Bound feature flag evaluation aggregation memory

[leoromanovsky]

Motivation

The current count limits do not bound retained heap when full evaluation data is enabled. JOL measured a typical 256-field bucket at 60,021 bytes. It measured a maximum legal bucket at 297,589 bytes. At the global count limit, these profiles project to 7.3 GiB and 36.3 GiB. This memory can remain until the 10-second flush.

Changes and Decisions

• Enforce a conservative 64 MiB estimated byte budget across full and degraded aggregation buckets.
• Merge existing buckets without a new byte reservation.
• Route a new full bucket to degraded aggregation when the byte budget is full.
• Drop the evaluation only when a degraded bucket cannot fit.
• Keep the existing count limits as secondary guards.
• Report byte-budget degradation and drops separately from cardinality limits.

Validation

JOL measured the retained object graph after the first byte-budget degradation:

• Typical 256-field context: 735 full buckets and one degraded bucket retained 44,110,800 bytes.
• Maximum 256-field context: 118 full buckets and one degraded bucket retained 35,115,816 bytes.
• The estimator reserved 67,108,848 bytes and 66,840,328 bytes for these profiles.
• Smaller profiles reached the per-flag limit first. Protected data retained 2,466,400 bytes. A typical 10-field context retained 26,866,360 bytes.

[datadog-datadog-prod-us1]

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 79.58% (+21.23%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 21b8641 | Docs | Datadog PR Page | Give us feedback!}