KV data from Consul
   • Upgrade to a Consul version that replaces BoltDB with bbolt, in Q1
   • Redesign cache deployment so the cache system can be brought up quickly from a standing start
   • Deploy Nomad enhancements that make it easier to turn up large jobs after a long period of unavailability
   • Use mechanisms for faster machine configuration changes
   • Reintroduce streaming only after a new implementation has been tested at Roblox's scale and workload

   Lessons

   Two mostly unrelated problems deep inside Consul, contention in streaming under very high read and write load and BoltDB freelist growth, were hard to diagnose, and that difficulty was the main reason the outage lasted so long. A single Consul cluster serving every workload made the impact larger, and monitoring that depended on Consul hid the cause when it was needed most. Faster hardware did not help and may have hurt, because the contention was worse on NUMA machines. HashiCorp had benchmarked streaming at a similar scale, but the problem needed both a large number of streams and a high churn rate. Restarting a platform from a fully stopped state, including warming caches, took a long time and needed tools built for bootstrapping rather than incremental change.

   View case
3. IL-0279 One failed Redis Cluster replica left only about 1 in 21 calls to Grab's Apollo service succeeding for a minute

   Grab2019

   Symptoms

   On 15 February 2019 a slave node of the AWS ElastiCache Redis Cluster behind Apollo, Grab's driver-side state machine for transport bookings, failed and had to be replaced. For about one minute only roughly 1 in 21 calls to Apollo succeeded, a call failure rate above 95% against the Redis Cluster, and Grab rides were significantly down.

   High error rateStale endpoint routing

   Trigger

   A slave node of Apollo's Redis Cluster failed and needed to be replaced.

   Root cause

   Software defect

   Component

   In-memory cache

   Contributing factor

   Unverified assumption

   Lessons

   Having several replicas did not protect Apollo, because the client side, not the Redis Cluster server, caused the failure. Grab suggests three ways to avoid it: a separate Redis Cluster client used only for pipelining with RouteByLatency set to true, so pipelined reads go through the master nodes when latency is under 1ms, watching master CPU because they take more load; checking the error of each command in a pipeline instead of failing the whole call, so successful queries are not affected; and changing the Go-Redis library so that the pipeline path reloads the nodes' state when an error happens, as defaultProcessor already does.

   View case
4. IL-0388 A Google Cloud live migration took a Discord Redis primary offline, and the failover broke API instances and then guilds

   Discord2017

   Symptoms

   At about 14:01 PDT on 13 October 2017, Discord engineers were alerted to anomalies by real-time metrics and monitoring. A Redis instance that was the primary of a highly available cluster used by the API services had dropped offline after Google Cloud Platform migrated it automatically, and some API instances did not handle the failover to the secondary node correctly.

   Failover failureHigh latencyService outage

   Trigger

   Google Cloud Platform automatically migrated the Redis instance that was the primary of a highly available cluster used by the API services, and the node incorrectly dropped offline.

   Root cause

   Software defect

   Component

   In-memory cacheDatabaseCDN and edge network

   Contributing factor

   Unpatched known bugOverlapping failuresMonitoring blind spot

   Detection

   Real-time metrics and monitoring alerted engineers at 14:01 PDT. The second phase was seen at 15:41 as another anomaly in monitoring, and end user reports of servers appearing offline were escalated at 15:42.

   Remediation

   • Ran a rolling restart of all API instances to clear the Redis connection issues
   • Corrected the misconfigured edge caching rule for the expensive API route
   • Disabled the problem API endpoint on the edge network
   • Recovered the Cassandra cluster
   • Ran a second rolling restart of all API instances for the latency left by the Cassandra issues
   • Removed the misbehaving guilds node from the service hashring
   • Forcefully restarted the misbehaving guilds node
   • Restarted the whole service, forcing all users to reconnect
   • Blocked traffic that relied on the failing push component to reduce its load
   • Corrected the API setting that lets users reconnect quickly and safely

   Action items

   • Fix the known issues with Redis failover within the coming weeks
   • Change the cached route so this type of error does not degrade other services
   • Monitor the signals that would have warned of the cascading failure
   • Improve the reliability and failure handling of the affected services
   • Roll out an update of the library with the suspected flaws

   View case

By root cause1