Stability Improvements: OOMKill Prevention and Stress Relief Tuning

[VinozzZ]

Problem

Refinery experiences high memory pressure during peak traffic periods, leading to pods being OOMKilled or remaining stuck in Stress Relief mode for extended durations. This issue has become more prominent since the upgrade to version 2.9

Ideas

• Improve Redistribution Strategy During Scale-Up Events #1532
• Improve HealthCheck Signaling Under High Load #1531
• feat: Tell Load Balancers to Back Off #1530 (comment)
• Improve eviction strategies when memory usage nears MaxAlloc
• Tune Stress Relief mode calcualtion
• Backpressure controls on ingestion to reduce burst impact

[VinozzZ]

We have made significant improvements on Refinery's memory usage through unmarshaling only defined sampling key fields. The improvements will be included in the upcoming 3.0 release. Closing this issue.