Reduce Feature Flags rule evaluation allocations

[leoromanovsky]

Motivation

Regex rules compiled and normalized the expression during every evaluation. JDK 11 JMH measured 341 ns and 2,256 bytes per operation.

Shard rules created an MD5 helper, a combined input string, and a hexadecimal string. JDK 11 JMH measured 345 ns and 1,456 bytes per operation.

The static control measured 80 ns and 368 bytes per operation.

Changes and Decisions

• Compile valid regex expressions during UFC preprocessing.
• Keep invalid expressions uncached so evaluation preserves PARSE_ERROR behavior.
• Cache immutable shard salt-prefix bytes during UFC preprocessing.
• Reuse one MD5 instance per evaluation thread without shared mutation.
• Read the first four digest bytes directly and preserve the Eppo shard algorithm.
• Create no scheduler or worker thread.
• Keep a JMH benchmark for static, regex, and shard rules.

Validation

JDK 11 single-thread JMH results:

• Regex: 341 to 122 ns and 2,256 to 568 bytes per operation.
• Shard: 345 to 232 ns and 1,456 to 448 bytes per operation.
• Static control: 80 to 85 ns and 368 bytes per operation.

Four-thread JDK 26 JMH results:

• Regex: 251 to 66 ns and 2,032 to 504 bytes per operation.
• Shard: 286 to 125 ns and 1,344 to 448 bytes per operation.
• Static control: 43 to 44 ns and 368 bytes per operation.

The shard parity test preserves the Go and Eppo vectors for eve, user-1, alice, and bob.

[datadog-datadog-prod-us1]

🎯 Code Coverage (details)
• Patch Coverage: 44.90%
• Overall Coverage: 77.52% (+19.17%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 43ba527 | Docs | Datadog PR Page | Give us feedback!}