How We Extracted 65% of Shopify API Calls from a Node Monolith Using Shadow Routing, Cutting P99 Latency by 82% and Saving $4k/Month

ad', details: validation.error.flatten() }); }

const payload: InventoryRequest = validation.data; const shadowEnabled = req.headers['x-shadow-enabled'] === 'true';

// Execute Monolith (Primary path) const monolithStart = performance.now(); let monolithRes; try { monolithRes = await callMonolithService(payload); } catch (err) { // Critical: Monolith failure must not block checkout console.error('Monolith failure, failing safe', err); return res.status(502).json({ error: 'Inventory service unavailable' }); } const monolithLatency = performance.now() - monolithStart;

// Shadow Execution (Fire-and-forget with timeout) let shadowResult: ShadowResult | null = null; if (shadowEnabled) { shadowResult = await executeShadow(payload, monolithRes, monolithLatency); // Emit metrics to OTEL here emitMetric('shadow.comparison', shadowResult); }

// Return Monolith response return res.json(monolithRes); }

async function executeShadow( payload: InventoryRequest, monolithRes: unknown, monolithLatency: number ): Promise<ShadowResult> { const moduleStart = performance.now(); try { // Call new Go-based Inventory Module const moduleRes = await fetch('http://inventory-module:8080/v1/check', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify(payload), signal: AbortSignal.timeout(50), // Strict timeout to prevent drag });

const moduleData = await moduleRes.json();
const moduleLatency = performance.now() - moduleStart;

// Deep comparison logic
const match = deepEqual(monolithRes, moduleData);

return {
  monolithLatency,
  moduleLatency,
  match,
  monolithData: monolithRes,
  moduleData,
};

} catch (err) { // Shadow failures are non-fatal but must be logged console.error('Shadow execution failed', err); return { monolithLatency, moduleLatency: -1, match: false, monolithData: monolithRes, moduleData: null, }; } }

// Mock monolith call async function callMonolithService(payload: InventoryRequest) { // Implementation omitted for brevity return { available: true, price: 29.99 }; }

function deepEqual(a: unknown, b: unknown): boolean { // Simplified deep equal; in prod use 'lodash.isequal' or 'fast-deep-equal' return JSON.stringify(a) === JSON.stringify(b); }

### Step 2: High-Throughput Inventory Worker

The monolith's inventory logic was slow due to ORM overhead. We rewrote this in Go using `pgx` for direct driver access. This worker consumes Shopify webhooks and updates the local PostgreSQL 17 instance. It implements a token bucket rate limiter to respect Shopify's API constraints.

**`inventory_worker.go`**
```go
package main

import (
	"context"
	"crypto/hmac"
	"crypto/sha256"
	"encoding/base64"
	"encoding/json"
	"fmt"
	"log"
	"net/http"
	"os"
	"time"

	"github.com/jackc/pgx/v5/pgxpool"
	"golang.org/x/time/rate"
)

type ShopifyWebhook struct {
	ID       int64  `json:"id"`
	Title    string `json:"title"`
	Variants []struct {
		ID        int64 `json:"id"`
		Inventory int   `json:"inventory_quantity"`
	} `json:"variants"`
}

var dbPool *pgxpool.Pool
var limiter = rate.NewLimiter(rate.Every(time.Second/40), 40) // 40 req/sec

func main() {
	// Init DB Pool (PostgreSQL 17)
	var err error
	dbPool, err = pgxpool.New(context.Background(), os.Getenv("DATABASE_URL"))
	if err != nil {
		log.Fatalf("Unable to create connection pool: %v", err)
	}
	defer dbPool.Close()

	http.HandleFunc("/webhook/shopify/products/update", handleWebhook)
	log.Println("Worker listening on :8080")
	log.Fatal(http.ListenAndServe(":8080", nil))
}

func handleWebhook(w http.ResponseWriter, r *http.Request) {
	// 1. Verify HMAC
	hmacHeader := r.Header.Get("X-Shopify-Hmac-Sha256")
	if !verifyHmac(r.Body, hmacHeader, os.Getenv("SHOPIFY_WEBHOOK_SECRET")) {
		http.Error(w, "Invalid HMAC", http.StatusUnauthorized)
		return
	}

	// 2. Parse Payload
	var webhook ShopifyWebhook
	if err := json.NewDecoder(r.Body).Decode(&webhook); err != nil {
		http.Error(w, "Bad JSON", http.StatusBadRequest)
		return
	}

	// 3. Upsert Inventory with Conflict Resolution
	// Shopify can send updates before creates in rapid succession.
	// We use INSERT ... ON CONFLICT to handle this safely.
	for _, v := range webhook.Variants {
		query := `
			INSERT INTO inventory (shopify_variant_id, quantity, updated_at)
			VALUES ($1, $2, NOW())
			ON CONFLICT (shopify_variant_id) 
			DO UPDATE SET quantity = EXCLUDED.quantity, updated_at = NOW()
		`
		_, err := dbPool.Exec(context.Background(), query, v.ID, v.Inventory)
		if err != nil {
			log.Printf("Failed to upsert variant %d: %v", v.ID, err)
			// In prod, push to dead-letter queue for reconciliation
		}
	}

	w.WriteHeader(http.StatusOK)
}

func verifyHmac(body io.ReadCloser, header string, secret string) bool {
	// HMAC verification logic
	// ... implementation ...
	return true
}

Step 3: State Reconciliation Loop

Webhooks can be lost or arrive out of order. We implemented a reconciliation worker that runs every 5 minutes. It queries Shopify for all variants and diffs them against the local DB. This is the safety net that guarantees consistency.

reconciler.ts

import { Client } from 'shopify-api-node'; // Using shopify-api-node v3.10.0
import { Pool } from 'pg'; // pg v8.12.0
import { z } from 'zod';

const ShopifyVariantSchema = z.object({
  id: z.number(),
  inventory_quantity: z.number(),
});

export async function runReconciliation() {
  const shopify = new Client({
    shopName: process.env.SHOPIFY_SHOP,
    accessToken: process.env.SHOPIFY_ACCESS_TOKEN,
    apiVersion: '2024-10',
  });

  const pool = new Pool({ connectionString: process.env.DATABASE_URL });

  try {
    // 1. Fetch all variants from Shopify (Paginated)
    const shopifyVariants = await fetchAllShopifyVariants(shopify);
    
    // 2. Fetch local variants
    const localRes = await pool.query('SELECT shopify_variant_id, quantity FROM inventory');
    const localMap = new Map(
      localRes.rows.map(r => [r.shopify_variant_id, r.quantity])
    );

    // 3. Diff and Fix
    let driftCount = 0;
    for (const variant of shopifyVariants) {
      const localQty = localMap.get(variant.id);
      if (localQty !== variant.inventory_quantity) {
        // Drift detected. Update local DB.
        // Use UPSERT to handle missing records
        await pool.query(
          `INSERT INTO inventory (shopify_variant_id, quantity, updated_at) 
           VALUES ($1, $2, NOW()) 
           ON CONFLICT (shopify_variant_id) DO UPDATE SET quantity = $2, updated_at = NOW()`,
          [variant.id, variant.inventory_quantity]
        );
        driftCount++;
      }
    }

    console.log(`Reconciliation complete. Fixed ${driftCount} drifted records.`);
  } catch (err) {
    console.error('Reconciliation failed', err);
    // Alert on failure
  } finally {
    await pool.end();
  }
}

async function fetchAllShopifyVariants(shopify: Client) {
  // Implementation of pagination using 'since_id'
  // Returns array of ShopifyVariantSchema
  return [];
}

Pitfall Guide

During migration, we encountered specific failures that aren't covered in Shopify docs. Here is how to debug them.

Real Production Failures

1. The "Ghost" Cart Token

   • Symptom: ShopifyBuyError: Cart token is invalid or expired during shadow routing.
   • Root Cause: The monolith and the new module used different session stores. The shadow router passed the monolith's session token to the new module, which rejected it.
   • Fix: We implemented a Token Migration Layer in the router. If the token is monolith-formatted, we decode it, extract the cartId, and re-sign it for the module context before shadowing.
   • Code Snippet: const moduleToken = migrateToken(req.headers['x-cart-token']);

2. Webhook Burst Overload

   • Symptom: pq: deadlock detected in PostgreSQL 17 during flash sales.
   • Root Cause: Shopify sends multiple products/update webhooks for a single product change (e.g., updating title triggers inventory update webhook). The Go worker processed them concurrently, causing row-level deadlocks on the inventory table.
   • Fix: Added a Distributed Lock using Redis 7.4 SETNX with a 500ms TTL per variant_id. This serialized updates for the same variant without blocking unrelated variants.
   • Metric: Deadlocks dropped from 120/min to 0.

3. GraphQL Rate Limiting in Reconciliation

   • Symptom: ShopifyApiError: Throttled during reconciliation.
   • Root Cause: The reconciler queried variants one by one or used a large query that consumed too many cost points.
   • Fix: Switched to a Cursor-Based Bulk Query with a strict token bucket. We batched 250 variants per query.
   • Code: query { variants(first: 250, after: $cursor) { ... } }

Troubleshooting Table

Error Message / Symptom	Root Cause	Immediate Fix
STALE_CART_TOKEN	Session mismatch between monolith and module.	Implement token migration in router; ensure sticky sessions during shadow phase.
DUPLICATE_KEY on shopify_variant_id	Shopify webhook ordering race.	Use INSERT ... ON CONFLICT (UPSERT); add distributed lock per variant.
RATE_LIMIT_EXCEEDED	Burst traffic or aggressive polling.	Implement leaky bucket rate limiter; switch to webhooks; add jitter to retries.
Shadow Mismatch Rate > 0.1%	Logic drift or timing differences.	Check timezone handling; verify currency rounding logic; inspect deepEqual implementation.
Connection Refused on Module	Module crashed or K8s readiness probe failed.	Check module logs for OOM; verify HPA metrics; check RBAC permissions.

Production Bundle

Performance Metrics

After 6 months of incremental extraction using this pattern:

• Latency: P99 latency on /checkout dropped from 340ms to 61ms (82% reduction). The Go inventory module responds in <8ms P99.
• Throughput: The monolith CPU usage dropped by 45%. We downsized the monolith cluster from 12 c6g.4xlarge to 6 instances.
• Error Rate: Overselling incidents dropped from 14/month to 0. The reconciliation loop catches 99.9% of drift within 5 minutes.
• Deployment Frequency: Deployments went from 42 minutes to 4 minutes for the module. The monolith deploys are now isolated to non-inventory changes.

Cost Analysis

• Infrastructure Savings:
  • Monolith downsizing: Saved $2,100/month.
  • Database optimization: Moved from r6gd.xlarge to r6gd.large with read replicas; Saved $800/month.
  • Total Savings: $2,900/month.
• New Costs:
  • Go Worker Cluster: $400/month (2 t4g.small instances).
  • Redis Cluster: $150/month.
  • Net Savings: $2,350/month ($28,200/year).
• Productivity ROI:
  • Developer velocity increased by 3x. New features in inventory module ship in hours, not days.
  • Estimated value: 2 senior engineers saved 10 hours/week each = $4,000/month in engineering time.
  • Total ROI: $6,350/month.

Monitoring Setup

We use OpenTelemetry (OTEL) for end-to-end tracing.

1. Dashboard: "Shadow Router Health".
   • Metrics: shadow.match_rate, shadow.latency_delta, shadow.error_rate.
   • Alert: If shadow.match_rate < 99.9% for 5 minutes, page on-call.
2. Dashboard: "Shopify Sync Drift".
   • Metrics: reconciler.drift_count, webhook.processing_lag.
   • Alert: If webhook.processing_lag > 30s, scale Go workers.
3. Tracing: Every request carries a trace_id. We can trace a checkout request from the browser, through the router, into the monolith, and see the shadow call to the Go module.

Scaling Considerations

• Go Worker: Scales horizontally based on webhook queue depth. We use KEDA to scale based on Redis list length. At peak, we scale to 20 pods handling 15k webhooks/sec.
• Database: PostgreSQL 17 handles the write load easily due to UPSERT efficiency. We use connection pooling via pgbouncer (v1.22) to manage connections from the Go workers.
• Router: The Node.js router is stateless and scales via K8s HPA on CPU. It handles 5k req/sec per pod.

Actionable Checklist

1. Define Domain Boundary: Choose a domain with clear inputs/outputs (e.g., Inventory, Pricing). Avoid Checkout transaction logic initially.
2. Implement Shadow Router: Deploy the router with shadow mode off. Verify zero latency impact.
3. Build New Module: Write the module with idempotency and error handling. Use Go for high-throughput workers.
4. Enable Shadow Mode: Turn on shadow for 1% of traffic. Monitor mismatch rates.
5. Fix Drift: Iterate until shadow.match_rate is >99.99%.
6. Increase Traffic: Ramp shadow traffic to 10%, 50%, 100%.
7. Switch Primary: Flip feature flag to route traffic to module. Keep monolith as fallback.
8. Decommission: Once stable for 2 weeks, remove monolith code for that domain.

This pattern allowed us to modularize a critical Shopify integration without a single minute of downtime or data loss. The key is not just extracting code, but rigorously validating state equivalence before trusting the new system.