Auto-scaling Arqel panels on Laravel Cloud

Admin panels have a very different load profile from public-facing applications: traffic concentrated during business hours, export spikes at month-end, and long-running AI jobs. This guide covers how to size each component without waste.

Scaling components

A standard Arqel app has four processes that scale independently:

Process	Role	Scale metric
Web (PHP-FPM/Octane)	Inertia routes, table fetches, form submits	RPS + p95 latency
Queue worker	Bulk Actions, Exports, AI jobs, broadcasts	Queue depth + job duration
Reverb	WebSocket fan-out (Inertia event broadcasting)	Concurrent connections
Scheduler	php artisan schedule:run	(does not scale — fixed at 1×)

Web instances

Simple rule: 1 web instance handles ~50 concurrent admins on a typical Arqel (assuming Octane on, Postgres with pgbouncer, and Redis for cache). Without Octane, count on ~20.

Size	Avg RPS	Web instances	Instance type
Small (≤10 users)	≤5	1×	nano (256 MB / 0.25 vCPU)
Medium (10-50)	5-30	2×	small (512 MB / 0.5 vCPU)
Large (50-200)	30-100	4-6×	medium (1 GB / 1 vCPU)
XL (200+)	100+	8+ × auto	large (2 GB / 2 vCPU)

Recommended auto-scale rule (configure in Settings → Scaling):

web:
  min: 2
  max: 10
  scale_up_when:
    cpu_average: '> 70% por 2min'
    OR
    p95_latency: '> 800ms por 1min'
  scale_down_when:
    cpu_average: '< 25% por 10min'

Classic admin-panel spike: Monday 09:00 (everyone opens the app). Pre-warm to 4× at that time if you know the pattern (scheduled scaling).

Queue workers

Bulk Actions and Exports are the most common bottleneck in admin panels. A worker processes around 10 jobs/s for simple CRUD, but only 0.1 jobs/s for 50k-row XLSX exports.

Case	Workers	Dedicated queues
Small (no AI, rare exports)	1×	default
Medium (occasional AI, daily exports)	2×	default, exports
Large (heavy AI, hourly exports)	4× + 2× exports	default, exports, ai
XL	8+ × default + 4× exports + 4× ai	dedicated

cloud.yaml configuration:

services:
  worker:
    command: php artisan queue:work redis --queue=default,exports --tries=3 --max-time=3600
    instances:
      min: 1
      max: 4
  ai_worker:
    command: php artisan queue:work redis --queue=ai --tries=2 --timeout=600
    instances:
      min: 1
      max: 2

Use separate queues so a large Export does not block fast mutations like "Approve order".

Concurrency on AI jobs

@arqel-dev/ai jobs (LLM calls via ClaudeProvider or OpenAIProvider) have high latency (2-30s) and per-call cost. Throttle with Laravel's RateLimited middleware:

public function middleware(): array
{
    return [new RateLimited('ai-tier')];
}

And configure in AppServiceProvider:

RateLimiter::for('ai-tier', fn () => Limit::perMinute(60));

Reverb (WebSocket)

Reverb is stateful — each connection sticks to one instance. To scale horizontally, use Redis pub/sub (already enabled in the cloud:export template).

Concurrent connections	Reverb instances	Memory/instance
≤500	1×	512 MB
500-2,000	2×	1 GB
2,000-10,000	4×	2 GB
10,000+	8+ × with sticky session	4 GB

Important: Laravel Cloud's load balancer already handles sticky sessions via the _reverb_sid cookie. Without it, connections drop on rescale.

services:
  reverb:
    command: php artisan reverb:start --host=0.0.0.0 --port=8080
    instances:
      min: 1
      max: 4
    sticky_session: true

Database connection pool

Laravel Cloud's Postgres ships with pgbouncer in front. Limits:

Tier	Max connections	Recommended for web/worker
Hobby	25	2× web + 1× worker
Starter	100	4× web + 2× worker
Production	400	10× web + 8× worker
Scale	1000+	no practical limit

Each Octane worker opens a persistent connection. If you have 4 web instances × 4 Octane workers = 16 connections from web alone. Add queue workers (typically 1 per process) and size the Postgres tier accordingly.

Symptom of an exhausted pool: SQLSTATE[08006] FATAL: sorry, too many clients already. Immediate fix: bump the Postgres tier. Long-term fix: set database.connections.pgsql.options.persistent = false on workers (but latency suffers).

Recommendations by application size

Small — startup or internal tool (≤10 users)

web:    { min: 1, max: 1, type: nano }
worker: { min: 1, max: 1, type: nano }
reverb: { min: 1, max: 1, type: nano }
postgres: hobby
redis: hobby

Approximate cost: $15-25/month (no AI tokens).

Medium — small/medium business (10-50 users)

web:    { min: 2, max: 4, type: small }
worker: { min: 2, max: 4, type: small }
reverb: { min: 1, max: 2, type: small }
postgres: starter
redis: starter

Approximate cost: $80-150/month.

Large — SaaS product (50-200 users)

web:        { min: 4, max: 8, type: medium }
worker:     { min: 4, max: 8, type: medium }
ai_worker:  { min: 2, max: 4, type: medium }
reverb:     { min: 2, max: 4, type: medium, sticky_session: true }
postgres: production
redis: production

Approximate cost: $400-700/month.

Monitoring

Metrics worth alerting on in Laravel Cloud:

• Web p95 > 1s for 5min → bump web.min or optimize the query.
• Queue depth > 1000 for 2min → bump worker.max.
• Reverb dropped connections > 5% → missing sticky session or slow Redis.
• Postgres CPU > 80% for 10min → tier upgrade or missing index.

Integrate with Sentry (recommended) or native Cloud Logs.

Next steps

• Estimate exact costs for your scenario → cost-estimation.md.
• Compare with other platforms → comparison-other-hosts.md.