GitHub: dbos_experiments/exp20

Experiment 20: DBOS Queue Concurrency Performance

Overview

This experiment measures DBOS queue performance with different concurrency settings. It enqueues multiple workflows and tracks execution time to understand how queue concurrency parameters affect throughput.

Queue Concurrency Parameters

queue = Queue("my_queue", concurrency=10, worker_concurrency=5)

• concurrency: Global maximum across all DBOS processes (default: unlimited)
• worker_concurrency: Per-process maximum (default: unlimited)

The effective limit is the minimum of these two values when running a single process.

Key Metrics

The experiment tracks:

• Enqueue time: How long it takes to submit all workflows to the queue
• Total execution time: Real wall-clock time from start to completion
• Workflow delays: Cumulative time spent in workflow steps
• Completion order: Which workflows finish first (using asyncio.as_completed)

Running the Experiment

python exp20/main.py

Adjusting Parameters

To test different configurations, modify:

# Number of workflows to enqueue
n_workflows = 10

# Queue concurrency limits
queue = Queue("my_queue", concurrency=10, worker_concurrency=5)

# Per-workflow step count and delay
total_delay = sum([await fetch_url(fake.url()) for _ in range(5)])  # 5 steps
delay = 0.2  # 0.2s per step in fetch_url

Performance Considerations

Optimal Concurrency Settings

1. CPU-bound workflows: Set worker_concurrency ≈ number of CPU cores
2. I/O-bound workflows: Higher values (10-50+) for better throughput
3. Rate-limited APIs: Match concurrency to API limits to avoid throttling
4. Resource constraints: Lower values to control memory/connection usage

Expected Behavior

With n_workflows=10, worker_concurrency=5, and 1 second per workflow:

• First 5 workflows execute immediately
• Remaining 5 wait in queue
• Second batch executes as first batch completes
• Total time: ~2 seconds (2 batches × 1 second)

Taking Advantage of DBOS Queues

Benefits:

• Automatic recovery: Enqueued workflows survive process restarts
• Persistent state: Queue state stored in database
• Controlled resource usage: Prevent overwhelming downstream services
• Fair scheduling: FIFO order with optional priority
• Distributed execution: Global concurrency limits across multiple processes

Best Practices:

1. Use enqueue_async for non-blocking submission
2. Process results with asyncio.as_completed for real-time feedback
3. Set worker_concurrency based on bottleneck (CPU, I/O, or external API)
4. Monitor queue depth and adjust concurrency if workflows pile up
5. Use multiple queues to separate different types of work

Database Requirement

Requires PostgreSQL connection. Set in config:

config: DBOSConfig = {
    "name": "dbos-starter",
    "database_url": "postgresql://user:pass@localhost:5432/dbname",
}

Example Output

All Workflows Enqueued: total_workflows=10, real_time_enqueue_elapsed=0.05s
Workflow Result (as completed): instance_id=3, total_delay=1.0s
Workflow Result (as completed): instance_id=1, total_delay=1.0s
...
All Workflows Completed: real_time_elapsed=2.1s, total_main_delay=10.0s

The real_time_elapsed shows wall-clock time while total_main_delay shows cumulative workflow time, demonstrating the parallelization benefit.

Recent changes

• 2025-11-30

  9485293

  Add exp20: DBOS queue concurrency performance experiment