Performance regression: v5 is 2–7× slower than v3.8.0 for local evaluation

[gagantrivedi]

Benchmarked flagsmith 5.0.1 vs 3.8.0 against the same environment (edge API, enable_local_evaluation=True, 262 features, single-threaded, after 200-iter warmup).

p50 per call

Operation	v3.8.0	v5.0.1	Regression
get_identity_flags	0.39 ms	0.82 ms	2.1×
get_environment_flags	0.11 ms	0.79 ms	7.4×
is_feature_enabled (cached Flags)	0.08 µs	0.08 µs	—

Throughput drops from ~2,570 → ~1,215 identity evals/sec per core. Latency variance also grew (stdev ~10 µs → ~220 µs).

Repro

from time import perf_counter
from flagsmith import Flagsmith

client = Flagsmith(
    environment_key="<server-side-key>",
    api_url="https://edge.api.flagsmith.com/api/v1/",
    enable_local_evaluation=True,
    environment_refresh_interval_seconds=3600,
)
client.update_environment()

# warmup
for _ in range(200):
    client.get_identity_flags(identifier="anonymous", traits={"venue_id": "12345"})

for n in (100, 1_000, 10_000):
    t0 = perf_counter()
    for _ in range(n):
        client.get_identity_flags(identifier="anonymous", traits={"venue_id": "12345"})
    elapsed = perf_counter() - t0
    print(f"{n:>6} iters: {elapsed*1000:8.1f} ms ({elapsed/n*1e6:7.2f} µs/call)")

Environment: flag-engine 10.0.1, Python 3.12, macOS.

[khvn26]

We've released #200 to mitigate this, but we still need to address the issue described here. The rough plan is to profile the repro benchmark to find the exact bottleneck in evaluation. Keep in mind we need to add .all_flags() now.