APM Server version (apm-server version): 8.2+

Description of the problem including expected versus actual behavior:

When Elasticsearch ingestion is significantly lower than the APM processing rate, and given enough CPU processing power for APM Server to decode 200 concurrent intake requests, the APM Server may run out of memory in some cases.

Steps to reproduce:

1. Create a CPU Optimized ESS deployment with a small backing Elasticsearch (16gb) and a 1GB APM Server.
2. Run apmbench against it
3. Watch it run out of memory.