batch-processing（批处理）

Batch Processing

30-40% throughput improvement by batching database operations with graceful fallback.

When to Use This Skill

• Processing multiple related records (invoices, orders, events)
• Network latency is significant (cloud databases)
• Writes are independent (no inter-record dependencies)
• You can implement fallback for reliability

Core Concepts

Sequential processing is slow because each item requires multiple DB round trips. The solution is to collect all data first, then execute batch operations:

Sequential (slow):
Item 1 → DB → DB → DB
Item 2 → DB → DB → DB
Item 3 → DB → DB → DB

Batched (fast):
Item 1 → collect
Item 2 → collect
Item 3 → collect
All items → BATCH INSERT

Key insight: Sequential mapping (fuzzy matching needs context), but batched writes (independent operations).

Implementation

Python

from decimal import Decimal
from typing import Dict, List
import time

class BatchProcessor:
    """
    Batch-optimized processor with fallback
    """

    def process_batch(self, items: List[Dict], user_id: str) -> Dict:
        start_time = time.perf_counter()

        # Collectors for batch operations
        transactions_to_create = []
        inventory_updates = {}
        failed_items = []
        items_processed = 0

        # Step 1: Process mappings sequentially (context-dependent)
        for idx, item in enumerate(items, 1):
            try:
                # Business logic that needs context
                mapping = self.find_or_create_mapping(item)

                # Collect for batch insert
                transactions_to_create.append({
                    "user_id": user_id,
                    "item_id": mapping['item_id'],
                    "quantity": float(item['quantity']),
                    "unit_cost": float(item['unit_price']),
                })

                # Aggregate inventory updates by item
                item_id = mapping['item_id']
                if item_id not in inventory_updates:
                    inventory_updates[item_id] = Decimal('0')
                inventory_updates[item_id] += Decimal(str(item['quantity']))

                items_processed += 1

            except Exception as e:
                failed_items.append({
                    "line": idx,
                    "error": str(e)
                })
                continue

        # Step 2: BATCH INSERT transactions
        if transactions_to_create:
            try:
                self.client.table("transactions").insert(
                    transactions_to_create
                ).execute()
            except Exception as e:
                # CRITICAL: Fallback to sequential on batch failure
                return self._fallback_to_sequential(items, user_id)

        # Step 3: BATCH UPDATE inventory (aggregate first)
        if inventory_updates:
            self._batch_update_inventory(inventory_updates)

        return {
            "status": "partial_success" if failed_items else "success",
            "items_processed": items_processed,
            "items_failed": len(failed_items),
            "failed_items": failed_items or None,
            "processing_time_seconds": round(time.perf_counter() - start_time, 2)
        }

    def _batch_update_inventory(self, updates: Dict[str, Decimal]):
        """Batch query, individual updates (Supabase limitation)"""
        item_ids = list(updates.keys())

        # Get current quantities in one query
        current = self.client.table("inventory").select(
            "id, quantity"
        ).in_("id", item_ids).execute()

        # Apply updates
        for item in current.data:
            item_id = item['id']
            new_qty = Decimal(str(item['quantity'])) + updates[item_id]
            self.client.table("inventory").update({
                "quantity": float(new_qty)
            }).eq("id", item_id).execute()

    def _fallback_to_sequential(self, items: List[Dict], user_id: str) -> Dict:
        """Fallback ensures data integrity when batch fails"""
        logger.warning("Falling back to sequential processing")
        # Process one at a time
        for item in items:
            self.process_single(item, user_id)

TypeScript

interface BatchResult {
  status: 'success' | 'partial_success' | 'failed';
  itemsProcessed: number;
  itemsFailed: number;
  failedItems?: { line: number; error: string }[];
  processingTimeMs: number;
}

class BatchProcessor {
  async processBatch(items: Item[], userId: string): Promise<BatchResult> {
    const startTime = Date.now();

    const transactionsToCreate: Transaction[] = [];
    const inventoryUpdates = new Map<string, number>();
    const failedItems: { line: number; error: string }[] = [];
    let itemsProcessed = 0;

    // Step 1: Process mappings sequentially
    for (let idx = 0; idx < items.length; idx++) {
      try {
        const mapping = await this.findOrCreateMapping(items[idx]);

        transactionsToCreate.push({
          userId,
          itemId: mapping.itemId,
          quantity: items[idx].quantity,
          unitCost: items[idx].unitPrice,
        });

        // Aggregate updates
        const current = inventoryUpdates.get(mapping.itemId) || 0;
        inventoryUpdates.set(mapping.itemId, current + items[idx].quantity);

        itemsProcessed++;
      } catch (error) {
        failedItems.push({ line: idx + 1, error: error.message });
      }
    }

    // Step 2: Batch insert
    if (transactionsToCreate.length > 0) {
      try {
        await this.db.transactions.insertMany(transactionsToCreate);
      } catch (error) {
        return this.fallbackToSequential(items, userId);
      }
    }

    // Step 3: Batch update inventory
    await this.batchUpdateInventory(inventoryUpdates);

    return {
      status: failedItems.length > 0 ? 'partial_success' : 'success',
      itemsProcessed,
      itemsFailed: failedItems.length,
      failedItems: failedItems.length > 0 ? failedItems : undefined,
      processingTimeMs: Date.now() - startTime,
    };
  }
}

Usage Examples

Invoice Processing

processor = BatchProcessor()

result = processor.process_batch(
    items=invoice_data['line_items'],
    user_id=user_id
)

if result['status'] == 'partial_success':
    logger.warning(f"Some items failed: {result['failed_items']}")

Best Practices

1. Sequential mapping, batched writes - fuzzy matching needs context, writes don't
2. Always implement fallback - batch operations can fail, sequential is reliable
3. Aggregate before update - combine multiple updates to same record
4. Handle partial success - one bad item shouldn't fail the entire batch
5. Chunk large batches - 500 records max to avoid timeouts

Common Mistakes

• Batching operations that depend on each other's results
• No fallback when batch operations fail
• Not aggregating updates to the same record
• Collecting too many records before writing (memory pressure)
• Not logging individual items when batch fails (lose context)

Related Patterns

• checkpoint-resume - Resume processing after failures
• idempotency - Prevent duplicate processing on retry
• dead-letter-queue - Handle failed items