kim/litellm-config
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: finalize smart pool tracking with hysteresis (handle jumps, drops, and expiry)

Browse Source
This commit is contained in:
kim
2026-01-29 23:17:39 +08:00
parent 2d849dc62d
commit cd35a193b6
1 changed files with 43 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
volc_limit_sync.py
+43 -19
View File
@@ -999,39 +999,63 @@ def active_monitor_loop():
dynamic_key = f"limit:dynamic:volcengine:remaining:{target}"
# ==========================================
# 🏔️ Peak Tracking & Stable Limit
# 🏔️ Smart Peak Tracking (Auto-Correction)
# ==========================================
# Logic Change: Track "Total Daily Budget" instead of just "Remaining".
# Total Budget = Current Usage + Current Remaining.
# This handles mid-day restarts correctly.
# Strategy:
# 1. Capture Jumps (Recovery): If Balance jumps UP, trust it immediately. (Fixes latency)
# 2. Capture Drops (Expiry): If Balance drops HUGE (>20%), trust it. (Fixes expiry)
# - Why trust drops? Because if it's a false drop (glitch), Step 1 will fix it
# automatically on the next sync when it "Jumps" back up.
total_pool = current_usage + balance
peak_key = f"volc:peak_pool:{today_str}:{target}"
current_peak = float(redis_client.get(peak_key) or 0)
if total_pool > current_peak:
# Short-term memory (Last seen value)
last_pool_key = f"volc:last_seen_pool:{target}"
last_pool_val = redis_client.get(last_pool_key)
last_pool = float(last_pool_val) if last_pool_val else 0.0
# Long-term memory (Day's High Water Mark)
old_peak_val = redis_client.get(peak_key)
historical_peak = float(old_peak_val) if old_peak_val else 0.0
current_peak = historical_peak
# 1. JUMP DETECTION (Recharge / Latency Resolved)
# If pool jumps significantly (>5%) from last seen, it's a fresh, true value.
if total_pool > (last_pool * 1.05) and last_pool > 0:
if DEBUG_MODE:
print(
f"DEBUG: New Peak Pool for {target}: {int(total_pool):,} (Usg:{int(current_usage):,} + Rem:{int(balance):,})",
f"DEBUG: 📈 Balance Jump ({int(last_pool):,} -> {int(total_pool):,}) - Locking New Peak.",
flush=True,
)
redis_client.setex(
peak_key, 172800, int(total_pool)
) # 48h retention
current_peak = total_pool
redis_client.setex(peak_key, 172800, int(current_peak))
# Fallback: If peak is somehow less than balance (shouldn't happen with logic above), fix it
if current_peak < balance:
current_peak = balance
# 2. DROP DETECTION (Resource Expiry)
# If pool drops MASSIVELY (>20%) below historical peak, it's likely an expiry.
# We accept this drop. If it was just a glitch, Rule #1 will fix it next time it bounces back.
elif (
total_pool < (historical_peak * 0.8) and historical_peak > 0
):
if DEBUG_MODE:
print(
f"DEBUG: 📉 Expiry/Drop Detected ({int(historical_peak):,} -> {int(total_pool):,}) - Adjusting Limit Down.",
flush=True,
)
current_peak = total_pool
redis_client.setex(peak_key, 172800, int(current_peak))
# 3. INITIALIZATION / DRIFT
elif historical_peak == 0 or total_pool > historical_peak:
current_peak = total_pool
redis_client.setex(peak_key, 172800, int(current_peak))
# Always update "Last Seen" for next comparison
redis_client.setex(last_pool_key, 86400, int(total_pool))
# Calculate Stable Limit
# Limit = Peak Pool - Buffer
# Buffer = Max(Peak * 15%, 250,000)
safety_margin = max(current_peak * 0.15, 250000)
# The Limit is the "Max Allowed Cumulative Usage"
# Max Usage = Total Pool - Buffer
stable_limit = int(max(0, current_peak - safety_margin))
stable_limit_key = (