Trendlyric publishes around the clock with zero human intervention. The whole pipeline - ingest, classify, summarise, illustrate, publish - runs on a single managed VPS. People assume you need Kubernetes. You don't.
What the pipeline actually does
Five steps. Ingest polls global news feeds on a continuous schedule. Classify tags each story by topic, region, and freshness, and runs a duplicate check against the last 30 days of published content. Summarise uses an LLM to rewrite the story in editorial voice, with a citation back to the source. Illustrate matches an image from a licensed library based on entities mentioned. Publish pushes to WordPress via REST, pings the sitemap, purges the CDN, and auto-posts to social.
Why a single VPS is enough
The whole stack is boring on purpose. A managed Hetzner VPS, Postgres for state, Redis for the queue, supervisord for processes, nginx out front. No autoscaling, no service mesh, no Kafka. The job throughput is well within what a single 4-vCPU box can handle, and it has been for years.
The unsexy alternative would have cost ten times more in serverless invocations and idle compute. Boring infra wins when load is steady.
What we'd do differently
One thing - we'd add a hot standby earlier. Trendlyric has had two unplanned outages in nine years; both were resolved in under 20 minutes, but a warm replica would have made them invisible. We added one in 2024.