Good breakdown. One thing that decided it for me beyond cost: what happens on a bad deploy. With Vercel a rollback is one click. On a self-hosted setup it depends on the tool, so I test rollback and a database restore before moving anything real. I run 15+ sites this way and built Levelrail for the parts I kept missing (audit log, database backups with restore, a small footprint on cheap VPSs): github.com/glincker/levelrail
Agreed, rollback is the part people skip testing until they need it. Coolify does keep previous images, so rolling the app back is quick, but its docs are clear that a rollback doesn't restore databases or persistent files. So I test the two separately: roll the app back, then restore a dump into a scratch database, before any real traffic moves over. Schema migrations are where they come apart, which is why I keep migrations backward compatible for at least one deploy.
I'd decide based on the runtime shape before comparing the headline prices. If the app needs long-running workers, scheduled jobs, and a queue, the frontend host is only one part of the deployment. I'd map where the database, workers, and scheduler live—and how failures are observed—before choosing a platform.
That's the right order, Vlad. Once workers, a scheduler and a queue are in the picture, the frontend host is the easy part, and on Vercel those pieces usually end up somewhere else anyway. That's why a lot of teams split it: frontend on Vercel, API, workers and data on a server. Good call on observability too. On a single Coolify box you have to add it yourself, because nothing tells you a worker died unless you set that up.
One thing I'd add is that process health alone isn't enough. A worker can be running but stuck or falling behind. I'd also track queue depth, the age of the oldest pending job, retries, and failed jobs. That gives you a much better picture of whether the background system is actually making progress.
Agreed, a worker can be up and still not doing anything useful. Of those, the age of the oldest pending job is the one I'd alert on first. Depth can spike during a normal burst and clear on its own, but if the oldest job keeps getting older, nothing is draining the queue. Failed and retried counts are the next thing I'd watch, since a job that fails and retries forever can look like progress from the outside.
I'd track that per queue rather than globally, especially when jobs have different priorities. A healthy high-volume queue shouldn't hide a stalled critical one. And retries need a clear max-attempt or dead-letter policy, otherwise the system can keep looking busy without actually completing the work.
Yes, per queue makes sense. A busy email queue can easily hide a stuck payments queue if you only look at totals. And agreed on retries: cap the attempts and move what's left to a dead-letter queue, then alert when that queue isn't empty. Otherwise a job that fails every time just keeps the workers busy forever.
Freelance fullstack developer creating fast, modern, and refined applications. I focus on simplicity, performance, and real value for users. I also write tech-related articles here.
The distinction between pricing models and runtime constraints is explained really well here.
Everyone talks about the cost difference (flat VPS vs. seat + usage billing), but the operational dealbreakers are usually what happens after deployment:
Colocated databases & background workers: Having Postgres and a queue worker running on the same private Docker network with zero latency and zero data-transfer fees is huge compared to stitching together external serverless DBs and function timeouts.
Execution limits: Being able to run long-running tasks without watching the 300s/800s serverless countdown clock completely changes how you architect backend jobs.
That said, Vercel’s friction-free PR previews and global edge CDN out of the box are still unmatched for purely frontend-heavy teams without ops bandwidth. Putting Cloudflare in front of a Coolify box bridges a lot of that gap, but you definitely have to accept being on-call for your own hardware. Great, balanced comparison!
Thanks, John. Agreed on both, and being on call for your own box is the real price of the cheaper bill. Coolify does do PR previews too, but you wire up the GitHub App and a wildcard domain yourself, so "friction-free" is still the right word for Vercel's.
The backups point is the one I would check first. A single Coolify server is fine for a small app, but only if the database dump is tested off-site, not just scheduled.
Agreed, Rishabh. A scheduled dump you've never restored doesn't prove much. Coolify can send its database backups to S3-compatible storage, and I'd restore one into a scratch database every so often to check it actually comes back. It matters more on Coolify because a rollback there doesn't restore the database or persistent files.
That scratch-restore habit is the part I will adopt. A backup that has never been restored is only a hope, not a safety net, so I will log each restore result next to the deploy.
We're a place where coders share, stay up-to-date and grow their careers.
Good breakdown. One thing that decided it for me beyond cost: what happens on a bad deploy. With Vercel a rollback is one click. On a self-hosted setup it depends on the tool, so I test rollback and a database restore before moving anything real. I run 15+ sites this way and built Levelrail for the parts I kept missing (audit log, database backups with restore, a small footprint on cheap VPSs): github.com/glincker/levelrail
Agreed, rollback is the part people skip testing until they need it. Coolify does keep previous images, so rolling the app back is quick, but its docs are clear that a rollback doesn't restore databases or persistent files. So I test the two separately: roll the app back, then restore a dump into a scratch database, before any real traffic moves over. Schema migrations are where they come apart, which is why I keep migrations backward compatible for at least one deploy.
I'd decide based on the runtime shape before comparing the headline prices. If the app needs long-running workers, scheduled jobs, and a queue, the frontend host is only one part of the deployment. I'd map where the database, workers, and scheduler live—and how failures are observed—before choosing a platform.
That's the right order, Vlad. Once workers, a scheduler and a queue are in the picture, the frontend host is the easy part, and on Vercel those pieces usually end up somewhere else anyway. That's why a lot of teams split it: frontend on Vercel, API, workers and data on a server. Good call on observability too. On a single Coolify box you have to add it yourself, because nothing tells you a worker died unless you set that up.
One thing I'd add is that process health alone isn't enough. A worker can be running but stuck or falling behind. I'd also track queue depth, the age of the oldest pending job, retries, and failed jobs. That gives you a much better picture of whether the background system is actually making progress.
Agreed, a worker can be up and still not doing anything useful. Of those, the age of the oldest pending job is the one I'd alert on first. Depth can spike during a normal burst and clear on its own, but if the oldest job keeps getting older, nothing is draining the queue. Failed and retried counts are the next thing I'd watch, since a job that fails and retries forever can look like progress from the outside.
I'd track that per queue rather than globally, especially when jobs have different priorities. A healthy high-volume queue shouldn't hide a stalled critical one. And retries need a clear max-attempt or dead-letter policy, otherwise the system can keep looking busy without actually completing the work.
Yes, per queue makes sense. A busy email queue can easily hide a stuck payments queue if you only look at totals. And agreed on retries: cap the attempts and move what's left to a dead-letter queue, then alert when that queue isn't empty. Otherwise a job that fails every time just keeps the workers busy forever.
The distinction between pricing models and runtime constraints is explained really well here.
Everyone talks about the cost difference (flat VPS vs. seat + usage billing), but the operational dealbreakers are usually what happens after deployment:
That said, Vercel’s friction-free PR previews and global edge CDN out of the box are still unmatched for purely frontend-heavy teams without ops bandwidth. Putting Cloudflare in front of a Coolify box bridges a lot of that gap, but you definitely have to accept being on-call for your own hardware. Great, balanced comparison!
Thanks, John. Agreed on both, and being on call for your own box is the real price of the cheaper bill. Coolify does do PR previews too, but you wire up the GitHub App and a wildcard domain yourself, so "friction-free" is still the right word for Vercel's.
The backups point is the one I would check first. A single Coolify server is fine for a small app, but only if the database dump is tested off-site, not just scheduled.
Agreed, Rishabh. A scheduled dump you've never restored doesn't prove much. Coolify can send its database backups to S3-compatible storage, and I'd restore one into a scratch database every so often to check it actually comes back. It matters more on Coolify because a rollback there doesn't restore the database or persistent files.
That scratch-restore habit is the part I will adopt. A backup that has never been restored is only a hope, not a safety net, so I will log each restore result next to the deploy.