mirror of https://github.com/langgenius/dify.git
docs: remove outdated document
This commit is contained in:
-LAN-
parent
103a9a4e67
commit
2adf5d0eee
|
|
@ -1,173 +0,0 @@
|
|||
# GraphEngine Worker Pool Configuration
|
||||
|
||||
## Overview
|
||||
|
||||
The GraphEngine now supports **dynamic worker pool management** to optimize performance and resource usage. Instead of a fixed 10-worker pool, the engine can:
|
||||
|
||||
1. **Start with optimal worker count** based on graph complexity
|
||||
1. **Scale up** when workload increases
|
||||
1. **Scale down** when workers are idle
|
||||
1. **Respect configurable min/max limits**
|
||||
|
||||
## Benefits
|
||||
|
||||
- **Resource Efficiency**: Uses fewer workers for simple sequential workflows
|
||||
- **Better Performance**: Scales up for parallel-heavy workflows
|
||||
- **Gevent Optimization**: Works efficiently with Gevent's greenlet model
|
||||
- **Memory Savings**: Reduces memory footprint for simple workflows
|
||||
|
||||
## Configuration
|
||||
|
||||
### Configuration Variables (via dify_config)
|
||||
|
||||
| Variable | Default | Description |
|
||||
|----------|---------|-------------|
|
||||
| `GRAPH_ENGINE_MIN_WORKERS` | 1 | Minimum number of workers per engine |
|
||||
| `GRAPH_ENGINE_MAX_WORKERS` | 10 | Maximum number of workers per engine |
|
||||
| `GRAPH_ENGINE_SCALE_UP_THRESHOLD` | 3 | Queue depth that triggers scale up |
|
||||
| `GRAPH_ENGINE_SCALE_DOWN_IDLE_TIME` | 5.0 | Seconds of idle time before scaling down |
|
||||
|
||||
### Example Configurations
|
||||
|
||||
#### Low-Resource Environment
|
||||
|
||||
```bash
|
||||
export GRAPH_ENGINE_MIN_WORKERS=1
|
||||
export GRAPH_ENGINE_MAX_WORKERS=3
|
||||
export GRAPH_ENGINE_SCALE_UP_THRESHOLD=2
|
||||
export GRAPH_ENGINE_SCALE_DOWN_IDLE_TIME=3.0
|
||||
```
|
||||
|
||||
#### High-Performance Environment
|
||||
|
||||
```bash
|
||||
export GRAPH_ENGINE_MIN_WORKERS=2
|
||||
export GRAPH_ENGINE_MAX_WORKERS=20
|
||||
export GRAPH_ENGINE_SCALE_UP_THRESHOLD=5
|
||||
export GRAPH_ENGINE_SCALE_DOWN_IDLE_TIME=10.0
|
||||
```
|
||||
|
||||
#### Default (Balanced)
|
||||
|
||||
```bash
|
||||
# Uses defaults: min=1, max=10, threshold=3, idle_time=5.0
|
||||
```
|
||||
|
||||
## How It Works
|
||||
|
||||
### Initial Worker Calculation
|
||||
|
||||
The engine analyzes the graph structure at startup:
|
||||
|
||||
- **Sequential graphs** (no branches): 1 worker
|
||||
- **Limited parallelism** (few branches): 2 workers
|
||||
- **Moderate parallelism**: 3 workers
|
||||
- **High parallelism** (many branches): 5 workers
|
||||
|
||||
### Dynamic Scaling
|
||||
|
||||
During execution:
|
||||
|
||||
1. **Scale Up** triggers when:
|
||||
|
||||
- Queue depth exceeds `SCALE_UP_THRESHOLD`
|
||||
- All workers are busy and queue has items
|
||||
- Not at `MAX_WORKERS` limit
|
||||
|
||||
1. **Scale Down** triggers when:
|
||||
|
||||
- Worker idle for more than `SCALE_DOWN_IDLE_TIME` seconds
|
||||
- Above `MIN_WORKERS` limit
|
||||
|
||||
### Gevent Compatibility
|
||||
|
||||
Since Gevent patches threading to use greenlets:
|
||||
|
||||
- Workers are lightweight coroutines, not OS threads
|
||||
- Dynamic scaling has minimal overhead
|
||||
- Can efficiently handle many concurrent workers
|
||||
|
||||
## Migration Guide
|
||||
|
||||
### Before (Fixed 10 Workers)
|
||||
|
||||
```python
|
||||
# Every GraphEngine instance created 10 workers
|
||||
# Resource waste for simple workflows
|
||||
# No adaptation to workload
|
||||
```
|
||||
|
||||
### After (Dynamic Workers)
|
||||
|
||||
```python
|
||||
# GraphEngine creates 1-5 initial workers based on graph
|
||||
# Scales up/down based on workload
|
||||
# Configurable via environment variables
|
||||
```
|
||||
|
||||
### Backward Compatibility
|
||||
|
||||
The default configuration (`max=10`) maintains compatibility with existing deployments. To get the old behavior exactly:
|
||||
|
||||
```bash
|
||||
export GRAPH_ENGINE_MIN_WORKERS=10
|
||||
export GRAPH_ENGINE_MAX_WORKERS=10
|
||||
```
|
||||
|
||||
## Performance Impact
|
||||
|
||||
### Memory Usage
|
||||
|
||||
- **Simple workflows**: ~80% reduction (1 vs 10 workers)
|
||||
- **Complex workflows**: Similar or slightly better
|
||||
|
||||
### Execution Time
|
||||
|
||||
- **Sequential workflows**: No change
|
||||
- **Parallel workflows**: Improved with proper scaling
|
||||
- **Bursty workloads**: Better adaptation
|
||||
|
||||
### Example Metrics
|
||||
|
||||
| Workflow Type | Old (10 workers) | New (Dynamic) | Improvement |
|
||||
|--------------|------------------|---------------|-------------|
|
||||
| Sequential | 10 workers idle | 1 worker active | 90% fewer workers |
|
||||
| 3-way parallel | 7 workers idle | 3 workers active | 70% fewer workers |
|
||||
| Heavy parallel | 10 workers busy | 10+ workers (scales up) | Better throughput |
|
||||
|
||||
## Monitoring
|
||||
|
||||
Log messages indicate scaling activity:
|
||||
|
||||
```shell
|
||||
INFO: GraphEngine initialized with 2 workers (min: 1, max: 10)
|
||||
INFO: Scaled up workers: 2 -> 3 (queue_depth: 4)
|
||||
INFO: Scaled down workers: 3 -> 2 (removed 1 idle workers)
|
||||
```
|
||||
|
||||
## Best Practices
|
||||
|
||||
1. **Start with defaults** - They work well for most cases
|
||||
1. **Monitor queue depth** - Adjust `SCALE_UP_THRESHOLD` if queues back up
|
||||
1. **Consider workload patterns**:
|
||||
- Bursty: Lower `SCALE_DOWN_IDLE_TIME`
|
||||
- Steady: Higher `SCALE_DOWN_IDLE_TIME`
|
||||
1. **Test with your workloads** - Measure and tune
|
||||
|
||||
## Troubleshooting
|
||||
|
||||
### Workers not scaling up
|
||||
|
||||
- Check `GRAPH_ENGINE_MAX_WORKERS` limit
|
||||
- Verify queue depth exceeds threshold
|
||||
- Check logs for scaling messages
|
||||
|
||||
### Workers scaling down too quickly
|
||||
|
||||
- Increase `GRAPH_ENGINE_SCALE_DOWN_IDLE_TIME`
|
||||
- Consider workload patterns
|
||||
|
||||
### Out of memory
|
||||
|
||||
- Reduce `GRAPH_ENGINE_MAX_WORKERS`
|
||||
- Check for memory leaks in nodes
|
||||
Loading…
Reference in New Issue