FastAPI & Streamlit - Deployment & Interaction
This combination allows scientists to move from a Jupyter Notebook to a production-ready system. FastAPI handles the backend (model serving, data processing), while Streamlit provides the frontend (interactive widgets, real-time plotting).
FIRST: Verify Prerequisites
pip install fastapi uvicorn streamlit pydantic
When to Use
FastAPI:
- Serving Machine Learning models as REST APIs.
- Creating microservices for heavy scientific computations.
- Building backends that require high concurrency (async/await).
- Automatically generating API documentation (Swagger/Redoc).
Streamlit:
- Building interactive dashboards for data exploration.
- Creating "Apps" to demonstrate scientific results to non-technical stakeholders.
- Rapid prototyping of UIs for internal tools.
- Visualizing complex datasets with interactive sliders, maps, and charts.
Reference Documentation
- FastAPI docs: https://fastapi.tiangolo.com/
- Streamlit docs: https://docs.streamlit.io/
- Search patterns:
fastapi.app,pydantic.BaseModel,st.slider,st.cache_data,st.sidebar
Core Principles
FastAPI: Type Safety and Async
FastAPI is built on Pydantic for data validation and Starlette for web capabilities. Every input is validated against Python type hints. It is one of the fastest Python frameworks thanks to async/await.
Streamlit: Execution Model
Streamlit scripts run from top to bottom every time a user interacts with a widget. It uses a "magic" caching system to prevent expensive scientific functions from re-running unnecessarily.
Quick Reference
Installation
pip install fastapi uvicorn streamlit pydantic
Standard Imports
# FastAPI
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
# Streamlit
import streamlit as st
import requests # To communicate with FastAPI
Basic Pattern - FastAPI Model Server
# main_api.py
from fastapi import FastAPI
from pydantic import BaseModel
app = FastAPI()
class ModelInput(BaseModel):
temperature: float
pressure: float
@app.post("/predict")
def predict(data: ModelInput):
# Imagine a complex physical model here
result = data.temperature * 0.5 + data.pressure * 0.2
return {"prediction": result}
# Run with: uvicorn main_api:app --reload
Basic Pattern - Streamlit Dashboard
# main_ui.py
import streamlit as st
import pandas as pd
st.title("Scientific Data Explorer")
# 1. Widgets for input
val = st.slider("Select a threshold", 0.0, 100.0, 50.0)
# 2. Logic/Processing
df = pd.DataFrame({"x": range(100), "y": [x**2 for x in range(100)]})
filtered_df = df[df["y"] > val]
# 3. Visualization
st.line_chart(filtered_df)
st.write(f"Points above threshold: {len(filtered_df)}")
# Run with: streamlit run main_ui.py
Critical Rules
✅ DO
- Use Pydantic Schemas (FastAPI) - Always define your API inputs and outputs using classes inheriting from
BaseModel. - Use st.cache_data (Streamlit) - Wrap heavy data loading or heavy math functions with
@st.cache_datato keep the UI responsive. - Use Async/Await (FastAPI) - For I/O bound tasks (database, API calls), use
async defto maximize throughput. - Set Page Config (Streamlit) - Use
st.set_page_config(layout="wide")for scientific dashboards that need space for plots. - Handle Exceptions - Use FastAPI's
HTTPExceptionto return clear error codes (400, 404, 500) to the user. - Modularize - Keep your scientific logic in a separate file/package, imported by both API and UI.
❌ DON'T
- Don't Run Heavy Logic in UI Thread - In Streamlit, if a function takes >1s, it must be cached or the UI will feel broken.
- Don't Block the Async Loop (FastAPI) - If a function is CPU-intensive (e.g., heavy NumPy math), use standard
definstead ofasync def; FastAPI will run it in a thread pool. - Don't Store Sensitive Data in UI Code - Use environment variables or
.streamlit/secrets.toml. - Don't Over-nest Widgets - Streamlit's "top-down" execution gets confusing if the UI logic is too complex.
Anti-Patterns (NEVER)
# ❌ BAD: Manual JSON parsing in FastAPI
# @app.post("/data")
# def handle_data(raw_json: dict):
# val = raw_json.get("value") # No validation!
# ✅ GOOD: Pydantic validation
class DataPoint(BaseModel):
value: float
@app.post("/data")
def handle_data(data: DataPoint):
return data.value # Guaranteed to be a float
# ❌ BAD: Loading data in every Streamlit rerun
# data = pd.read_csv("massive_data.csv") # Re-reads every time you move a slider!
# ✅ GOOD: Caching
@st.cache_data
def load_massive_data():
return pd.read_csv("massive_data.csv")
data = load_massive_data()
FastAPI: Advanced Features
Dependency Injection (e.g., Database/Model loading)
from functools import lru_cache
@lru_cache()
def load_model():
# Load your PyTorch or Scikit-learn model here
return MyHeavyModel().load("weights.pt")
@app.get("/status")
def get_status(model = Depends(load_model)):
return {"model_version": model.version}
Background Tasks (Long-running computations)
from fastapi import BackgroundTasks
def solve_pde_task(params):
# Long FEniCS simulation
pass
@app.post("/run-sim")
def run_simulation(params: Params, background_tasks: BackgroundTasks):
background_tasks.add_task(solve_pde_task, params)
return {"message": "Simulation started in background"}
Streamlit: Layout and Interaction
Multi-column and Sidebars
st.sidebar.header("Settings")
mode = st.sidebar.selectbox("Model Mode", ["Fast", "Accurate"])
col1, col2 = st.columns(2)
with col1:
st.header("Input Parameters")
temp = st.number_input("Temperature (K)")
with col2:
st.header("Results Visualization")
# Plotly/Matplotlib chart
st.plotly_chart(fig)
Session State (Keeping track of user data)
if 'results_history' not in st.session_state:
st.session_state.results_history = []
if st.button("Run Experiment"):
res = run_model()
st.session_state.results_history.append(res)
st.write(f"History length: {len(st.session_state.results_history)}")
Practical Workflows
1. Scientific Model Serving (FastAPI + PyTorch)
import torch
from fastapi import FastAPI
from pydantic import BaseModel
app = FastAPI()
model = torch.load("model.pth")
model.eval()
class PredictionRequest(BaseModel):
features: list[float]
@app.post("/v1/predict")
def get_prediction(req: PredictionRequest):
input_tensor = torch.tensor([req.features])
with torch.no_grad():
output = model(input_tensor)
return {"class": output.argmax().item(), "confidence": output.max().item()}
2. Interactive Data Cleaning Tool (Streamlit + Polars)
import streamlit as st
import polars as pl
st.title("Data Cleaner")
uploaded_file = st.file_uploader("Choose a CSV file")
if uploaded_file:
df = pl.read_csv(uploaded_file)
st.write("Original Data Summary", df.describe())
col_to_drop = st.multiselect("Drop columns", df.columns)
if st.button("Clean Data"):
df_clean = df.drop(col_to_drop).drop_nulls()
st.dataframe(df_clean)
st.download_button("Download Clean CSV", df_clean.write_csv(), "clean.csv")
3. Real-time Monitoring App
import streamlit as st
import time
placeholder = st.empty()
for i in range(100):
with placeholder.container():
st.metric("Current Sensor Reading", f"{get_val()} units")
st.progress(i + 1)
time.sleep(1)
Performance Optimization
1. FastAPI: Uvicorn Workers
For production, run with multiple workers to handle more requests.
uvicorn main:app --workers 4
2. Streamlit: st.cache_resource
Use cache_resource for objects that should stay in memory across users/sessions, like Database connections or ML models.
@st.cache_resource
def get_database_connection():
return create_engine("postgresql://...")
3. Streamlit: PyArrow
Streamlit uses Apache Arrow for data exchange. Ensuring your data is in Arrow-compatible formats (like Polars or Pandas) makes UI rendering instant.
Common Pitfalls and Solutions
FastAPI: Input Data Validation Errors
If a user sends a string where a float is expected, FastAPI returns 422 Unprocessable Entity.
# ✅ Solution: Wrap Pydantic models in try-except if needed,
# but usually, let FastAPI handle it and customize exception_handlers.
Streamlit: The "Double Rerun"
Sometimes widgets trigger multiple reruns.
# ✅ Solution: Use st.form to group widgets so the script
# only reruns once when the "Submit" button is clicked.
with st.form("my_form"):
# ... inputs ...
submitted = st.form_submit_button("Submit")
Deployment Port Conflict
By default, Streamlit uses 8501 and FastAPI (Uvicorn) uses 8000.
# ✅ Solution: Be explicit in Docker/Compose files about ports.
Best Practices
- Separate Concerns - Keep scientific logic separate from API/UI code for reusability
- Type Everything - Use Pydantic models for all FastAPI endpoints to catch errors early
- Cache Aggressively - In Streamlit, cache any computation that takes >100ms
- Use Async Wisely - FastAPI async is great for I/O, but CPU-bound tasks should be sync
- Test Both Separately - Test your FastAPI endpoints with
httpxorrequests, test Streamlit UI manually - Document APIs - FastAPI auto-generates docs, but add docstrings to your Pydantic models
- Handle Errors Gracefully - Both frameworks have good error handling; use it
- Monitor Performance - Use FastAPI's built-in metrics and Streamlit's execution time display
The FastAPI + Streamlit stack is the "Last Mile" of scientific computing. It transforms raw code into accessible tools, making your models useful to the rest of the world.