Zero-restart deploys
Apply a definition — the gateway picks up the new shape in place. In-flight requests finish on the old shape; new ones land on the new.
Platform teams · Runtime
Deploy at lunch. Don't tell ops.
Every definition, policy, and transform hot-loads on the running gateway. No restarts, no rolling reboots, no lost in-flight requests — and the distributed cache survives every pod reschedule.
The problem
Most gateways pretend a deploy is a small thing. It isn't.
A redeploy means a rolling restart. A rolling restart means warm caches go cold, connection pools rebuild, JWKs re-fetch, and tail latency spikes for ten minutes. Engineers learn to deploy on Fridays at 04:00 — which means they don't deploy. Apinizer's runtime treats a deploy as a config refresh: the gateway picks up the new shape in place, and the distributed cache holds steady.
- 0
restarts
per definition apply
- <200ms
propagation
across Workers
- 0
in-flight requests lost
Capabilities
What Apinizer does here
Distributed cache
Hazelcast-backed cache shared across every Worker. Survives reschedules; invalidations propagate cluster-wide in under a second.
Coordinated invalidation
Bust a cache key once; every Worker drops it. No more 'one node has the stale value' tickets.
Per-key TTLs
Cache strategies per endpoint, per consumer, per response variant. Fast lanes for hot data, conservative TTLs for sensitive lookups.
Pre-warmed pools
Connection pools, JWK sets, OAuth introspection caches — all kept warm across deploys. New pods inherit warm state from peers.
Cache analytics
Hit rate, miss rate, eviction reason — per endpoint, per region. Tune what's worth caching before it becomes a cost problem.
Use cases
In production, this looks like…
- Retail
Istanbul e-commerce ships 40 policy changes during Black Friday
Rate limits, headers, A/B routes — all hot-deployed during peak. Zero rolling restarts, zero abandoned carts attributable to platform.
40 deploys, 0 restarts
- Banking
Frankfurt private bank refreshes JWKs without a maintenance window
Identity provider rotates keys. The gateway picks up the new JWKs in place; in-flight tokens validate on the old key, new ones on the new.
- Public transit
Amsterdam transit authority caches schedule API across 6 Workers
97% cache hit rate on schedule lookups during peak commute. Invalidation on disruption fans out in 400ms.
97% hit rate
- Insurance
Stockholm insurer A/B-tests a new pricing endpoint mid-day
Hot-deploy carves 5% of traffic to the new shape. Telemetry confirms; the next apply moves the cutover to 100%. No deploy window negotiated.
- Banking
Warsaw bank serves loan calculator from cache during marketing campaign
Calculator API absorbs 12k RPS during a TV ad. Backend never sees more than 200 RPS; cache shoulders the rest with per-input TTLs.
60× backend reduction
- Telecom
Prague carrier ships pod reschedules without losing cache warmth
Hazelcast keeps state across the cluster. When a pod reschedules, the new pod inherits the cache from peers — no cold-start storm to the backend.
- Media
Milan publisher invalidates 4M cache keys per breaking story
Editorial publishes; the gateway invalidates affected keys cluster-wide. CDN pulls fresh in under a second; readers never see stale headlines.
- E-government
Baku ministry refreshes 200 endpoints during business hours
APIops applies land hot. The platform team stops scheduling deploys for nights; the operations runbook loses its longest section.
How it works
Apply, propagate, serve — in that order, every time.
- Step 01
Apply
Push a definition via UI, APIops, or pipeline. The Manager validates and broadcasts to every Worker.
- Step 02
Propagate
Workers receive the change in under a second. The cache layer keeps hot keys; nothing is dropped.
- Step 03
Serve
In-flight requests complete on the old shape. New requests land on the new — no restart, no warm-up, no spike.
- Step 04
Observe
Analytics shows the cutover, cache hit rates, and tail latency in real time. Roll back is another apply.
Recommended products
What we'd reach for first
Cache
Hazelcast-backed distributed cache with coordinated invalidation across every Worker.
Open the Cache pageAPI Gateway
The runtime that hot-loads definitions, policies, and transforms without dropping a request.
Open the Gateway pageAnalytics Engine
Real-time evidence that a deploy did what you expected — and rollback evidence when it didn't.
Open the Analytics pageMonitoring
Synthetic probes that fire seconds after an apply — confirm green before the team sees the alert.
Open the Monitoring pageResources
Keep going
Hot deploy semantics
What apply, propagate, serve looks like under the hood — and why it doesn't drop requests.
Cache product
Hazelcast-backed distributed cache with coordinated invalidation and per-key TTLs.
Analytics Engine
Live deploy cutover, hit rate, and tail-latency telemetry — per definition.
Architecture overview
How the Manager publishes and Workers pick up changes without a restart.
APIops manifests
Hot deploys driven from Git — the same apply that promotes across environments.
Changelog
Recent runtime improvements — cache, hot deploys, and propagation guarantees.
Deploy without the dread
Push at noon. Sleep at midnight.
A 30-minute walkthrough — apply, propagate, cache, observe — on a Kubernetes of your choice.