mguilbert/matheoguilbert.fr
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update blog article number 3
Deploy Vercel / deploy-preprod (push) Successful in 1m31s
Details
Deploy Vercel / deploy-production (push) Has been skipped
Details

Browse Source
This commit is contained in:
Mathéo GUILBERT
2026-06-12 16:00:48 +02:00
parent abed8a9cf4
commit 7d7fc36938
3 changed files with 286 additions and 99 deletions
Download Patch File Download Diff File Expand all files Collapse all files
content/blog/en/03-Deploying-an-infrastructure-for-a-macos-app-distribution.mdx
+163 -74
View File
@@ -1,140 +1,229 @@
---
ready: true
translationKey: deploying-an-infrastructure-for-a-macos-app-distribution
title: Deploying a macOS app distribution infrastructure for the price of a monthly dinner
title: Deploying an Infrastructure for a macOS App Distribution for the Price of a Restaurant Meal per Month
description: Long live self-hosting!
date: 2026-05-21
category: Architecture
tags:
- Experience
- TechStack
- Technologies
readTime: 5 min
summary:
- Introduction
- "The specifications: Exploit Open Source"
- "The core of the problem: making public and private coexist"
- "Overview: How it runs day-to-day?"
- The monitoring and the backups
- The costs in all that
- "The Specifications: Leveraging Open Source"
- "The Core of the Problem: Making Public and Private Coexist"
- "Overview: How Does It Run Daily?"
- Monitoring and Backups
- The Costs in All This
cover: /images/blog/03/deploying-an-infrastructure-for-a-macos-app-distribution.png
---
## Introduction
A few weeks ago at the time of writing this article, I launched into the development of Thence, a macOS application that remembers a developer's project context to save them time and energy when they resume it after a break.
A few weeks ago at the time I am writing this article, I launched into the development of Thence, a macOS application that memorizes a developer's project context to save them time and energy when they resume it after a break.
The application code as such is only the tip of the iceberg. Very quickly, the reality on the ground catches up with you: to make a product live, you need a whole ecosystem around it. A space for the community, a software license management system, and internal tools to manage everything and make the right decisions for the product's evolution.
The code of the application as such is only the tip of the iceberg. Very quickly, the reality of the field catches up with you: to make a product live, you need an entire ecosystem around it. A space for the community, a software license management system, and internal tools to pilot everything and make the right decisions for the evolution of the product.
If we turn to all kinds of SaaS, each specializing in a particular task, with the versatility of the necessary systems, we can quickly find ourselves with a hefty bill. Being a student at that time, and having much better plans for my money, I made a radical decision: self-host as much as possible.
If we turn to SaaS of all kinds, each specialized in one task in particular, with the versatility of the necessary systems, you quickly end up with a salty bill. Being a student at that time, and having much better projects for my money, I took a radical decision: self-host as much as possible.
In this article, I invite you to explore the architecture of my VPS (Virtual Private Server). We will see how I managed to make public and private tools coexist on one and the same machine, the technical choices behind each brick, and how this "system D" (resourceful) approach allowed me to build a reliable, scalable, and available architecture, for not too much money.
In this article, I propose to go through the architecture of my VPS (Virtual Private Server). We will see how I managed to make public and private tools coexist on one and the same machine, the technical choices behind each brick and how this "system D" approach allowed me to build a reliable, scalable and available architecture, for not too expensive.
## The Specifications: Leveraging Open Source
The goal was not to host services for the pleasure of testing them, but indeed to meet a real business need for Thence. For each need, I searched for and selected the best free, self-hostable solution in the best-case scenario, capable of running efficiently in Docker containers, without letting myself bury under technical debt.
## The specifications: Exploit Open Source
### Code Hosting: Gitea
Goodbye GitHub or GitLab, I wanted to maintain the sovereignty of my code and manage my own CI/CD runners. Gitea is lightweight and does the job, so it stood out as an obvious choice.
The objective was not to host services for the pleasure of testing them, but bel and bien to respond to a real business need for Thence. For each need I searched and selected the best free solution, self-hostable in the best of cases, capable of running efficiently in Docker containers, without however making me collapse under technical debt.
### Automation: n8n
To manage my blog's RSS feeds and my newsletter, I needed a conductor. I had wanted to test n8n for a while, it was the perfect opportunity!
### Distribution and licenses: Keygen
### Distribution and Licenses: Keygen
For a paid desktop application, software license management is the sinews of war. I needed a system capable of generating keys, managing activations, and ensuring that a user doesn't deploy the application on 10 different machines with a single subscription. I deployed the self-hosted version of Keygen for this.
For a paid desktop application, software license management is the nerve of the war. I needed a system capable of generating keys, managing activations and ensuring that a user does not deploy the application on 10 different machines with a single subscription. I deployed the self-hosted version of Keygen for that.
### Support and Community: Discourse
Rather than opening a Discord server that is difficult for Google to index, or managing an unmanageable amount of customer support emails visible only to me, I chose Discourse. It's a forum, exposed to the Internet, which allows me to structure discussions with users, publish roadmaps, and exchange with the community around the application in general.
### Support and community: Discourse
More precisely, I use it for several things at once:
* **Pre-registering users:** It is by inviting them to this forum in a dedicated group that I was able to find the beta users of the application, those with whom I am building the MVP (Minimum Viable Product), still at the time of writing this article.
* **Creating a public knowledge base self-fueled by users:** This is the second strong point of this kind of tool. People ask questions, I answer them, and other people who ask themselves the same questions read the discussions to find their answers. It is a self-fueled FAQ that answers users' questions directly, and not questions that I think users will ask.
Rather than opening a Discord server difficultly indexable by Google, or managing a quantity of support client emails intractable and visible only by me, I chose Discourse. It is a forum, exposed on the Internet, which allows me to structure discussions with users, publish roadmaps, exchange with the community around the application in a general way.
### Deployment and Monitoring: Komodo
Deploying applications galore in Docker containers is good. But being able to do it graphically, monitor their health status, and perform their updates in a few clicks is better! Not that typing commands by hand particularly annoys me, on the contrary, but it's mostly tiring, longer. This is where Komodo comes in, my control center to manage the vast majority of my self-hosted applications serenely.
More precisely, I use it for several things at the same time:
* **Pre-register users:** It is by inviting them to this forum in a dedicated group that I was able to find the beta users of the application, those with whom I am building the MVP (Minimum Viable Product), still at the moment I am writing the article.
* **Create a public knowledge base self-fueled by users:** This is the second strong point of this kind of tool. People ask questions, I answer them, and other people who ask themselves the same questions read the discussions to find their answers. It is a self-fueled FAQ that responds to the questions of the users directly, and not to the questions that I think the users are going to ask themselves.
### Deployment and monitoring: Komodo
Deploying applications galore in Docker containers is good. But being able to do it graphically, monitor their health state, perform their updates in a few clicks is better! Not that typing commands by hand annoys me particularly, on the contrary, but it is especially tiring, longer. This is where Komodo intervenes, my control center to pilot the vast majority of my self-hosted applications serenely.
---
## The core of the problem: making public and private coexist
You will have perhaps noticed it, I spoke of services accessible to the public like Discourse, but also of services that must imperatively remain private like Komodo or Keygen. This is why a good cloisonnement of the two (public and private) is needed so as not to expose me, myself and the data of the users, to obvious and important flaws.
## The Core of the Problem: Making Public and Private Coexist
You may have noticed, I talked about services accessible to the public like Discourse, but also services that must imperatively remain private like Komodo or Keygen. That is why a good partitioning of the two (public and private) is necessary so as not to expose myself, and user data, to obvious and significant vulnerabilities.
Here is how I secured and organized this traffic.
### Who orchestrates the traffic? The duo Vercel and Caddy
### Who Orchestrates the Traffic? The Vercel and Caddy Duo
One of the challenges when making several services cohabit on a single server is managing traffic and SSL certificates (HTTPS). In my architecture, I set up a cascading proxy system.
One of the challenges when making several services cohabit on a single server is the management of traffic and SSL certificates (HTTPS). In my architecture, I set up a cascading proxy system.
#### Public Traffic: Vercel at the DNS, Caddy at the Routing
For everything accessible by users (like the Discourse forum), the architecture is as follows:
#### The Public traffic: Vercel in first line, Caddy at the switching
My main domain name is managed at Vercel. I declared a type A record there for public subdomains (for example `forum.thence.app`) pointing to the public IP of my VPS.
For everything that is accessible by users (like the Discourse forum) the path is the following:
It is therefore my VPS that receives the connection. Once the request arrives, Caddy takes over. I chose Caddy, among other things, for its native Automatic HTTPS feature. It automatically provides and renews SSL certificates without my intervention.
Vercel handles the entry. My public domain names point to Vercel. It is its infrastructure that takes the connection first. Vercel handles the public SSL certificate, ensures that the connection is secure, then cleanly redirects the traffic to the IP of my VPS.
Here is an extract from my Caddyfile for the forum:
Once the request arrives on my VPS, it is Caddy that takes the relay. It inspects the subdomain received: if the request targets `forum.thence.app`, Caddy sends it to the Discourse container.
File successfully created: deploying-an-infrastructure-for-a-macos-app-distribution.md
This approach allows me to benefit from the power and security of Vercel in frontal, while keeping total flexibility on my VPS thanks to Caddy to dispatch the traffic to my different Docker containers.
```caddy
forum.thence.app {
reverse_proxy 127.0.0.1:9080 {
header_up Host {host}
header_up X-Real-IP {remote_host}
header_up X-Forwarded-For {remote_host}
header_up X-Forwarded-Proto {scheme}
header_up X-Forwarded-Host {host}
}
#### The Private traffic: The Caddy + Tailscale safe
encode gzip zstd
}
For services that never need to be exposed on the public Web (like my Komodo administration interface), I applied the principle of Zero Trust by combining Caddy and Tailscale. Out of the question that these flows pass through the Internet or through Vercel.
```
Tailscale is a secure mesh VPN based on the WireGuard protocol. By installing Tailscale on my VPS and on my computers, my server obtains a unique private IP address within my secure network (my tailnet), as well as a private domain name.
No mention of certificates to manage, Caddy reads the request, manages HTTPS, transmits the correct headers so the application keeps the origin IP, and redirects the flow to the local port of my Discourse container.
Here, Caddy adopts a totally different behavior:
* I ask it to listen only on the private network interface of Tailscale for these sensitive services.
* Caddy will directly ask for an SSL certificate from the local Tailscale daemon to secure access in HTTPS.
#### Private Traffic: The Caddy + Tailscale Safe
#### The result
For services that never need to be exposed to the public Web (like my Komodo administration interface), I applied the Zero Trust principle by combining Caddy and Tailscale. Out of the question for these flows to pass through the Internet or through Vercel.
If a user or a robot tries to type the URL of my Komodo from the public Web, he runs into a wall: the domain does not exist and the port is closed. For me to be able to administer my infrastructure, I must obligatorily activate Tailscale on my computer. As soon as I am part of the private network, Caddy recognizes my machine, validates the Tailscale HTTPS, and gives me access to my private containers.
Tailscale is a secure mesh VPN based on the WireGuard protocol. By installing Tailscale on my VPS and on my computers, my server gets a unique private IP address within my secure network (my tailnet).
---
For these services, the Caddy configuration adopts a totally closed approach:
## Overview: How it runs day-to-day?
```caddy
vps-939ea86a.tail8644df.ts.net {
tls /etc/caddy/certs/cert.crt /etc/caddy/certs/cert.key
reverse_proxy 127.0.0.1:9120
}
To completely understand how all these bricks cohabit without stepping on each other's feet, nothing beats a good diagram.
```
This block will never respond to a request on the Internet. It only listens on the private domain provided by Tailscale and uses locally generated certificates. If a robot scans my server's public IP, it will find absolutely nothing. To access my administration interface (here on local port 9120), I must imperatively activate Tailscale on my computer.
### Hybrid Cases: The Best of Both Worlds
#### Filtering by HTTP Route (The n8n Example)
Sometimes, a service must be partially public to receive data, but its administration interface must remain strictly inaccessible. This is the case for my n8n instance which must be able to receive webhooks from the outside.
Here is how I secure this flow:
```caddy
n8n.thence.app {
@preflight {
method OPTIONS
path /webhook/*
}
handle @preflight {
header Access-Control-Allow-Origin "[https://thence.app](https://thence.app)"
header Access-Control-Allow-Methods "GET, POST, OPTIONS"
header Access-Control-Allow-Headers "Content-Type, Authorization"
header Access-Control-Max-Age "86400"
respond "" 204
}
handle /webhook/* {
header Access-Control-Allow-Origin "[https://thence.app](https://thence.app)"
header Access-Control-Allow-Methods "GET, POST, OPTIONS"
header Access-Control-Allow-Headers "Content-Type, Authorization"
reverse_proxy [http://100.127.230.106:5678](http://100.127.230.106:5678)
}
handle {
respond "Forbidden" 403
}
}
```
Here, I handle:
* CORS "preflight" (OPTIONS) requests to authorize my main application to communicate with the API
* I authorize the reverse proxy only on the `/webhook/*` path
* The last handle block acts like an `else`: any other request will be returned a 403 Forbidden error. The tool can thus work serenely with the outside without ever exposing its back-office.
Here, I handle:
* CORS "preflight" (OPTIONS) requests to authorize my main application to communicate with the API
* I authorize the reverse proxy only on the `/webhook/*` path
* The last handle block acts like an `else`: any other request will be returned a 403 Forbidden error. The tool can thus work serenely with the outside without ever exposing its back-office.
#### Filtering by Protocol and Network (The Gitea Example)
For my Gitea instance, the hybrid need is different. I want the web interface to be publicly accessible to view what I'm doing, but I also want to lock down write access (clone and push).
To keep control over my source code, the rule is simple: I disable HTTP cloning and force SSH only on the Tailscale domain.
Caddy simply exposes the web interface:
```caddy
git.matheoguilbert.fr {
encode gzip zstd
reverse_proxy 127.0.0.1:3000
}
```
And on the Gitea side, I disable HTTP cloning and specify the SSH domain to use:
```ini
GITEA__server__DISABLE_HTTP_GIT=true
GITEA__server__SSH_DOMAIN=vps-939ea86a.tail8644df.ts.net
GITEA__server__SSH_PORT=222
```
## Overview: How Does It Run Daily?
To understand well how all these bricks cohabit without stepping on each other's toes, nothing beats a good diagram.
```mermaid
graph TD
%% Flux Public
Internet([Le Web Public]) --> Vercel[Vercel]
Vercel --> Caddy
%% Public Flow
Internet([The Public Web]) -.->|DNS Resolution| VercelDNS[Vercel DNS]
Internet -->|Direct HTTPS Traffic| Caddy
%% Flux Privé
MonMac([Mon Mac]) --> Tailscale[Tailscale VPN]
%% Private Flow
MonMac([My Mac]) --> Tailscale[Tailscale VPN]
Tailscale --> Caddy
%% Le VPS et Caddy
subgraph Mon VPS
%% The VPS and Caddy
subgraph Mon VPS [My VPS]
Caddy[Reverse Proxy: Caddy]
%% Dispatching vers Docker
%% Dispatching to Docker
Caddy -->|Public| Discourse[Discourse Forum]
Caddy -->|Public| Keygen[Keygen Licences]
Caddy -->|Privé| Keygen2[Keygen Admin]
Caddy -->|Privé| Komodo[Komodo Admin]
Caddy -->|Public| Keygen[Keygen Licenses]
Caddy -->|Hybrid| N8N[n8n Webhooks]
Caddy -->|Hybrid| Gitea[Gitea Web]
Caddy -->|Private| Komodo[Komodo Admin]
end
```
---
## The monitoring and the backups
## Monitoring and Backups
An infrastructure is only viable if it is monitored and backed up.
* **The administration:** This is where Komodo takes all its meaning. Since my PC (via Tailscale), I have access to a dashboard that allows me to see the health state of each container, to consult the logs in one click and to restart a service if necessary.
* **Administration:** This is where Komodo makes perfect sense. From my PC (via Tailscale), I have access to a dashboard that allows me to see the health status of each container, view logs in one click, and restart a service if necessary.
* **Backup Strategy:** The trap of self-hosting is losing everything if the server crashes. To avoid this, I automated my backups with Restic, an open-source tool that natively encrypts and deduplicates data. Every night, a Bash script performs a hot dump of my databases (via `mongodump`), then Restic directly backs up my Docker volumes (`/var/lib/docker/volumes`) to my S3 storage. The script even includes a retention policy (`restic forget --keep-daily 7`) to automatically purge backups older than a week and optimize storage costs.
* **The Backup strategy:** The trap of self-hosting is to lose everything if the server crashes. I have therefore automated the backup of Docker volumes. Each night, a script encrypts these data and sends them to an S3 compatible object storage.
## The Costs in All This
---
For the server itself, I went through OVH, it's quite reliable and not too expensive. I pay around €10.20 per month.
## The costs in all that
For Thence's public domain name, I went through Vercel, which I use to host the website. I pay €14.99 per year.
* For the server itself, I went through OVH, it is quite reliable and not too expensive. I pay around **10.20€ per month**.
For S3 storage, I have 1 TB with the premium plan of Next.ink, an independent French tech journal. I pay €8 per month.
* For the public domain name of Thence, I went through Vercel, with which I host the website. I pay 14.99€ per year.**.
That makes a total of €33.19 per month, which is quite a low amount for the resource I get with it. It will take a lot of traffic on Thence to reach saturation, and at that point, I don't believe the financial side will be a real obstacle.
* For the S3 storage, I have 1 To with the premium plan of Next.ink, a French and independent tech journal. I pay 8€ per month.**.
On that note, thank you for reading this far and see you next time in another article.
That makes a total of 33.19€ per month, it is a quite low amount for the resource I have with that. It will take a lot of traffic on Thence to arrive at saturation, and at that moment, I do not believe that the financial side will be a real break.
On this, thank you for having read until here and to the next one in another article.
**Mathéo G**
Mathéo G