Pinterest encountered a significant performance issue during the migration of its search infrastructure, Manas, to Kubernetes, where one in a million search requests experienced latency spikes. The investigation revealed that cAdvisor’s memory monitoring processes were causing excessive contention, leading to these delays. The team resolved the issue by disabling a specific metric in cAdvisor, allowing them to continue their migration efforts without compromising performance.

By implementing a php-fpm-exporter in a Kubernetes environment, the author identified severe underutilization of PHP-FPM processes due to a misconfigured shared configuration file. After analyzing the traffic patterns and adjusting the PHP-FPM settings accordingly, memory utilization was reduced by over 80% without sacrificing performance. The article emphasizes the importance of customizing configurations based on specific application needs rather than relying on default settings.