Implement streaming pipeline endpoint for batch processing

- Added `/run/streaming-pipeline` endpoint to process data in batches from either a database or CSV file. - Introduced `BatchResponse` model for structured responses. - Updated README with new endpoint details, including parameters and example usage. - Enhanced error handling and logging during batch processing. - Ensured data preprocessing and NaN handling in analysis functions.
2025-05-02 14:25:31 +01:00
parent 5767f55686
commit 9c429caa56
10 changed files with 246 additions and 11 deletions
@@ -119,6 +119,32 @@ uvicorn salary_analytics.api:app --reload

 6. **Pipeline**
   - `POST /run/pipeline`: Run complete pipeline
+   - `POST /run/streaming-pipeline`: Run pipeline in batches
+     - Parameters:
+       - `source`: Data source ('db' or 'csv')
+       - `file`: CSV file (required if source is 'csv')
+       - `batch_size`: Number of rows to process in each batch (default: 10000)
+     - Example:
+       ```bash
+       # Run streaming pipeline from database
+       curl -X POST "http://localhost:8000/run/streaming-pipeline?source=db&batch_size=5000"
+       
+       # Run streaming pipeline from CSV
+       curl -X POST "http://localhost:8000/run/streaming-pipeline?source=csv&batch_size=5000" -F "file=@path/to/your/file.csv"
+       ```
+     - Response:
+       ```json
+       [
+         {
+           "batch_number": 1,
+           "total_batches": 10,
+           "processed_rows": 5000,
+           "results_path": "/app/output/csv/batch_results_20240315_123456/batch_1_results.csv",
+           "message": "Successfully processed batch 1 of 10"
+         },
+         // ... more batch responses ...
+       ]
+       ```

 ### Workflow

@@ -127,6 +153,12 @@ uvicorn salary_analytics.api:app --reload
 3. Run any of the analysis endpoints
 4. Generate and download reports as needed

+For large datasets, use the streaming pipeline endpoint:
+1. Start the API server
+2. Run the streaming pipeline with appropriate batch size
+3. Monitor batch processing progress
+4. Access results in the batch results directory
+
 Note: All analysis endpoints require data to be loaded first. If you try to run any analysis without loading data, you'll receive a 400 error with a message to load data first.

 ## Docker Deployment
@@ -6,15 +6,16 @@ from fastapi import FastAPI, HTTPException, BackgroundTasks, UploadFile, File
 from fastapi.responses import FileResponse
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
-from typing import Optional, Dict
+from typing import Optional, Dict, List
 import os
 import socket
 import logging
 import pandas as pd
 import tempfile
-
+from datetime import datetime
+from sqlalchemy import text
 from .main import SalaryAnalyticsPipeline
-from .config import OUTPUT_PATHS
+from .config import OUTPUT_PATHS, TABLE_NAME
 from .data_loader import DataLoader
 from .salary_predictor import SalaryPredictor
 from .salary_earner_analyzer import SalaryEarnerAnalyzer
@@ -56,6 +57,14 @@ class AnalysisResponse(BaseModel):
    data: Optional[Dict] = None
    file_path: Optional[str] = None

+class BatchResponse(BaseModel):
+    """Response model for batch processing."""
+    batch_number: int
+    total_batches: int
+    processed_rows: int
+    results_path: str
+    message: str
+
 def check_data_loaded():
    """Check if data is loaded before running analytics."""
    if pipeline.df is None:
@@ -278,3 +287,171 @@ async def load_data(source: str = "db", file: UploadFile = None):
    except Exception as e:
        logger.error(f"Error loading data: {str(e)}")
        raise HTTPException(status_code=500, detail=str(e))
+
+@app.post("/run/streaming-pipeline", response_model=List[BatchResponse])
+async def run_streaming_pipeline(source: str = "db", file: UploadFile = None, batch_size: int = 10000):
+    """
+    Run the complete salary analytics pipeline in batches.
+    
+    Args:
+        source (str): Source of data ('db' or 'csv')
+        file (UploadFile): CSV file to load (required if source is 'csv')
+        batch_size (int): Number of rows to process in each batch
+    
+    Returns:
+        List[BatchResponse]: List of responses for each batch processed
+    """
+    try:
+        if source not in ['db', 'csv']:
+            raise HTTPException(status_code=400, detail="Source must be either 'db' or 'csv'")
+        
+        if source == 'csv' and not file:
+            raise HTTPException(status_code=400, detail="File must be provided when loading from CSV")
+        
+        # Initialize data loader
+        data_loader = DataLoader()
+        data_loader.chunk_size = batch_size
+        
+        # Create output directory for batch results
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        batch_output_dir = os.path.join(os.path.dirname(OUTPUT_PATHS['final_table']), f"batch_results_{timestamp}")
+        os.makedirs(batch_output_dir, exist_ok=True)
+        
+        responses = []
+        batch_number = 0
+        
+        def preprocess_chunk(chunk):
+            """Preprocess a chunk of data with the same logic as DataLoader."""
+            # Convert dates
+            chunk['trx_start_date'] = pd.to_datetime(chunk['trx_start_date'])
+            chunk['trx_end_date'] = pd.to_datetime(chunk['trx_end_date'])
+            
+            chunk = chunk.rename(columns={
+                    'd1': 'trx_type',
+                    'd2': 'trx_subtype',
+                    'd3': 'initiated_by',
+                    'd4': 'customer_id'
+                    })
+            chunk = chunk.dropna()
+            
+            return chunk
+        
+        if source == 'csv':
+            # Save uploaded file temporarily
+            with tempfile.NamedTemporaryFile(delete=False, suffix='.csv') as temp_file:
+                content = await file.read()
+                temp_file.write(content)
+                temp_file_path = temp_file.name
+            
+            try:
+                # Process CSV in chunks
+                for chunk in pd.read_csv(temp_file_path, chunksize=batch_size):
+                    batch_number += 1
+                    logger.info(f"Processing batch {batch_number}")
+                    
+                    # Preprocess chunk
+                    chunk = preprocess_chunk(chunk)
+                    
+                    # Run pipeline on chunk
+                    pipeline = SalaryAnalyticsPipeline()
+                    pipeline.df = chunk
+                    
+                    try:
+                        # Run analyses
+                        pipeline.run_keyword_analysis()
+                        pipeline.run_consistent_amount_analysis()
+                        pipeline.run_transaction_type_analysis()
+                        
+                        # Generate reports
+                        reports = pipeline.generate_salary_earner_reports()
+                        
+                        # Save batch results
+                        batch_results_path = os.path.join(batch_output_dir, f"batch_{batch_number}_results.csv")
+                        reports['final_table'].to_csv(batch_results_path, index=False)
+                        
+                        responses.append(BatchResponse(
+                            batch_number=batch_number,
+                            total_batches=-1,  # Unknown for CSV
+                            processed_rows=len(chunk),
+                            results_path=batch_results_path,
+                            message=f"Successfully processed batch {batch_number}"
+                        ))
+                    except Exception as e:
+                        logger.error(f"Error processing batch {batch_number}: {str(e)}")
+                        responses.append(BatchResponse(
+                            batch_number=batch_number,
+                            total_batches=-1,
+                            processed_rows=len(chunk),
+                            results_path="",
+                            message=f"Error processing batch {batch_number}: {str(e)}"
+                        ))
+            finally:
+                # Clean up temporary file
+                os.unlink(temp_file_path)
+        else:
+            # Process database in chunks
+            if not data_loader.connect():
+                raise HTTPException(status_code=500, detail="Failed to connect to database")
+            
+            # Get total row count
+            with data_loader.engine.connect() as conn:
+                count_query = text(f"SELECT COUNT(*) FROM {TABLE_NAME}")
+                total_rows = conn.execute(count_query).scalar()
+            
+            total_batches = (total_rows + batch_size - 1) // batch_size
+            offset = 0
+            
+            while offset < total_rows:
+                batch_number += 1
+                logger.info(f"Processing batch {batch_number} of {total_batches}")
+                
+                # Load chunk from database
+                query = f"SELECT * FROM {TABLE_NAME} LIMIT {batch_size} OFFSET {offset}"
+                chunk = pd.read_sql(query, data_loader.engine)
+                
+                if chunk.empty:
+                    break
+                
+                # Preprocess chunk
+                chunk = preprocess_chunk(chunk)
+                
+                # Run pipeline on chunk
+                pipeline = SalaryAnalyticsPipeline()
+                pipeline.df = chunk
+                
+                try:
+                    # Run analyses
+                    pipeline.run_keyword_analysis()
+                    pipeline.run_consistent_amount_analysis()
+                    pipeline.run_transaction_type_analysis()
+                    
+                    # Generate reports
+                    reports = pipeline.generate_salary_earner_reports()
+                    
+                    # Save batch results
+                    batch_results_path = os.path.join(batch_output_dir, f"batch_{batch_number}_results.csv")
+                    reports['final_table'].to_csv(batch_results_path, index=False)
+                    
+                    responses.append(BatchResponse(
+                        batch_number=batch_number,
+                        total_batches=total_batches,
+                        processed_rows=len(chunk),
+                        results_path=batch_results_path,
+                        message=f"Successfully processed batch {batch_number} of {total_batches}"
+                    ))
+                except Exception as e:
+                    logger.error(f"Error processing batch {batch_number}: {str(e)}")
+                    responses.append(BatchResponse(
+                        batch_number=batch_number,
+                        total_batches=total_batches,
+                        processed_rows=len(chunk),
+                        results_path="",
+                        message=f"Error processing batch {batch_number}: {str(e)}"
+                    ))
+                
+                offset += batch_size
+        
+        return responses
+    except Exception as e:
+        logger.error(f"Error in streaming pipeline: {str(e)}")
+        raise HTTPException(status_code=500, detail=str(e)) 
@@ -69,6 +69,7 @@ class DataLoader:
                        'd4': 'customer_id'
                    })
                
+                chunk = chunk.dropna()
                chunks.append(chunk)
            
            # Combine all chunks
@@ -127,6 +128,7 @@ class DataLoader:
                    'd4': 'customer_id'
                })
                
+                chunk = chunk.dropna()
                chunks.append(chunk)
                offset += self.chunk_size
                
@@ -18,7 +18,7 @@ class SalaryEarnerAnalyzer:
    def filter_venn_section(self, **kwargs):
        """Filter accounts based on specified combinations of hypothesis flags."""
        valid_columns = {'is_salary_related', 'is_consistent_amount', 'is_salary_type'}
-        df1 = self.df[self.df['initiated_by'] == 'C']
+        df1 = self.df[self.df['initiated_by'] == 'C'].copy()
        
        invalid_keys = set(kwargs.keys()) - valid_columns
        if invalid_keys:
@@ -28,7 +28,13 @@ class SalaryEarnerAnalyzer:
        for key, value in kwargs.items():
            condition &= (df1[key] == value)

-        return df1[condition]
+        filtered_df = df1[condition]
+        
+        # Drop any rows with NaN values in critical columns
+        critical_cols = ['accountid', 'trx_start_date', 'amount']
+        filtered_df = filtered_df.dropna(subset=critical_cols)
+        
+        return filtered_df

    def plot_hypothesis_overlap(self, hypothesis1_df, hypothesis3_df, hypothesis4_df, account_col='accountid'):
        """Plot and save Venn diagram showing overlap between hypotheses."""
@@ -47,21 +53,37 @@ class SalaryEarnerAnalyzer:
        """Generate a table of salary earners with their metrics."""
        results = []
        for accountid, group in all_three_hypotheses.groupby('accountid'):
+            # Skip if group is empty
+            if group.empty:
+                continue
+                
            # Calculate required metrics
            num_months = len(group)
-            last_6_months = group[group['trx_start_date'] >= (datetime.now() - timedelta(days=180))]
-            least_inflow = last_6_months['amount'].min()
-            avg_salary = group['amount'].mean()
            
-            # Calculate days since last transaction
+            # Handle last 6 months calculation
+            last_6_months = group[group['trx_start_date'] >= (datetime.now() - timedelta(days=180))]
+            if last_6_months.empty:
+                least_inflow = 0
+            else:
+                least_inflow = last_6_months['amount'].min()
+            
+            # Handle average salary calculation
+            if group['amount'].notna().any():
+                avg_salary = group['amount'].mean()
+            else:
+                avg_salary = 0
+            
+            # Calculate days_since_last_trx with NaN handling
            group['days_since_last_trx'] = group['trx_start_date'].diff().dt.days
            median_interval = group['days_since_last_trx'].median()
+            if pd.isna(median_interval):
+                median_interval = 30  # Default to 30 days if no interval data

            last_date = group['trx_start_date'].max()
            next_date = last_date + timedelta(days=median_interval)
            next_amount = avg_salary

-            # Boolean flags
+            # Boolean flags with NaN handling
            days_since_last = (datetime.now() - last_date).days
            has_45d = days_since_last <= 45
            has_2m = len(group[group['trx_start_date'] >= (datetime.now() - timedelta(days=60))]) >= 2
@@ -78,7 +100,9 @@ class SalaryEarnerAnalyzer:
            })

        final_df = pd.DataFrame(results)
-        final_df = final_df.dropna()
+        # Drop rows where all numeric columns are NaN
+        numeric_cols = ['num_months', 'least_inflow_6m', 'avg_monthly_salary', 'estimated_next_amount']
+        final_df = final_df.dropna(subset=numeric_cols, how='all')
        return final_df

    def analyze_salary_earners(self, final_df):