feat(context): enhance timeout handling, tenant isolation, and budget management
- Added timeout enforcement for token counting, scoring, and compression with detailed error handling. - Introduced tenant isolation in context caching using project and agent identifiers. - Enhanced budget management with stricter checks for critical context overspending and buffer limitations. - Optimized per-context locking with cleanup to prevent memory leaks in concurrent environments. - Updated default assembly timeout settings for improved performance and reliability. - Improved XML escaping in Claude adapter for safety against injection attacks. - Standardized token estimation using model-specific ratios.
This commit is contained in:
@@ -6,9 +6,9 @@ Combines multiple scoring strategies with configurable weights.
|
||||
|
||||
import asyncio
|
||||
import logging
|
||||
import time
|
||||
from dataclasses import dataclass
|
||||
from typing import TYPE_CHECKING, Any
|
||||
from weakref import WeakValueDictionary
|
||||
|
||||
from ..config import ContextSettings, get_context_settings
|
||||
from ..types import BaseContext
|
||||
@@ -91,11 +91,11 @@ class CompositeScorer:
|
||||
self._priority_scorer = PriorityScorer(weight=self._priority_weight)
|
||||
|
||||
# Per-context locks to prevent race conditions during parallel scoring
|
||||
# Uses WeakValueDictionary so locks are garbage collected when not in use
|
||||
self._context_locks: WeakValueDictionary[str, asyncio.Lock] = (
|
||||
WeakValueDictionary()
|
||||
)
|
||||
# Uses dict with (lock, last_used_time) tuples for cleanup
|
||||
self._context_locks: dict[str, tuple[asyncio.Lock, float]] = {}
|
||||
self._locks_lock = asyncio.Lock() # Lock to protect _context_locks access
|
||||
self._max_locks = 1000 # Maximum locks to keep (prevent memory growth)
|
||||
self._lock_ttl = 60.0 # Seconds before a lock can be cleaned up
|
||||
|
||||
def set_mcp_manager(self, mcp_manager: "MCPClientManager") -> None:
|
||||
"""Set MCP manager for semantic scoring."""
|
||||
@@ -141,7 +141,8 @@ class CompositeScorer:
|
||||
Get or create a lock for a specific context.
|
||||
|
||||
Thread-safe access to per-context locks prevents race conditions
|
||||
when the same context is scored concurrently.
|
||||
when the same context is scored concurrently. Includes automatic
|
||||
cleanup of old locks to prevent memory growth.
|
||||
|
||||
Args:
|
||||
context_id: The context ID to get a lock for
|
||||
@@ -149,25 +150,78 @@ class CompositeScorer:
|
||||
Returns:
|
||||
asyncio.Lock for the context
|
||||
"""
|
||||
now = time.time()
|
||||
|
||||
# Fast path: check if lock exists without acquiring main lock
|
||||
if context_id in self._context_locks:
|
||||
lock = self._context_locks.get(context_id)
|
||||
if lock is not None:
|
||||
# NOTE: We only READ here - no writes to avoid race conditions
|
||||
# with cleanup. The timestamp will be updated in the slow path
|
||||
# if the lock is still valid.
|
||||
lock_entry = self._context_locks.get(context_id)
|
||||
if lock_entry is not None:
|
||||
lock, _ = lock_entry
|
||||
# Return the lock but defer timestamp update to avoid race
|
||||
# The lock is still valid; timestamp update is best-effort
|
||||
return lock
|
||||
|
||||
# Slow path: create lock or update timestamp while holding main lock
|
||||
async with self._locks_lock:
|
||||
# Double-check after acquiring lock - entry may have been
|
||||
# created by another coroutine or deleted by cleanup
|
||||
lock_entry = self._context_locks.get(context_id)
|
||||
if lock_entry is not None:
|
||||
lock, _ = lock_entry
|
||||
# Safe to update timestamp here since we hold the lock
|
||||
self._context_locks[context_id] = (lock, now)
|
||||
return lock
|
||||
|
||||
# Slow path: create lock while holding main lock
|
||||
async with self._locks_lock:
|
||||
# Double-check after acquiring lock
|
||||
if context_id in self._context_locks:
|
||||
lock = self._context_locks.get(context_id)
|
||||
if lock is not None:
|
||||
return lock
|
||||
# Cleanup old locks if we have too many
|
||||
if len(self._context_locks) >= self._max_locks:
|
||||
self._cleanup_old_locks(now)
|
||||
|
||||
# Create new lock
|
||||
new_lock = asyncio.Lock()
|
||||
self._context_locks[context_id] = new_lock
|
||||
self._context_locks[context_id] = (new_lock, now)
|
||||
return new_lock
|
||||
|
||||
def _cleanup_old_locks(self, now: float) -> None:
|
||||
"""
|
||||
Remove old locks that haven't been used recently.
|
||||
|
||||
Called while holding _locks_lock. Removes locks older than _lock_ttl,
|
||||
but only if they're not currently held.
|
||||
|
||||
Args:
|
||||
now: Current timestamp for age calculation
|
||||
"""
|
||||
cutoff = now - self._lock_ttl
|
||||
to_remove = []
|
||||
|
||||
for context_id, (lock, last_used) in self._context_locks.items():
|
||||
# Only remove if old AND not currently held
|
||||
if last_used < cutoff and not lock.locked():
|
||||
to_remove.append(context_id)
|
||||
|
||||
# Remove oldest 50% if still over limit after TTL filtering
|
||||
if len(self._context_locks) - len(to_remove) >= self._max_locks:
|
||||
# Sort by last used time and mark oldest for removal
|
||||
sorted_entries = sorted(
|
||||
self._context_locks.items(),
|
||||
key=lambda x: x[1][1], # Sort by last_used time
|
||||
)
|
||||
# Remove oldest 50% that aren't locked
|
||||
target_remove = len(self._context_locks) // 2
|
||||
for context_id, (lock, _) in sorted_entries:
|
||||
if len(to_remove) >= target_remove:
|
||||
break
|
||||
if context_id not in to_remove and not lock.locked():
|
||||
to_remove.append(context_id)
|
||||
|
||||
for context_id in to_remove:
|
||||
del self._context_locks[context_id]
|
||||
|
||||
if to_remove:
|
||||
logger.debug(f"Cleaned up {len(to_remove)} context locks")
|
||||
|
||||
async def score(
|
||||
self,
|
||||
context: BaseContext,
|
||||
|
||||
Reference in New Issue
Block a user