TLDR Tech / Data
Drastically Reducing Out-of-Memory Errors in Apache Spark at Pinterest
Pinterest developed the Auto Memory Retries feature, which automatically retries OOM-failed tasks first by increasing CPU allocation to reduce contention, then launching bigger executors with 2x/3x/4x resource profiles scaling memory, overhead, and off-heap for Gluten jobs. This resulted in a 96% reduction in OOM failures, compute cost savings, and reduced manual tuning efforts on common culprits like severe data skew or unpredictable per-task memory spikes.
Read the full story