58 lines
1.5 KiB
Markdown
58 lines
1.5 KiB
Markdown
---
|
|
title: Use Streams for Real-Time Data
|
|
impact: MEDIUM
|
|
impactDescription: Enables real-time progress and LLM streaming
|
|
tags: streaming, write_stream, read_stream, realtime
|
|
---
|
|
|
|
## Use Streams for Real-Time Data
|
|
|
|
Workflows can stream data in real-time to clients. Useful for LLM responses, progress reporting, or long-running results.
|
|
|
|
**Incorrect (returning all data at end):**
|
|
|
|
```python
|
|
@DBOS.workflow()
|
|
def llm_workflow(prompt):
|
|
# Client waits for entire response
|
|
response = call_llm(prompt)
|
|
return response
|
|
```
|
|
|
|
**Correct (streaming results):**
|
|
|
|
```python
|
|
@DBOS.workflow()
|
|
def llm_workflow(prompt):
|
|
for chunk in call_llm_streaming(prompt):
|
|
DBOS.write_stream("response", chunk)
|
|
DBOS.close_stream("response")
|
|
return "complete"
|
|
|
|
# Client reads stream
|
|
@app.get("/stream/{workflow_id}")
|
|
def stream_response(workflow_id: str):
|
|
def generate():
|
|
for value in DBOS.read_stream(workflow_id, "response"):
|
|
yield value
|
|
return StreamingResponse(generate())
|
|
```
|
|
|
|
Stream characteristics:
|
|
- Streams are immutable and append-only
|
|
- Writes from workflows happen exactly-once
|
|
- Writes from steps happen at-least-once (may duplicate on retry)
|
|
- Streams auto-close when workflow terminates
|
|
|
|
Close streams explicitly when done:
|
|
|
|
```python
|
|
@DBOS.workflow()
|
|
def producer():
|
|
DBOS.write_stream("data", {"step": 1})
|
|
DBOS.write_stream("data", {"step": 2})
|
|
DBOS.close_stream("data") # Signal completion
|
|
```
|
|
|
|
Reference: [Workflow Streaming](https://docs.dbos.dev/python/tutorials/workflow-communication#workflow-streaming)
|