HTTP 503 Service Unavailable (maintenance & overload)

The service is temporarily not taking your request — maintenance, overload, or no healthy backend. Often you should wait and retry.

What 503 is saying

5xx = server-side. 503 is the polite “closed for a bit” signal. Good operators return it deliberately during deploys or when refusing load is safer than melting the database.

HTTP/1.1 503 Service Unavailable
Retry-After: 120
Content-Type: application/json

{"error":"service_unavailable","message":"Try again in 2 minutes"}

503 vs lookalikes

Some CDNs map “origin down” to 521/522/523 — same idea as unavailable, different banners.

Typical causes

• Maintenance / deploy — process stopped; LB returns 503 by design.
• No healthy backends — health checks failing; target group empty; Kubernetes endpoints blank.
• Overload — thread/worker pool full; intentionally shedding load.
• Circuit breaker open — sidecar (Envoy, app library) stops calling a failing dependency and fails fast with 503.
• Draining — instance marked unhealthy for shutdown; new traffic gets 503 while in-flight finish.

Retry-After

Same rules as 429: integer seconds or an HTTP-date. Clients should wait at least that long. If missing, use exponential backoff with jitter and a capped attempt count.

curl -sI https://example.com/api/health
# Look for HTTP/2 503 and Retry-After

Operator checklist

1. Return 503 (not 500) for planned unavailability; include Retry-After when you know the window.
2. Health checks should fail soft during drain so the LB stops sending new work.
3. Capacity: autoscale on queue depth / p99 latency, not only CPU.
4. Fail fast on broken dependencies (circuit breaker) instead of holding every worker until 504.
5. Status page + synthetic checks so humans know it is you, not their Wi-Fi.

Client checklist

1. Honor Retry-After; add jitter.
2. Treat 503 as transient for idempotent reads; show a calm “try again shortly” UI.
3. Do not spin a tight retry loop — that worsens overload.
4. Distinguish 503 (server busy) from 429 (you personally are too chatty).

Kid-friendly sibling: Wirebook — when sites get tired (queues, rate limits, sharing the load).

Related

Cloudflare Error 521 · Cloudflare Error 522 · Cloudflare Error 524 · HTTP 504 Gateway Timeout · 502 Bad Gateway · HTTP 429 Too Many Requests · ETIMEDOUT · net::ERR_CONNECTION_TIMED_OUT · TypeError: Failed to fetch · Anvil HTTP status · Wirebook: when sites get tired

Questions

Published 6 November 2026.