<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Tomas Suarez on Medium]]></title>
        <description><![CDATA[Stories by Tomas Suarez on Medium]]></description>
        <link>https://medium.com/@datatomas?source=rss-e0afe4ea32bf------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*NdlE02esBv_SHAOLJOyr1A.png</url>
            <title>Stories by Tomas Suarez on Medium</title>
            <link>https://medium.com/@datatomas?source=rss-e0afe4ea32bf------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 03:24:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@datatomas/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Launch Your Website for Free with Cloudflare and Upgrade When You Grow]]></title>
            <link>https://medium.com/@datatomas/launch-your-website-for-free-with-cloudflare-and-upgrade-when-you-grow-4282e821169d?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/4282e821169d</guid>
            <category><![CDATA[cloudflare]]></category>
            <category><![CDATA[cloudflare-workers]]></category>
            <category><![CDATA[email-marketing-lists]]></category>
            <category><![CDATA[cdn]]></category>
            <category><![CDATA[dns]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 18:57:20 GMT</pubDate>
            <atom:updated>2026-11-15T18:57:20.638Z</atom:updated>
            <content:encoded><![CDATA[<p>A practical launch checklist for a small to medium site, from first deploy to a trustworthy custom email address.</p><p>A small website has three separate jobs: serve its pages, handle messages from visitors, and send email from its own domain. Cloudflare can host the site and provide the building blocks for a contact form. A mailbox or email service handles delivery. Keeping those jobs clear makes the setup easier to troubleshoot.</p><ol><li><strong>Deploy the site with Cloudflare Pages</strong></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j68Sqtcbl7Rnp2-IYhMtdQ.png" /><figcaption>Notice fee request is up to 100,000 basically you won’t get charged until your business is relevant which is one of the few perks I love about cloudfare.</figcaption></figure><p>In the Cloudflare dashboard, open Workers &amp; Pages and create a Pages project. Connect a GitHub or GitLab repository if you want each push to trigger a deployment. You can also start with a direct upload for a site that is already built. Choose the deployment method deliberately: Cloudflare does not let a Git-integrated Pages project switch to Direct Upload later.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fhU_2hTdbaNqBv_t9NNZAQ.png" /><figcaption>Cloudflare’s create-app choices. From Gitops to static deployments.</figcaption></figure><p>When authorizing the Git integration, grant access only to the repositories the site needs. Select the production branch, then configure the build using your project’s actual scripts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CSikdKvxIlDC7F8fRFjsLg.png" /><figcaption>Let’s keep it POLP by allowing cloudfare to interact only with needed repositories.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oIUSi_tDmvVR6Lm-5Ac3yQ.png" /><figcaption>For a typical Vite app, the build command is npm run build and the output directory is dist; another framework may use a different command or directory. If the site lives inside a larger repository, set its root directory as well. Check the generated output locally before deploying.</figcaption></figure><p>After the first build, open the pages.dev preview URL and test the site on both desktop and mobile. Confirm that links, images, and any client-side routes work. If the build fails, read the deployment log first: a wrong output directory, missing environment variable, or mismatched Node version is often the cause.</p><p>Where Workers fit: Pages serves the site. A Pages Function or separate Worker can run server-side code for a form or an API. A plain static page does not need a Worker simply to exist.</p><p><strong>2. Add your domain and configuration</strong></p><p>In the Pages project’s Custom domains settings, add the domain you own and follow Cloudflare’s DNS instructions. Verify both the root domain and www if you plan to use both, and choose one canonical address for links. Leave mail-related MX records alone unless you are intentionally changing who receives your email.</p><p>Use environment variables for configuration that differs between preview and production. Keep private API credentials in the project’s secrets or encrypted variable settings, never in client-side JavaScript or a public repository. A contact form might need a FORM_DESTINATION address on the server side; exposing that variable to browser code defeats the purpose.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L8XTvlMb1RqQNzdELbOoVQ.png" /><figcaption>Make sure to have previously bought a domain so you can use it as custom domain for your worker.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jXViXb1jotXsNfHbidZhjg.png" /><figcaption>When adding email forwarding for a domain address or setting up a contact form, add a FORM_DESTINATION variable with the real inbox where you want messages delivered, such as Zoho or Gmail. The EMAIL binding sends contact form notifications, while FORM_LIMITER helps limit spam. To forward mail sent directly to your domain address, create a separate Cloudflare Email Routing rule.</figcaption></figure><p>A form needs more than an HTML submit button. Add a Pages Function or Worker endpoint that validates the fields, limits their length, and rejects malformed requests. Add spam protection such as Cloudflare Turnstile and rate limiting before sending or forwarding messages. Return a clear success or error state to the visitor, and test that a submitted message reaches the intended inbox.</p><p>Cloudflare Email Routing can forward incoming mail sent to an address such as hello@example.com to a mailbox you already use. It is an incoming-mail tool; it does not by itself send your form submissions or newsletters. For outgoing form notifications, connect an email provider or another supported delivery service to your server-side endpoint. Keep that provider’s key in a secret.</p><p>If you want to send from hello@example.com, first verify the domain with your mailbox or sending provider. It will supply DNS records and a verification token. Copy its exact host names and values. The examples below use example.com and placeholders; never publish real tokens or API keys.</p><p>• SPF lists authorized senders in a TXT record. Publish only one SPF record at a given domain name. If several services send mail, combine their entries according to the providers’ instructions.</p><p>• DKIM signs outgoing messages. Enable it in the sending provider and add the selector records it gives you, usually CNAME or TXT records.</p><p>• DMARC checks whether SPF or DKIM aligns with the visible From domain. Start with a monitoring TXT record at _dmarc.example.com, such as v=DMARC1; p=none; rua=mailto:dmarc-reports@example.com. Review the reports before moving to p=quarantine or p=reject.</p><p>If a newsletter platform sends on your behalf, verify its custom sending domain too. Send a test campaign and inspect the message headers for SPF, DKIM, and DMARC results. A passing SPF check alone may still fail DMARC if its domain does not align with the visible From address.</p><p><strong>Treat BIMI as an optional finishing touch</strong></p><p>BIMI can let participating inboxes display your brand logo beside authenticated mail. It does not improve delivery by itself, and display is never guaranteed. First get SPF, DKIM, and aligned DMARC working. BIMI requires DMARC enforcement (p=quarantine or p=reject, including relevant subdomains), a compliant SVG Tiny PS logo hosted at a stable HTTPS URL, and a BIMI TXT record at default._bimi.example.com.</p><p>Some inbox providers also require a Verified Mark Certificate or Common Mark Certificate. Follow the current requirements of your target inbox provider and certificate issuer before paying for one. If you have a certificate, the BIMI record can point to it with an a= tag; otherwise use only the fields your provider supports.</p><p><strong>Before you call it done</strong></p><p>Before the final launch check, enable observability for your Worker. Logs, traces, and issues help you see whether form submissions reach the Worker and where a failed request stops.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SiQQ6gnB9MAkQbOA4D93aA.png" /><figcaption>Worker observability: Turn on logs and traces, deploy the changes, then submit a test form. Check the results here if the message does not reach your inbox.</figcaption></figure><p>• [ ] A push to the production branch deploys the expected build.</p><p>• [ ] The custom domain resolves and the site works on mobile.</p><p>• [ ] The contact form rejects bad input and spam, then delivers a real test message.</p><p>• [ ] Secrets stay in server-side configuration.</p><p>• [ ] The sending domain is verified; SPF, DKIM, and DMARC pass on a real test email.</p><p>• [ ] Mail MX records still point to the intended inbox provider.</p><p>A launch is complete when the full path works: a visitor can load the site, send a message, and receive a reply from an authenticated address. Test that path end to end before sharing the URL.</p><p>Sources and further reading</p><p>• <a href="https://developers.cloudflare.com/pages/get-started/git-integration/">Cloudflare Pages: Git integration</a></p><p>• <a href="https://developers.cloudflare.com/pages/configuration/build-configuration/">Cloudflare Pages: Build configuration</a></p><p>• <a href="https://developers.cloudflare.com/workers/platform/limits/">Cloudflare Workers: Limits and pricing</a></p><ul><li><a href="https://bimigroup.org/implementation-guide/">BIMI Group: Implementation guide</a></li></ul><p>Connect with me:</p><p><strong>🥊</strong><a href="https://fightprophet.com/"><strong>Fight Prophet — EDGE AI &amp; MMA Intelligence</strong></a><strong><br></strong>💼 [Upwork: Tomas Suarez](<a href="https://www.upwork.com/freelancers/~01">Link)</a><br> 🌐 [Freelancer: datatomas](<a href="https://www.freelancer.com/u/datatomas">Link</a>)<br> 💻 [GitHub: datatomas](<a href="https://github.com/datatomas">datatomas (Tomas Suarez)</a>)<br> 🧠 [Substack: datatomas](<a href="https://datatomas.substack.com/">Link</a>)<br> 🔗 [LinkedIn: Tomas Suarez] (<a href="https://www.linkedin.com/in/datatomas">Link</a>)<br> 📰 [<a href="https://medium.com/@datatomas">Medium: @datatomas</a>] (<a href="https://medium.com/@datatomas">Link</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4282e821169d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Wiring a Nano GPS to a micro drone and enabling GPS rescue]]></title>
            <link>https://medium.com/@datatomas/wiring-a-nano-gps-to-a-micro-drone-5c88b958b5a5?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c88b958b5a5</guid>
            <category><![CDATA[betaflight]]></category>
            <category><![CDATA[flywoo]]></category>
            <category><![CDATA[drones]]></category>
            <category><![CDATA[fpv]]></category>
            <category><![CDATA[gps]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 04:39:07 GMT</pubDate>
            <atom:updated>2026-10-29T17:08:57.706Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*WRSfkzPX7rP-dU9fOEjMvg.png" /><figcaption>Adding GPS to a frame under 3 inches is as much about finding space as connecting wires. On Bjorn 2.5, we paired a Flywoo Nano GPS with the SpeedyBee F405 Mini flight controller in the mini stack, moved the battery underneath, and kept room above the frame for the GPS. (Buy the gps here <a href="https://flywoo.net/products/goku-gm10-nano-v3.1-gps?utm_source=chatgpt.com">https://flywoo.net/products/goku-gm10-nano-v3.1-gps</a>)</figcaption></figure><p>Small frame, six wires: fitting a Flywoo Nano GPS to Bjorn 2.5, from wiring and battery placement to the Betaflight OSD.</p><h3>Why the Nano fits</h3><p>The Flywoo GOKU GM10 Nano is a compact option for small builds. Flywoo lists the Nano V3 with compass at 12 × 17 × 5 mm and 2.6 g, compared with an 18 × 18 mm footprint for its GM10 Mini models. That smaller footprint helps on sub-3-inch frames where the camera, video antenna, and battery strap already compete for space. The larger model in this comparison is officially called Mini, rather than Micro; sizes across other brands vary.</p><p>The wiring diagram below covers the six-wire Nano V3 with compass. Check the actual module marking before soldering: the four-wire Nano V3.1 has no SDA/SCL compass connection.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HcaRfJel8mLBzpee5gM3NA.png" /><figcaption>With power disconnected and props removed, connect GPS SDA (SD) to the FC’s DA pad, SCL (SC) to CL, VCC (5V) to 4V5, and GND to G. Cross the serial signals: GPS TX goes to R2, and GPS RX goes to T2. The compass uses I2C on DA/CL; the GPS position data travels over UART2.</figcaption></figure><p>We used R2/T2 on Bjorn 2.5, as reported by the builder. SpeedyBee’s standard GPS illustration uses UART6, so match Betaflight’s GPS port to the UART actually wired. The diagram is a connection schematic, not a physical pad-order drawing. Flywoo specifies a 3.3–5 V supply for this Nano variant, and SpeedyBee shows its 4V5 rail for GPS power. Verify labels, polarity, and the installed board revision before connecting power.</p><h3>Battery above or below?</h3><p>The top-versus-bottom choice here concerns the battery. A top-mounted pack occupies valuable space above the frame and needs a strap route that leaves the GPS antenna clear. Moving the pack underneath, as shown in the photo, frees that upper space for the GPS installation.</p><p>A bottom-mounted pack needs secure retention, protection from landing impacts, and enough ground clearance. Check balance with the actual battery installed. This layout gives us more room up top; improved handling or GPS reception has not yet been measured.</p><h3>A printed holder or tape for the first test</h3><p>A small 3D-printed GPS holder is useful for repeatable placement, strain relief, and keeping the module clear of the battery strap. Fit it to the exact Nano variant and leave the antenna facing the sky.</p><p>For a temporary local reception test, double-sided mounting tape can hold the GPS in place without waiting for a print. Use an electrically insulating mounting layer, keep exposed contacts away from conductive carbon, and secure the wires so they cannot pull the module loose. Check adhesion before each test. A taped installation is a practical way to assess placement during stationary, props-removed tests; it does not establish a secure flight mount.</p><h3>Check reception and the OSD</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dmet7Xs0HTmqKWajh-16VA.png" /><figcaption>betaflight osd config</figcaption></figure><p>After checking the wiring, open the <strong>Ports</strong> tab in Betaflight and select <strong>GPS</strong> under Sensor Input for the UART connected to the module. On Bjorn 2.5, we used <strong>UART2</strong>, with GPS TX connected to R2 and GPS RX connected to T2. Save and reboot, then open the <strong>GPS</strong> tab. Enable <strong>GPS for navigation and telemetry</strong>, select <strong>UBLOX</strong> as the protocol, and turn <strong>Auto Config ON</strong>. UBLOX is the communication format between the Flywoo receiver and the flight controller; Auto Config lets Betaflight configure the receiver automatically.</p><p>In the same GPS tab, turn <strong>Use Galileo ON</strong>. Galileo is Europe’s global satellite navigation system, and its signals are available worldwide, including Colombia. The Flywoo GM10 Nano supports Galileo alongside GPS and other satellite constellations. Giving the receiver additional satellites to work with can improve satellite availability and position reliability, particularly when part of the sky is obstructed. It does not guarantee a faster fix or eliminate interference, so antenna placement and a clear sky view remain important. No separate receiver, subscription, or compass is needed to receive Galileo signals.</p><p>For our initial reception test in Colombia, we left <strong>Ground Assistance Type</strong> at <strong>None</strong>. This setting controls optional regional SBAS corrections and is separate from Galileo. Save and reboot again, then test outdoors with the props removed. In the GPS tab, wait for <strong>3D Fix = True</strong>, check that the satellite count increases, and watch the positional PDOP decrease from its initial value. A satellite row marked “locked” alone does not establish a valid position fix. Once reception is working, move on to displaying the GPS information in the OSD.</p><p>The GPS latitude and longitude elements can then be added to the OSD. Betaflight provides several coordinate display formats, and the choice only changes how the same position is shown on screen:</p><p><strong>Decimal degrees with 7 decimal places</strong> provides the most detailed numeric representation, for example 4.6533321, -74.0836527. Seven decimal places represent a theoretical coordinate resolution of roughly one centimeter, although this does not mean the GPS itself is accurate to one centimeter. A small drone GPS will normally have real-world positioning uncertainty measured in meters.</p><p><strong>Decimal degrees with 4 decimal places</strong> produces a shorter coordinate such as 4.6533, -74.0837. This is easier to fit into a crowded FPV OSD but gives roughly 11-meter coordinate resolution. It is useful when the goal is quickly identifying the approximate location of the aircraft rather than recording a highly detailed coordinate.</p><p><strong>Degrees, minutes and seconds (DMS)</strong> displays a traditional geographic coordinate such as 4°39&#39;12.0&quot;N 74°05&#39;01.1&quot;W. This format is familiar from maps, aviation and navigation, but it consumes more OSD space than decimal degrees.</p><p><strong>Open Location Code</strong>, also known as a <strong>Plus Code</strong>, converts the location into a compact location identifier rather than displaying raw latitude and longitude. This can be convenient when communicating a location to another person because the code is shorter and less error-prone to read over radio or type manually.</p><p>For an FPV or recovery-oriented build, <strong>7-digit decimal degrees</strong> is a good default because the coordinates can be copied directly into mapping software. If OSD space is limited, the <strong>4-digit format</strong> is considerably cleaner. Plus Codes are particularly useful when the goal is quickly sharing the drone’s location with another person.</p><p>The OSD can also display <strong>satellite count and GPS ground speed</strong> alongside latitude and longitude. Remember that the values displayed in the Betaflight OSD preview are examples used to position the elements; they are not a live GPS fix.</p><p>Finally, test the installation outdoors with an unobstructed view of the sky. Wait for a <strong>3D fix</strong>, confirm that latitude and longitude update correctly, and verify satellite count and speed before relying on the GPS for flight features such as GPS Rescue. If the GM10 version includes a compass, compass detection, orientation and calibration should also be verified separately.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8z6AV1V4G3uIK98puPXNjA.png" /><figcaption>image taken from flywoo.com all their gps and wiring</figcaption></figure><h4>Bonus: GPS rescue and newer hold modes</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kT4UOH7p658w-sBjY-kU5w.png" /><figcaption>Betaflight Modes enabled in 2026.2 firmware Alt hold, pos hold and with gps gps rescue.</figcaption></figure><p>Betaflight supports GPS Rescue from version 4.4 onward. In the <strong>Failsafe</strong> tab, set the Stage 2 procedure to <strong>GPS Rescue</strong> if you want the quad to use Rescue after losing radio control. You can also assign GPS Rescue to a switch in the <strong>Modes</strong> tab for testing or manual activation. Follow the <a href="https://github.com/betaflight/betaflight.com/blob/master/docs/wiki/guides/current/GPS-Rescue.md">Betaflight GPS Rescue guide</a> to configure and tune it for your quad.</p><p>With GPS Rescue configured, Betaflight normally prevents arming until the GPS has a valid position fix and reaches the configured minimum satellite count. That wait is intentional: Rescue needs a recorded home point to navigate back to.</p><p>Betaflight also provides an <strong>Allow arming without fix</strong> option. Use caution: if you arm before a home point is set, GPS Rescue cannot return the quad home if failsafe occurs. Betaflight’s guide warns that Rescue may instead perform a landing sequence and disarm. For a recovery setup, wait for a valid fix before arming.</p><p>Betaflight 2025.12 introduced <strong>Altitude Hold</strong> and <strong>Position Hold</strong>. If these modes appear in the <strong>Modes</strong> tab after updating, assign each to an AUX switch range and save the configuration. Altitude Hold needs an accelerometer and an altitude source. Position Hold needs GPS and an accelerometer; a working, calibrated magnetometer is strongly recommended. Test in a safe, open area because hold modes change how the quad responds to the sticks. <a href="https://betaflight.com/docs/wiki/guides/current/Position-Hold-2025-12">Position Hold guide</a></p><p>In the screenshot, <strong>POS HOLD</strong> is assigned to AUX 3 over the 1700–2100 range. <strong>GPS RESCUE</strong> is assigned to AUX 4 over the same range. Check that these switch ranges match your radio, and avoid putting failsafe activation on a switch range shared with another mode.</p><h3>Connect with me</h3><p>Thanks for reading. I write about practical engineering across <strong>Linux, Kubernetes, cloud infrastructure, DevOps, Python automation, AI/ML systems, and hardware projects </strong>, usually based on things I am actually building, testing, breaking, and fixing.</p><p>If you found this useful, you can follow my work here:</p><p>💻 <strong>GitHub:</strong> <a href="https://github.com/datatomas">datatomas</a><br>🔗 <strong>LinkedIn:</strong> <a href="https://www.linkedin.com/in/datatomas">Tomas Suarez</a><br>🧠 <strong>Substack:</strong> <a href="https://datatomas.substack.com/">datatomas</a><br>💼 <strong>Upwork:</strong> <a href="https://www.upwork.com/freelancers/~01">Tomas Suarez</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c88b958b5a5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Basics configs for a LuckFox Pico]]></title>
            <link>https://medium.com/@datatomas/getting-a-shell-on-a-luckfox-pico-0c49a40d7ed1?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c49a40d7ed1</guid>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[luckfox]]></category>
            <category><![CDATA[picoctf]]></category>
            <category><![CDATA[buildroot]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 03:50:59 GMT</pubDate>
            <atom:updated>2026-10-28T03:08:07.297Z</atom:updated>
            <content:encoded><![CDATA[<p>A LuckFox Pico Mini B is a $10-ish ARM Linux board. You plug it into your<br>laptop with a USB cable, and it becomes a network card. No WiFi, no Ethernet, no router. Everything below is plain Linux, the same commands work on any distro.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TuXH65APWSyl6z7J4a2O7w.png" /></figure><p>The LuckFox Pico runs <strong>Buildroot</strong>, not Debian or Ubuntu. That means it behaves more like a minimal embedded Linux device than a Raspberry Pi-style general-purpose computer.</p><p>The main differences are:</p><ul><li>It uses BusyBox /etc/init.d scripts instead of systemd.</li><li>There is no apt install; software is usually copied directly to the device or included when rebuilding the image.</li><li>You normally work as root, so there is no typical sudo workflow.</li><li>RAM is extremely limited at around <strong>32 MB</strong>, compared with gigabytes on a Raspberry Pi or Orange Pi.</li><li>Networking is much more minimal.</li><li>Initial access is commonly through ADB/USB rather than SSH.</li></ul><p>Because of this, some common Pi habits do not transfer directly.</p><p>There is no raspi-config, no nmtui, no .local hostname discovery by default, and SSH may not be enabled initially.</p><p>For this project, that minimal environment is actually useful. The goal is not to turn the LuckFox into a full Linux server, but to run only the components we need: the camera pipeline, the small C AprilTag detector, and communication with the flight controller.</p><h3>Step 1: confirm the board appears over USB</h3><p>Connect the LuckFox and check the available interfaces:</p><pre>ip -br link<br>dmesg | tail -20</pre><p>The board should appear as a USB-connected device or network interface depending on the firmware configuration.ip -br link<br>dmesg | tail -20</p><p>You will see something like `enxc21d3abb330b` plus a `cdc_ether`/`rndis_host` line in `dmesg`. That is the board’s USB Ethernet gadget. The board’s side of<br>the link is `usb0`, with the address `172.32.0.93/24` baked into the firmware.</p><p><strong>Two things bite here:</strong></p><p>The interface name changes every boot. The board has no MAC address<br>burned in, so it makes one up at each boot (`using random self ethernet<br>address`). Linux names USB network cards `enx` + the MAC, so a profile bound to `enxc21d3abb330b` breaks the next time you unplug it.</p><p>There is no DHCP server on either end. Nobody hands out an address. You have to set one by hand on your machine.</p><h4>Step 2: give your laptop an address on the same subnet</h4><pre>sudo ip addr add 172.32.0.100/24 dev enxc21d3abb330b<br>sudo ip link set enxc21d3abb330b up<br>ping -c3 172.32.0.93</pre><p>That is the whole network. `172.32.0.93` is the board, `172.32.0.100` is you.</p><h4>Step 3: get in ADB, not SSH</h4><p>ADB is Android’s debug bridge, and the LuckFox firmware runs `adbd` on the USB gadget. The name says Android; this board is not Android. It just uses the transport.</p><pre>sudo apt install adb<br>adb devices                  # board appears as a USB serial number<br>adb -d shell                 # root shell, no login, no password<br>adb -d push file.txt /mnt/sdcard/</pre><p>`-d` means “the one device on USB”. Once the network is up you can also go over<br>TCP, which is faster and lets you reconnect across a USB re-plug:</p><pre>adb connect 172.32.0.93:5555<br>adb -s 172.32.0.93:5555 shell</pre><p>If `adb connect` says `already connected` and then `device offline`, that is a stale entry from the last boot, not a broken board:</p><pre>adb disconnect 172.32.0.93:5555<br>adb kill-server<br>adb connect 172.32.0.93:555</pre><h4>3.1) UART connection to an ESP32</h4><p>The UART pins can be enabled and wired correctly, but communication still fails if the LuckFox and ESP32 use different baud rates. The baud rate is the speed of the wired serial link. Both ends must use the same settings; for example if your project uses <strong>115200 baud, 8 data bits, no parity, 1 stop bit (8N1)</strong>.</p><p>First, check which UART devices exist and how the pins are assigned:</p><pre>ls -l /dev/ttyS*<br>luckfox-config show</pre><p>To enable UART3 on a LuckFox Pico Mini A/B, run:</p><pre>luckfox-config</pre><p>Select <strong>Advanced Options → UART → UART3_M1 → Enable</strong>, then save or apply the configuration when prompted. On the Mini A/B mapping used in this project, GPIO1_D0 is UART3 TX and GPIO1_D1 is UART3 RX. Apply the saved configuration and verify the device:</p><pre>luckfox-config load<br>ls -l /dev/ttyS3<br>luckfox-config show</pre><p>Check the displayed pin map for your board before wiring; options and pin assignments can vary by model.</p><p>Configure the LuckFox UART and send a test line:</p><pre>stty 115200 raw -echo &lt; /dev/ttyS3<br>stty -a &lt; /dev/ttyS3<br>printf &#39;HELLO FROM LUCKFOX\n&#39; &gt; /dev/ttyS3</pre><p>stty -a should show speed 115200 baud and -echo. Connect LuckFox TX to ESP32 RX, LuckFox RX to ESP32 TX, and connect the grounds.</p><p>The ESP32 UART must also be started at 115200, for example:</p><pre>uartLink.begin(115200, SERIAL_8N1, RX_PIN, TX_PIN);</pre><p>Here, 115200 means 115,200 baud; it does not mean 1,152,000. Baud rate applies to the wired UART link. ESP-NOW is the wireless link from the first ESP32 to the second.</p><p>If /dev/ttyS3 does not exist, check the enabled UART configuration and device paths first.</p><h4>Step 4: make the board’s address survive reboot</h4><p>The `172.32.0.93` address lives on `usb0` and is added at boot by the image<br>but not reliably, and the board has no clock and no persistence you can trust. Buildroot runs every executable `/etc/init.d/S*` script at boot in numeric order. Drop in `S99` and you are last in line:</p><pre>cat &gt; /etc/init.d/S99usbnet &lt;&lt;&#39;EOF&#39;<br>#!/bin/sh<br>[ -d /sys/class/net/usb0 ] || exit 0<br>ip link set usb0 up<br>ip addr add 172.32.0.93/24 dev usb0 2&gt;/dev/null<br>exit 0<br>EOF<br>chmod +x /etc/init.d/S99usbnet</pre><p>`2&gt;/dev/null` because `ip addr add` fails on the second boot if the address is already there, and that is fine. `[ -d /sys/class/net/usb0 ] || exit 0` because<br>if the USB cable is unplugged the interface does not exist and the script must not block boot.</p><p>Verify without rebooting:</p><pre>sh /etc/init.d/S99usbnet<br>ip -4 addr show usb0</pre><p>This survives reboot because `/etc` is writable on the UBI root filesystem.<br>The risk is a firmware reflash, which wipes it.</p><h4>Step 5: make your laptop’s side persistent without auto-connecting</h4><p>The manual ip addr add command from Step 2 disappears after reboot, and the enx... interface name can change because the LuckFox presents a random MAC address.</p><p>On systems using NetworkManager, create a connection profile that matches the USB RNDIS driver instead of a specific interface name.</p><p>One important choice here: <strong>do not let NetworkManager automatically activate the profile.</strong> USB RNDIS devices can also be phones or other USB gadgets, so it is safer to enable the LuckFox connection only when you actually need it.</p><p>Create the profile:</p><pre>sudo tee /etc/NetworkManager/system-connections/luckfox-usb.nmconnection &gt;/dev/null &lt;&lt;&#39;EOF&#39;<br>[connection]<br>id=luckfox-usb<br>type=ethernet<br>autoconnect=false<br>match-device=driver:rndis_host</pre><pre>[ipv4]<br>method=manual<br>address1=172.32.0.100/24<br>never-default=true<br>dns-search=</pre><pre>[ipv6]<br>method=disabled<br>EOF</pre><p>Lock down the file permissions and reload NetworkManager:</p><pre>sudo chmod 600 /etc/NetworkManager/system-connections/luckfox-usb.nmconnection<br>sudo nmcli connection reload</pre><p>Now bring the LuckFox link up manually whenever the board is connected:</p><pre>sudo nmcli connection up luckfox-usb</pre><p>The important options are:</p><ul><li>match-device=driver:rndis_host — matches the USB network driver instead of a changing enx... interface name.</li><li>autoconnect=false — prevents NetworkManager from automatically activating this profile whenever an RNDIS device appears.</li><li>method=manual — uses the static address 172.32.0.100/24 instead of waiting for DHCP.</li><li>never-default=true — prevents the LuckFox USB link from becoming your machine’s default Internet route.</li><li>dns-search= — prevents DNS configuration from being associated with this isolated link.</li><li>method=disabled under IPv6 — disables IPv6 for this connection.</li><li>chmod 600 — required because NetworkManager ignores keyfiles that are readable by other users.</li></ul><p>You can verify that autoconnect is disabled:</p><pre>nmcli -f NAME,AUTOCONNECT connection show | grep luckfox</pre><p>Expected output:</p><pre>luckfox-usb      no</pre><p>You can also verify that your real Internet connection still owns the default route:</p><pre>ip route | grep default</pre><p>The default route should point to your normal Wi-Fi or Ethernet interface, not the LuckFox USB interface.</p><p>The LuckFox should only add a local route similar to:</p><pre>172.32.0.0/24 dev enx...</pre><p>From now on, the usual connection sequence is:</p><pre>sudo nmcli connection up luckfox-usb<br>adb connect 172.32.0.93:5555<br>adb -s 172.32.0.93:5555 shell</pre><p>This keeps the LuckFox connection persistent across reboots while still requiring you to enable it explicitly.</p><p><strong>Putting it together</strong></p><pre>sudo nmcli connection up luckfox-usb<br>adb connect 172.32.0.93:5555<br>adb -s 172.32.0.93:5555 shell</pre><p>Debian side: static `172.32.0.100/24` matched by driver. Board side:<br>`172.32.0.93/24` re-added at boot by `S99`. ADB as the transport. No router,<br>no DHCP, no SSH keys, no passwords.</p><p><strong>The one gotcha worth remembering</strong></p><p>Every LuckFox Pico ships with the same default 172.32.0.93. Two boards plugged in at once would therefore conflict if both kept the factory address.</p><p>For a second board, give it a different subnet. For example:</p><pre>Board 1: 172.32.0.93<br>Board 2: 172.33.0.93</pre><p>Each board’s /etc/init.d/S99bjorn-usbnet script should configure its own address at boot.</p><p>With two boards attached, the <strong>endpoint becomes the identity</strong>. Instead of relying on a changing USB interface name or random MAC address, you know which board you are talking to from the IP endpoint it reports.</p><p>You can verify each one directly:</p><pre>adb -s 172.32.0.93:5555 shell &#39;ip -4 addr show usb0&#39;<br>adb -s 172.33.0.93:5555 shell &#39;ip -4 addr show usb0&#39;</pre><p>The first command should show:</p><pre>172.32.0.93/24</pre><p>and the second:</p><pre>172.33.0.93/24</pre><p>That gives each board a stable identity even though its USB MAC address and Linux enx... interface name may change between boots.</p><p>On the laptop side, create a corresponding NetworkManager profile for each subnet:</p><pre>Board 1 → laptop 172.32.0.100/24<br>Board 2 → laptop 172.33.0.100/24</pre><p>If both boards are connected simultaneously, also pin each NetworkManager profile to the appropriate USB port path rather than matching only driver:rndis_host. Otherwise both profiles may try to match both boards.</p><p>The important idea is:</p><pre>USB name / MAC      = temporary<br>board IP endpoint   = stable identity</pre><p>So once the per-board boot configuration is in place, commands such as:</p><pre>adb -s 172.32.0.93:5555 shell<br>adb -s 172.33.0.93:5555 shell</pre><p>give you a predictable way to address each LuckFox individually.</p><h4>Other useful commands</h4><pre># Get the luckfox model you have<br>cat /proc/device-tree/model<br># enter luckfox configs<br>luckfox-config</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c49a40d7ed1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Kubernetes over-reservation tax explained]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@datatomas/the-kubernetes-over-reservation-tax-explained-ae4363dab79c?source=rss-e0afe4ea32bf------2"><img src="https://cdn-images-1.medium.com/max/1615/1*7u7eT3HNy36ZDeU2BW-5_Q.png" width="1615"></a></p><p class="medium-feed-snippet">Kubernetes schedules on promises, not reality. Here&apos;s how Grafana + Prometheus showed me my cluster was &quot;full&quot; at 0.2% actual usage.</p><p class="medium-feed-link"><a href="https://medium.com/@datatomas/the-kubernetes-over-reservation-tax-explained-ae4363dab79c?source=rss-e0afe4ea32bf------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@datatomas/the-kubernetes-over-reservation-tax-explained-ae4363dab79c?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae4363dab79c</guid>
            <category><![CDATA[grafana]]></category>
            <category><![CDATA[prometheus]]></category>
            <category><![CDATA[scheduler]]></category>
            <category><![CDATA[k3]]></category>
            <category><![CDATA[kubernetes]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Sat, 17 Oct 2026 17:31:04 GMT</pubDate>
            <atom:updated>2026-10-17T17:54:37.745Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Micro drones, big lessons]]></title>
            <link>https://medium.com/@datatomas/micro-drones-big-lessons-b042a4d1a439?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/b042a4d1a439</guid>
            <category><![CDATA[drones]]></category>
            <category><![CDATA[flight-controller]]></category>
            <category><![CDATA[fpv]]></category>
            <category><![CDATA[betaflight]]></category>
            <category><![CDATA[microdrones]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 23:31:01 GMT</pubDate>
            <atom:updated>2026-10-10T16:34:41.220Z</atom:updated>
            <content:encoded><![CDATA[<p>A short, hands-on guide to getting a Mobula 7, a 3-inch Apex and a 2.5-inch “Mini Bjorn” build set up in Betaflight.</p><p>Workshop story, not a sponsored review. These are our builds and mistakes, not universal settings. I kept learning the same lesson: a quad that will not lift or flies oddly may need a prop, motor-map, wiring or firmware check before it needs a new PID tune.</p><h3>Start with the motor diagram, not the diagonal</h3><p>On both our Mobula 7 and Apex 3, I ended up flying with motor rotation reversed in Betaflight (the props-out arrangement). In our setup, motors 1 and 4 used the regular propellers, while motors 2 and 3 used the reverse-marked propellers. I initially expected the opposite pattern. I put the wrong prop type on the 1/4 diagonal, and the quad tipped/flipped when I tried to lift it. The problem was the motor/prop match, not the PID values.</p><p>Betaflight’s “Motor direction is reversed” option changes the mixer’s expected yaw direction; it does not reverse ESC motor spin or fix motor order. With propellers removed, check that each Motors-tab slider spins the motor shown in its diagram, check the actual spin direction, then fit the matching propeller. The Happymodel Mobula 7 manual shows the 1/4 versus 2/3 prop pattern for its props-out example. Use the diagram and manual for the exact aircraft in front of you: prop direction is not a rule to copy betIen targets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/948/1*C2WgfoJqlhKESDGLkeIawA.png" /><figcaption>In my case while using the Diatone mamba mk4 f722 miniflight controller I needed also to reorder motors since I had to pointr it backwards to have the perfect fit on the apex 3. SO make sure to click on reorder motors within betaflight because they will be inverted.</figcaption></figure><h3>1S, 4S are different battery systems</h3><p>The Mobula 7 is a 1S micro whoop. “S” counts cells in series: a standard LiPo 1S pack is 3.7 V nominal and 4.2 V fully charged each flight can take at most 5–10 minutes with the gaoneng gnb 550mAh 1s 3.8v hv 100C lightIigh lipo I bought from ali express. ; A 4S pack is 14.8 V nominal and 16.8 V fully charged, which is giving me around 15–20 minutes flights. Our Betaflight screenshots show the contrast: about 3.78 V (1S) on the Mobula and 14.89 V (4S) on the 3-inch build. The voltage is not a matter of tuning use only the cell count and chemistry supported by that aircraft, ESC and motors. LiHV packs have a different full-charge voltage, so choose the matching charger setting.</p><p>My Mobula is about 26 g all-up; the 3-inch Apex build is over 233 g, based on our reported build Iights. The Mobula has taken a surprising number of small knocks in our use. The heavier Apex carries more momentum at the same speed, so a crash can transfer more energy into the frame, motors, camera and electronics. That is our experience with these builds, not a promise that a whoop cannot break or that every heavier quad will. The settings I’ve been ussing on my lipo charger are</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RwcvR2PQnA3own-WQEBmCQ.png" /><figcaption>Apex 3&#39;’ with 580 4s Battery 233 gr</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mu2msmHXukF39wncRBqbng.png" /><figcaption>Mobula 7 weight (w/o battery )</figcaption></figure><p>charging settings</p><p>Mobula 7 – 1S 550 mAh LiHV</p><ul><li>Nominal voltage: 3.8 V</li><li>* Charge mode: 1S LiHV</li><li>* Charge current: 0.55 A</li></ul><p>3S 580 mAh LiPo</p><ul><li>Nominal voltage: 11.1 V</li><li>* Charge mode: 3S LiPo</li><li>* Charge current: 0.4 A</li></ul><p>4S 580 mAh LiPo</p><ul><li>Nominal voltage: 14.8 V</li><li>* Charge mode: 4S LiPo</li><li>* Charge current: 0.6 A</li></ul><p>Quick reference:</p><p>S = number of cells in series</p><p>V = voltage</p><p>mAh = battery capacity</p><p>A = charging current</p><p>The Mobula battery is LiHV, so use the charger’s LiHV mode, not standard LiPo mode.</p><h3>Firmware belongs to the flight controller</h3><p>Our Mobula 7 uses the HAMO CRAZYBEEF4SX1280 target and a built-in ELRS receiver. The saved configurator screenshots show Betaflight 2026.6.2, while the local HEX filename says 2026.6.1; those are separate records, not proof that the image and screenshot are the same build. The build notes also call out selecting the gyro driver that matches the actual board revision.</p><p>The Apex screenshot shows target MAMBAF722_2022A running Betaflight 2026.6.2; an earlier screenshot shoId 4.5.1. The Bjorn 2.5 notes record a SpeedyBee F405 Mini running 4.4.2. When flashing, check that the selected firmware version supports the gyro on that specific board. For our Mobula 7’s CRAZYBEEF4SX1280 build, the gyro custom define I recorded was ACCGYRO_LSM6DSV16X. Confirm the board revision and gyro detection before flying, and don’t transfer a full CLI configuration betIen aircraft: their targets, wiring, ports, receivers and modes differ.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hQA4JyDOj5EvP3itwXfEzg.png" /><figcaption><em>Firmware Flasher: choose the exact target and build options for the flight controller and gyro.</em></figcaption></figure><h3>Ports, radio and video: follow the signal path</h3><p>The Mobula’s receiver is built in; our Apex and Bjorn builds use external ExpressLRS receivers configured for CRSF. For a UART receiver, connect receiver TX to flight-controller RX and receiver RX to flight-controller TX, with a shared ground and the correct supply voltage. In Ports, enable Serial RX on the UART actually used; in Receiver, choose Serial via UART and CRSF. Then confirm stick bars and channel order in Receiver before arming. A built-in SPI receiver follows a different setup pat, do not invent a UART for it.</p><p>For a UART-controlled analog VTX, SmartAudio uses a flight-controller TX pad connected to the VTX control wire. Select SmartAudio on that UART in Ports, then load the table for the exact VTX variant and verify channel and poIr in Video Transmitter. Our Mobula record uses Raceband R8 at 5917 MHz; its table’s poIr label is a configuration value, not a measured RF output. The Rush Tiny Tank and Rush Tank Mini are different VTX models, with different wiring, voltage and poIr details. Check the label and manufacturer manual before applying a table. Keep an antenna attached when poIring an analog VTX, and follow local frequency and poIr rules.</p><p>The firmware, Ports, Video Transmitter and OSD screens below make a useful setup trail: choose the target, assign the UARTs, confirm the VTX table, and put readable battery and flight information in the OSD.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KwYoZlh0j_gbse-gBYqwqA.png" /><figcaption><em>Ports: enable the receiver and VTX functions only on the UARTs used by this tiny mobula 7 aircraft.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rmJ93lrJg98cqZCuZ14c5Q.png" /><figcaption><em>Video Transmitter: verify the selected band, channel and table against the installed VTX.</em></figcaption></figure><h3>Board layouts and parts are not interchangeable</h3><p>The Apex uses a Mamba F722-based controller; the Bjorn 2.5 uses a SpeedyBee F405 Mini stack. On our SpeedyBee, the published wiring map made pad selection easier. The Mamba took more careful diagram checking, and the OSD shoId temperatures approaching 90°C during heavy use or when I had overconfigured the build. This is an observation from this build, not a brand-wide thermal comparison. Read the manual for the exact board revision before soldering or assigning UARTs.</p><p>I have in my inventory two separate motor sets: has iFlight XING2 1404 motors at 3800KV with 1.5 mm shafts, listed for 3S/4S; Drone 2 (Apex) has T-Motor F1404 motors at 4600KV with 1.5 mm shafts. KV is a motor speed constant, not a battery-safety rating. Check battery, ESC, propeller and frame limits together; a motor listing alone does not establish that an assembled drone is safe for 4S.</p><p>The 2.5-inch DJI-style carbon frame became my “mini bjorn” build: a compact custom project where I had to trace motor outputs and orientation against the flight-controller diagram instead of assuming the frame wiring would make it obvious. For video, the Rush Tiny Tank is the smaller model in our Bjorn notes; the Apex inventory lists a Rush Tank Mini. Neither name is a generic VTX profile.</p><p>Budget note: our inventory records the Mobula 7 at COP 356,105, with its 1S batteries listed separately. The Apex parts Ire bought as separate items and the tracked total is incomplete, so a precise build-to-build price comparison would be misleading. These are historical purchase records, not current quotes or affiliate recommendations.</p><h3>PID tuning comes after the build checks</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BJHqlfkUBlHDDBKKuUoYgg.png" /><figcaption>Default PID tuning that comes with firmware might need adjustment for your comfort</figcaption></figure><p>Keep the terms simple: P pushes the quad toward the commanded movement; I helps it hold its attitude against sustained errors; D damps rapid changes but can add motor heat when excessive; Feedforward shapes the immediate response to stick movement. Rates and expo set how much rotation the sticks request. If the quad merely feels too quick, adjust rates/expo before changing PIDs. If it oscillates, bounces, drifts or gets hot, diagnose with a clean build, good props, correct motor order, logs and small changes — not a copied tune.</p><p>For a first setup, keep the firmware’s safe baseline, verify board alignment, motor map, prop match, receiver failsafe, arming switch and flight modes with props removed. Make one tuning change at a time, take a short controlled flight, check motor temperature by hand after disarming, and save the result for that aircraft only. A conversation suggestion is not a measured tune.</p><p>There’s no single ideal rate profile; it depends on how you fly. To make forward/back stick movement feel less abrupt, adjust <strong>pitch rates</strong> in <strong>PID Tuning → Rateprofile Settings</strong>:</p><ul><li>For twitchiness during small stick movements, lower <strong>Pitch Center Sensitivity</strong> (pitch_rc_rate)</li><li>To make the response gentler through the middle of stick travel, increase <strong>Pitch Expo</strong> (pitch_expo) slightly.</li><li>If the drone rotates too quickly during large stick movements, lower <strong>Pitch Max Rate</strong> (pitch_srate).</li></ul><p>Make one small change at a time, check the rates graph, then test cautiously. The Betaflight Rate Calculator explains how the different rate types affect the curve and rotation speed. <a href="https://betaflight.com/docs/wiki/guides/current/Rate-Calculator">Rate Calculator</a></p><p>On the other hand if small throttle movements make the drone climb or drop too abruptly,</p><ul><li><strong>Throttle Expo</strong> controls how gentle throttle feels around the chosen part of the stick range. Higher Expo makes small throttle movements produce smaller power changes there, which helps with smooth hovering and altitude control.</li><li><strong>Throttle Mid </strong>positions that softer region.</li><li><strong>RC Rate</strong> controls initial stick sensitivity for roll, pitch, and yaw. Higher values make the drone start rotating faster from small stick movements. It does <strong>not</strong> affect throttle.</li><li><strong>Super Rate</strong> increases rotation speed near full stick deflection. It lets you keep controlled center-stick handling while still reaching fast flips and rolls.</li><li><strong>Roll/Pitch/Yaw Expo</strong> softens rotation around center stick. Higher Expo makes small steering inputs gentler, while the ends of the stick remain responsive.</li></ul><p>For the Apex 3&#39;:</p><ul><li><strong>Throttle Mid: 50</strong><br> This places the softest throttle response at half stick. It is only the center of the throttle curve, not the actual hover-throttle value. Once we know the Apex hovers, for example, around 42% stick, set Throttle Mid closer to 42.</li><li><strong>Throttle Expo: 30</strong><br> This makes throttle changes gentler around the selected midpoint. At 50 Mid, moving from 45% to 55% throttle produces a smaller motor-output change than before; far from that region, the response becomes progressively more direct. Moving Expo from 20 to 30 is a modest increase in fine control.</li><li><strong>Throttle RC Rate: none in Betaflight’s throttle curve</strong><br> “RC Rate” applies to the rotational axes: roll, pitch, and yaw. It does not change throttle. If you are seeing an RC Rate field, it controls how quickly the quad begins rotating when you move the relevant stick from center.</li></ul><p>Check the throttle graph and test one change at a time. Expo helps you manage throttle more smoothly, but it doesn’t automatically slow a fall when you cut throttle completely. A controlled descent still requires some throttle. Betaflight throttle settings</p><h3>Our quick setup order</h3><p>1. Remove props; save a private backup for this exact flight controller.</p><p>2. Confirm target, Betaflight version, gyro, board orientation and receiver type.</p><p>3. Check UART wiring and Receiver-tab stick/AUX movement; set modes and failsafe.</p><p>4. Set the exact VTX control port, table, channel and OSD; match the goggles.</p><p>5. Use the Motors tab to verify order and spin, then install the matching props.</p><p>6. Only after the physical checks, do a brief first hover and tune rates/PIDs from observed behavior.</p><p>Why the micro matters: a tiny whoop can be forgiving in a living room, but its 1S battery, light frame and integrated electronics demand their own setup. The 2.5- and 3-inch builds add voltage, mass and separate wiring choices. The habits transfer; the CLI values do not.</p><p><strong>OSD:</strong> Configure the on-screen display to show the information you need while flying. I enabled altitude, battery voltage and a battery level indicator, and added temperature so I could watch for overheating. Small builds can warm up quickly; in our experience, the Mamba setup reached around 90 °C during heavy use or when overconfigured. Treat that as an observation from this build, not a universal temperature, and investigate excessive heat rather than continuing to fly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/1*SVzJxuqhJXNK1_cFsm4yBQ.png" /><figcaption>Make sure you take into account the size of your screen, my SoloGood 4.3-inch 5.8GHz 40CH DVR FPV monitor cxan only handle about 2 buttons down and 2 up.</figcaption></figure><p><strong>Radio and switch mapping:</strong> On the RadioMaster Pocket, EdgeTX offers two ways to set up switch actions. You can assign a physical switch directly as the source for a channel on the <strong>Mixes</strong> page, or first create a named input on the <strong>Inputs</strong> page and then use that input as the source in a channel mix. Either way, the mix must send the action on an output channel: Betaflight maps channels 1–4 to the sticks, then channel 5 to AUX1, channel 6 to AUX2, and so on — so channel 9 is AUX5. I mapped AUX1 to ARM, AUX2 to Acro/Angle/Horizon, AUX3 to the beeper, AUX4 to Altitude Hold, and AUX5 to Flip Over After Crash. Betaflight receives the channel values, not the names, so check the radio’s channel monitor and Betaflight’s Receiver and Modes tabs with props removed. (<a href="https://manual.edgetx.org/color-radios/model-settings/inputs-mixes-and-outputs/inputs">EdgeTX Inputs</a>, <a href="https://manual.edgetx.org/color-radios/model-settings/inputs-mixes-and-outputs/mixes">EdgeTX Mixes</a>)</p><h3><strong>Sensors and modes:</strong></h3><p>Our Apex switch plan uses a separate ARM switch and a three-position flight-mode switch: <strong>Acro</strong> for manual flying, <strong>Angle</strong> for self-leveling, and <strong>Horizon</strong> for self-leveling near the stick center with more freedom for flips and rolls at the ends. I mapped AUX4 high to <strong>Altitude Hold</strong>, which also puts the quad in Angle mode, and reserved AUX5 for <strong>Flip Over After Crash</strong> after a crash. These are configured switch assignments; check the Modes tab and each switch position before flying. For Altitude Hold, confirm the Mamba’s barometer reports stable data in Sensors, use firmware built with ALTITUDE_HOLD, and calibrate hover_throttle from a real hover. Altitude Hold keeps height but does not stop horizontal drift. I haven’t added GPS for Position Hold; that’s a future open-sky project, while GPS-denied position holding would need more than LiDAR alone, such as optical flow paired with a downward range sensor. <a href="https://betaflight.com/docs/wiki/guides/current/Modes">Betaflight Modes</a>, <a href="https://betaflight.com/docs/wiki/guides/current/Position-Hold-2025-12">Altitude and Position Hold</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*CCC_h6F886w03CJQYo2zxg.png" /><figcaption>Notice how with the right firmware it will automatically identify available sensors.</figcaption></figure><p><strong>Radio and switch mapping:</strong> On the RadioMaster Pocket, EdgeTX offers two ways to set up switch actions. You can assign a physical switch directly as the source for a channel on the <strong>Mixes</strong> page, or first create a named input on the <strong>Inputs</strong> page and then use that input as the source in a channel mix. Either way, the mix must send the action on an output channel: Betaflight maps channels 1–4 to the sticks, then channel 5 to AUX1, channel 6 to AUX2, and so on — so channel 9 is AUX5. We mapped AUX1 to ARM, AUX2 to Acro/Angle/Horizon, AUX3 to the beeper, AUX4 to Altitude Hold, and AUX5 to Flip Over After Crash. Betaflight receives the channel values, not the names, so check the radio’s channel monitor and Betaflight’s Receiver and Modes tabs with props removed. (<a href="https://manual.edgetx.org/color-radios/model-settings/inputs-mixes-and-outputs/inputs">EdgeTX Inputs</a>, <a href="https://manual.edgetx.org/color-radios/model-settings/inputs-mixes-and-outputs/mixes">EdgeTX Mixes</a>)</p><h3>References and image notes</h3><p>Screenshots are from our Mobula 7 setup folder and Apex Betaflight capture. Build versions and Iights are our saved observations; they are not manufacturer-wide claims.</p><p>• Betaflight Setup Guide (https://betaflight.com/docs/wiki/getting-started/setup-guide) and Serial / UART guide (https://betaflight.com/docs/wiki/guides/current/Serial)</p><p>• Betaflight Motors tab (https://betaflight.com/docs/wiki/app/motors-tab), reversed motor direction (https://betaflight.com/docs/wiki/guides/current/Reversed-motor-direction), and PID Tuning tab (https://betaflight.com/docs/wiki/app/pid-tuning-tab)</p><p>• Happymodel Mobula 7 1S manual (https://www.happymodel.cn/wp-content/uploads/2022/05/Mobula7-and-Mobula7-HD-1S-PNP-and-TBS-Manual.pdf)</p><p>• Rush Tiny Tank specifications (https://rushfpv.net/en-fr/products/tank-tiny-vtx); SpeedyBee F405 Mini wiring manual (https://support.speedybee.cn/?i=186034740992675840&amp;l=en&amp;p=0&amp;s=teched&amp;t=a)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b042a4d1a439" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Migrating from ingress-nginx to Gateway API on Kubernetes (AKS)]]></title>
            <link>https://medium.com/@datatomas/migrating-from-ingress-nginx-to-gateway-api-on-kubernetes-aks-246984abfd58?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/246984abfd58</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[secretproviderclass]]></category>
            <category><![CDATA[nginx]]></category>
            <category><![CDATA[ak]]></category>
            <category><![CDATA[api-gateway]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Thu, 24 Sep 2026 20:48:56 GMT</pubDate>
            <atom:updated>2026-09-24T20:48:56.252Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0ypHceOnkkD4Jxg0skb05g.png" /><figcaption>figure 1:Cross Namespace Routing taken from Kubernetes Documentation <a href="https://gateway-api.sigs.k8s.io/guides/user-guides/multiple-ns/">Cross-Namespace routing | Gateway API</a></figcaption></figure><p>There are plenty of guides on swapping ingress-nginx for Gateway API. Almost none of them mention the two things that will actually take your service down during the cutover: your internal DNS record, and the Private Link Service that still points at the old load balancer frontend.</p><p>Here is the whole migration, in the order that keeps traffic alive.</p><p><strong>The shape of the problem</strong></p><p>The old path:</p><p>client -&gt; private endpoint -&gt; PLS -&gt; ILB frontend A -&gt; ingress-nginx -&gt; pods</p><p>The new path:</p><p>client -&gt; private endpoint -&gt; PLS -&gt; ILB frontend B -&gt; Gateway -&gt; pods</p><p>Same private endpoint. Same DNS name. Same certificate subject, if you plan it right. What changes is the middle, and every piece of the middle is a separate resource that has to be repointed deliberately.</p><p><strong>1. Install Gateway API and the controller</strong></p><p>Three ways to get there, pick one:</p><h4>A. Manual, minimal (no scripting overhead)</h4><p>Just the required kubectl applys, pinned to real versions (check latest tags first):</p><pre>GATEWAY_API_VERSION=v1.3.0   # check: https://github.com/kubernetes-sigs/gateway-api/releases<br>NGF_VERSION=v1.6.1           # check: https://github.com/nginx/nginx-gateway-fabric/releases<br><br>kubectl apply --server-side -f &quot;https://github.com/kubernetes-sigs/gateway-api/releases/download/${GATEWAY_API_VERSION}/standard-install.yaml&quot;<br>kubectl apply --server-side -f &quot;https://raw.githubusercontent.com/nginx/nginx-gateway-fabric/${NGF_VERSION}/deploy/crds.yaml&quot;<br>kubectl apply --server-side -f &quot;https://raw.githubusercontent.com/nginx/nginx-gateway-fabric/${NGF_VERSION}/deploy/azure/deploy.yaml&quot;</pre><p>Same commands your script already runs — this is just the bare sequence without the multi-cluster loop, logging, and skip-detection logic, for when you’re doing one cluster by hand.</p><h4>B. Helm (recommended if you’ll manage upgrades)</h4><pre>kubectl apply --server-side -f https://github.com/kubernetes-sigs/gateway-api/releases/download/v1.3.0/standard-install.yaml<br><br>helm install ngf oci://ghcr.io/nginx/charts/nginx-gateway-fabric \<br>  --create-namespace -n nginx-gateway</pre><p>Helm still doesn’t own the Gateway API CRDs (NGF’s chart assumes they exist), so that first line stays regardless of install method. Helm gives you helm upgrade/helm uninstall for free afterward instead of re-running raw manifests.</p><h4>C. AKS-managed Gateway API (no NGINX at all)</h4><p>If you’re open to not using NGINX specifically, AKS now has a first-party managed Gateway API implementation via the application routing add-on — but it’s Istio-based under the hood, not NGINX Gateway Fabric</p><pre>az aks approuting enable --resource-group &lt;rg&gt; --name &lt;cluster&gt; --enable-gateway-api</pre><p>This is worth knowing about because Microsoft is positioning it as the long-term successor to managed NGINX ingress (which loses Azure support after November 2026), but it’s a different data plane (Envoy/Istio, not NGINX), so only pick this if you don’t need NGINX-specific features.</p><p>Verify with either A or B:</p><pre>kubectl get crd gateways.gateway.networking.k8s.io<br>kubectl get pods -n nginx-gateway</pre><p><strong>1.2. Deploy the Gateway Object</strong></p><p>Gateway API denies cross-namespace references by default, and it applies that rule in <strong>two independent directions</strong> — which is the part that trips people up, because they’re configured in different places by different teams.</p><p><strong>Routes → Gateway</strong> is controlled by the Gateway, via allowedRoutes. from: Same (the default) accepts only routes in the Gateway&#39;s own namespace; from: All accepts everything; from: Selector accepts namespaces matching a label. The selector is the usual choice: the platform team owns the Gateway, app teams own their routes in their own namespaces, and the label is the handshake. Each HTTPRoute then points back with parentRefs naming the Gateway and its namespace.</p><p><strong>Gateway → Secret</strong> runs the opposite way, and allowedRoutes has no say in it. If certificateRefs includes a namespace field pointing elsewhere, the namespace that <em>owns the Secret</em> must publish a ReferenceGrant authorizing it:</p><pre>apiVersion: gateway.networking.k8s.io/v1beta1<br>kind: ReferenceGrant<br>metadata:<br>  name: allow-gw-tls<br>  namespace: ns-app          # el namespace DUEÑO del Secret<br>spec:<br>  from:<br>    - group: gateway.networking.k8s.io<br>      kind: Gateway<br>      namespace: nginx-gateway<br>  to:<br>    - group: &quot;&quot;<br>      kind: Secret<br>      name: tls-app</pre><p>Then label each namespace whose routes you want attached:</p><pre>kubectl label namespace ns-app gateway-access=allowed</pre><p>Each HTTPRoute then points back with parentRefs naming the Gateway and its namespace. This direction is all most setups need, and it&#39;s the one you&#39;ll configure every time.</p><p><strong>Gateway → Secret</strong> runs the opposite way, and allowedRoutes has no say in it. It only comes up <strong>if you need a cross-namespace reference, </strong>that is, if certificateRefs carries an explicit namespace field pointing at a Secret somewhere else. Keep the Secret beside the Gateway and omit that field, and none of this applies.</p><p>If you do need it, the namespace that <em>own</em></p><pre>apiVersion: gateway.networking.k8s.io/v1beta1<br>kind: ReferenceGrant<br>metadata:<br>  name: allow-gw-tls<br>  namespace: ns-app          # el namespace DUEÑO del Secret<br>spec:<br>  from:<br>    - group: gateway.networking.k8s.io<br>      kind: Gateway<br>      namespace: nginx-gateway<br>  to:<br>    - group: &quot;&quot;<br>      kind: Secret<br>      name: tls-app</pre><p>The direction is deliberate: the owner of a resource grants access to it, rather than the consumer claiming it. Per spec, a missing grant leaves the listener at ResolvedRefs=False with reason RefNotPermitted — though enforcement varies by implementation, and some controllers resolve the reference anyway. Don&#39;t rely on that; if you&#39;re referencing across namespaces, write the grant.</p><p>The same rule covers an HTTPRoute whose backendRefs point at a Service in another namespace — that&#39;s the other place you&#39;ll meet ReferenceGrant.</p><p>Simplest path stays: keep the Secret in the Gateway’s namespace and skip namespace in certificateRefs. It matters more than it looks when the cert comes from Key Vault, since the CSI driver can only create the synced Secret where the mounting pod runs — so a cross-namespace certificateRefs forces the sync pod into the app namespace <em>and</em> requires the grant. Two moving parts instead of zero.</p><p><strong>2. Enable the Key Vault CSI driver and workload identity</strong></p><p>This is what lets the cluster pull the TLS certificate from Key Vault instead of you pasting a base64 blob into a Secret.</p><pre>az aks enable-addons -g RG_APP -n aks-app-01 \ <br><br>  --addons azure-keyvault-secrets-provider <br><br>  <br><br>az aks update -g RG_APP -n aks-app-01 \ <br><br>  --enable-oidc-issuer --enable-workload-identity </pre><p>Verify before moving on, because the control plane reports these independently:</p><pre>az aks show -g RG_APP -n aks-app-01 \ <br><br>  --query &quot;{csi:addonProfiles.azureKeyvaultSecretsProvider.enabled, \ <br><br>wi:securityProfile.workloadIdentity.enabled, \ <br><br>oidc:oidcIssuerProfile.enabled}&quot; -o yaml </pre><p><strong>3. Create an identity and federate it</strong></p><p>The identity is what the CSI driver presents to Key Vault. Federation is what lets a Kubernetes ServiceAccount borrow that identity without a stored secret.</p><pre>az identity create -g RG_APP -n id-app-tls-01 -l eastus2 <br><br>  <br><br>CLIENT_ID=$(az identity show -g RG_APP -n id-app-tls-01 --query clientId -o tsv) <br><br>PRINCIPAL_ID=$(az identity show -g RG_APP -n id-app-tls-01 --query principalId -o tsv) <br><br>KV_ID=$(az keyvault show -n kv-app-01 --query id -o tsv) <br><br>  <br><br>az role assignment create --assignee-object-id $PRINCIPAL_ID \ <br><br>  --assignee-principal-type ServicePrincipal \ <br><br>  --role &quot;Key Vault Secrets User&quot; --scope &quot;$KV_ID&quot; <br><br>  <br><br>ISSUER=$(az aks show -g RG_APP -n aks-app-01 \ <br><br>  --query oidcIssuerProfile.issuerUrl -o tsv) <br><br>  <br><br>az identity federated-credential create \ <br><br>  --name fc-nginx-gateway-sa-kv-cert-sync \ <br><br>  --identity-name id-app-tls-01 -g RG_APP \ <br><br>  --issuer &quot;$ISSUER&quot; \ <br><br>  --subject &quot;system:serviceaccount:nginx-gateway:sa-kv-cert-sync&quot; \ <br><br>  --audiences api://AzureADTokenExchange </pre><p>Two things that bite here.</p><p><strong>The role is Secrets User, not Certificates User.</strong> The CSI driver reads a certificate through the secret endpoint, because that is the only endpoint that returns the private key alongside the cert.</p><p><strong>The subject string is exact.</strong> system:serviceaccount:&lt;namespace&gt;:&lt;serviceaccount&gt;. A typo does not fail here — it fails much later, at pod mount, with a 401 that tells you nothing.</p><p><strong>4. Import the certificate</strong></p><pre>az keyvault certificate import \ <br><br>  --vault-name kv-app-01 --name tls-app \ <br><br>  --file wildcard.pfx --password &#39;REDACTED&#39; </pre><p>Check the chain before you trust it. A leaf-only PFX will validate in your browser (which caches intermediates) and fail for API clients that do not:</p><pre>openssl pkcs12 -in wildcard.pfx -nokeys -passin pass:&#39;REDACTED&#39; \ <br><br>  | grep -c &quot;BEGIN CERTIFICATE&quot; </pre><p><strong>5. The SecretProviderClass</strong></p><p>This is the piece people get wrong most often, so here it is in full:</p><pre>apiVersion: secrets-store.csi.x-k8s.io/v1 <br><br>kind: SecretProviderClass <br><br>metadata: <br><br>  name: app-tls-keyvault <br><br>  namespace: nginx-gateway <br><br>spec: <br><br>  provider: azure <br><br>  secretObjects: <br><br>    - secretName: tls-app <br><br>      type: kubernetes.io/tls <br><br>      data: <br><br>        - objectName: tls-app <br><br>          key: tls.key <br><br>        - objectName: tls-app <br><br>          key: tls.crt <br><br>  parameters: <br><br>    usePodIdentity: &quot;false&quot; <br><br>    useVMManagedIdentity: &quot;false&quot; <br><br>    clientID: &quot;&lt;CLIENT_ID&gt;&quot; <br><br>    keyvaultName: kv-app-01 <br><br>    tenantId: &quot;&lt;TENANT_ID&gt;&quot; <br><br>    objects: | <br><br>      array: <br><br>        - | <br><br>          objectName: tls-app <br><br>          objectType: secret <br><br>          objectFormat: pem </pre><p>Three traps in that file.</p><p><strong>objectFormat: pem is mandatory</strong> if the certificate was imported from a PFX. Key Vault stores it with contentType application/x-pkcs12, and without this line you get binary PKCS#12 that the driver cannot split into tls.crt and tls.key.</p><p><strong>secretObjects is what creates the Kubernetes Secret.</strong> Without that block the certificate mounts as a file and no Secret ever appears, which means the Gateway has nothing to reference.</p><p><strong>The namespace has to match the Gateway’s.</strong> The CSI driver can only create the synced Secret in the namespace where the mounting pod runs. If your Gateway’s certificateRefs points somewhere else, you either move this or add a ReferenceGrant.</p><p><strong>6. The pod nobody tells you about</strong></p><p>Applying the SecretProviderClass does nothing on its own. The Secret only materializes while a pod is actively mounting the CSI volume — and the Gateway controller reads the Secret through the API, it does not mount anything.</p><p>So you need a pod whose only job is to hold the mount open:</p><pre>apiVersion: v1 <br><br>kind: ServiceAccount <br><br>metadata: <br><br>  name: sa-kv-cert-sync <br><br>  namespace: nginx-gateway <br><br>  annotations: <br><br>    azure.workload.identity/client-id: &quot;&lt;CLIENT_ID&gt;&quot; <br><br>--- <br><br>apiVersion: apps/v1 <br><br>kind: Deployment <br><br>metadata: <br><br>  name: kv-cert-sync <br><br>  namespace: nginx-gateway <br><br>spec: <br><br>  replicas: 1 <br><br>  selector: <br><br>    matchLabels: <br><br>      app: kv-cert-sync <br><br>  template: <br><br>    metadata: <br><br>      labels: <br><br>        app: kv-cert-sync <br><br>        azure.workload.identity/use: &quot;true&quot; <br><br>    spec: <br><br>      serviceAccountName: sa-kv-cert-sync <br><br>      securityContext: <br><br>        runAsNonRoot: true <br><br>        runAsUser: 65532 <br><br>        seccompProfile: <br><br>          type: RuntimeDefault <br><br>      containers: <br><br>        - name: pause <br><br>          image: mcr.microsoft.com/oss/kubernetes/pause:3.9 <br><br>          resources: <br><br>            requests: {cpu: 10m, memory: 16Mi} <br><br>            limits: {memory: 32Mi} <br><br>          securityContext: <br><br>            allowPrivilegeEscalation: false <br><br>            readOnlyRootFilesystem: true <br><br>            capabilities: <br><br>              drop: [&quot;ALL&quot;] <br><br>          volumeMounts: <br><br>            - name: kv-certs <br><br>              mountPath: /mnt/kv <br><br>              readOnly: true <br><br>      volumes: <br><br>        - name: kv-certs <br><br>          csi: <br><br>            driver: secrets-store.csi.k8s.io <br><br>            readOnly: true <br><br>            volumeAttributes: <br><br>              secretProviderClass: app-tls-keyvault </pre><p>The azure.workload.identity/use: “true” label on the <strong>pod template</strong> is what triggers the webhook that injects the token. Without it, the mount fails with a 403 and the annotation on the ServiceAccount alone will not save you.</p><p>Also: do not set automountServiceAccountToken: false on this pod. Workload identity needs the projected token.</p><p>Verify you got the right certificate, not just a running pod:</p><pre>kubectl get secret tls-app -n nginx-gateway \ <br><br>  -o jsonpath=&#39;{.data.tls\.crt}&#39; | base64 -d \ <br><br>  | openssl x509 -noout -subject -fingerprint -sha1 </pre><p>Compare that fingerprint to the one in Key Vault. A pod going Ready only proves the mount worked.</p><p><strong>7. Certificate rotation: the flags, and when to use them</strong></p><p>Rotation is off by default. Turning it on is one flag, but the decision is not automatic — here is both.</p><p><strong>The flags</strong></p><pre># at addon install time <br><br>az aks enable-addons -g RG_APP -n aks-app-01 \ <br><br>  --addons azure-keyvault-secrets-provider \ <br><br>  --enable-secret-rotation \ <br><br>  --rotation-poll-interval 2m <br><br>  <br><br># on a cluster where the addon is already enabled <br><br>az aks update -g RG_APP -n aks-app-01 \ <br><br>  --enable-secret-rotation --rotation-poll-interval 2m <br><br># turn it back off <br><br>az aks update -g RG_APP -n aks-app-01 --disable-secret-rotation <br><br>  <br><br># check the current state <br><br>az aks show -g RG_APP -n aks-app-01 \ <br><br>  --query addonProfiles.azureKeyvaultSecretsProvider.config -o yaml </pre><p>— rotation-poll-interval defaults to 2m and only takes effect together with — enable-secret-rotation. It accepts Go duration strings: 2m, 30m, 1h.</p><p><strong>When to rotate</strong></p><ul><li><strong>Certificates that renew on a schedule you do not control.</strong> An automated PKI or an ACME issuer writing straight into Key Vault — the whole point is that nobody is watching.</li><li><strong>Short-lived certificates.</strong> Anything on a 90-day or shorter cycle. The renewal cadence is faster than most change windows.</li><li><strong>Clusters where nobody owns the runbook.</strong> If the restart step depends on a specific person remembering, the automation is cheaper than the outage.</li><li><strong>Secrets that are genuinely rotated</strong>, like database credentials mounted alongside the cert.</li></ul><p><strong>When not to rotate</strong></p><ul><li><strong>Change-controlled production.</strong> If every production change needs a ticket, an automatic poll that swaps the serving certificate at 3am is an unapproved change. Rotate deliberately, inside the window.</li><li><strong>Long-lived certificates renewed by hand.</strong> A one-year OV cert renewed manually gains almost nothing from polling every two minutes, and you will do a controlled rollout anyway.</li><li><strong>Large clusters with many SecretProviderClasses.</strong> Each poll is a Key Vault request per pod per mounted object. On a big cluster with a short interval this gets you throttled, and throttling looks like random mount failures.</li><li><strong>When you need the swap to be auditable.</strong> A rollout restart leaves a deployment revision. A background poll leaves nothing obvious in the cluster’s history.</li></ul><blockquote>If rotation is off, the driver reads the certificate once at mount and never again. When the cert is renewed in Key Vault, the Secret keeps serving the old one until you restart the pod. Put this in the renewal runbook, next to the renewal step itself:</blockquote><pre>kubectl rollout restart deploy/kv-cert-sync -n nginx-gateway </pre><p>A useful middle ground: leave rotation off, and set a calendar reminder keyed to the certificate’s expiry date rather than to a polling interval.</p><p><strong>8. Repoint the Private Link Service</strong></p><p>This is the actual cutover.</p><p>Your PLS is bound to an ILB <strong>frontend IP configuration</strong>, not to a Kubernetes Service. The new Gateway created a second frontend on the same internal load balancer. List them:</p><pre>MCRG=MC_RG_APP_aks-app-01_eastus2 <br><br>az network lb frontend-ip list --lb-name kubernetes-internal -g $MCRG -o table </pre><p>You will see two. Match the IP to the Gateway’s Service EXTERNAL-IP, then repoint:</p><p>FE_NEW=$(az network lb frontend-ip show — lb-name kubernetes-internal \</p><p>-g $MCRG -n &lt;new-frontend-name&gt; — query id -o tsv)</p><p>az network private-link-service update -n pls-app-01 -g RG_APP \</p><p>— lb-frontend-ip-configs “$FE_NEW”</p><blockquote><strong>Update it. Never delete and recreate it.</strong> The PLS alias is what every consumer’s private endpoint is bound to. Recreating the PLS issues a new alias and breaks all of them, including ones in other subscriptions that you may not control or even know about.</blockquote><p>Check who is connected before you touch anything:</p><pre>az network private-link-service show -n pls-app-01 -g RG_APP \ <br><br>  --query &quot;{alias:alias, conns:privateEndpointConnections[].\ <br><br>{state:privateLinkServiceConnectionState.status, pe:privateEndpoint.id}}&quot; -o yaml </pre><p>Traffic moves the moment this command returns.</p><p><strong>10. Internal DNS</strong></p><p>If your clients resolve the hostname through an internal DNS server rather than Azure Private DNS, that A record is a separate system that knows nothing about any of the above.</p><p>NXDOMAIN means the name does not exist in the zone at all — not that it resolved to the wrong place. Before and after adding the record:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/870/1*ywNtAJ-ImgQOCNrfbZozCw.png" /><figcaption>figure 2: <em>The same query before and after the A record exists. NXDOMAIN is the zone saying the name was never there.</em></figcaption></figure><p><strong>Opening the DNS console on Windows</strong></p><p>The console is <strong>DNS Manager</strong>. Three ways in:</p><pre>dnsmgmt.msc                          :: Win+R, or any command prompt <br><br>dnsmgmt.msc /computer:dc01           :: target a specific server <br><br>  <br><br>:: Server Manager -&gt; Tools -&gt; DNS <br><br>:: Administrative Tools -&gt; DNS </pre><p>From a workstation you need RSAT installed; the console then prompts you for a server to connect to.</p><p><strong>Adding the record</strong></p><p>In the GUI: expand the server, <strong>Forward Lookup Zones</strong>, right-click the zone, <strong>New Host (A or AAAA)</strong>.</p><p>Or from PowerShell, which is what you want if this is going in a runbook:</p><pre>Add-DnsServerResourceRecordA -ZoneName &quot;contoso.net&quot; -Name &quot;api-example&quot; ` <br><br>  -IPv4Address &quot;10.20.30.40&quot; -TimeToLive 00:05:00 <br><br>  <br><br># verify against that server specifically, not your default resolver <br><br>Resolve-DnsName api-example.contoso.net -Server dc01 <br><br>Get-DnsServerResourceRecord -ZoneName &quot;contoso.net&quot; -Name &quot;api-example&quot; </pre><p>Point it at whatever the client actually reaches: the ILB IP for callers inside the VNet, or the private endpoint’s NIC IP for callers coming through Private Link. These are different addresses and it is easy to publish the wrong one.</p><blockquote>Drop the TTL to five minutes a day before the migration, not during it. A record cached at one hour will outlast your maintenance window, and there is nothing you can do about it from the server side. (Actually not sure of this tip I just read it somwhere but haven’t tested it , use with precaution)</blockquote><p><strong>11. Verify, then delete</strong></p><p>Check the certificate being served on the wire, not just the Secret:</p><pre>kubectl run tlstest --rm -it --image=nicolaka/netshoot --restart=Never -- \ <br><br>  openssl s_client -connect &lt;ilb-ip&gt;:443 -servername api-example.contoso.net \ <br><br>  &lt;/dev/null 2&gt;/dev/null | openssl x509 -noout -subject -fingerprint -sha1 </pre><p>Then check from a real client, through the private endpoint, with the real DNS name. Then watch your route metrics and application logs long enough to be convinced — traffic arriving is not the same as traffic succeeding.</p><p>Only then, remove ingress-nginx:</p><pre># 1. webhook first - otherwise it can block other admissions once the pod is gone <br><br>kubectl delete validatingwebhookconfiguration ingress-nginx-admission <br><br>  <br><br># 2. the stale Ingress object <br><br>kubectl -n ns-app delete ingress app-gateway <br><br>  <br><br># 3. the namespace - takes the deployment, both services and the LB with it <br><br>kubectl delete namespace ingress-nginx <br><br>  <br><br># 4. cluster-scoped leftovers <br><br>kubectl delete clusterrole ingress-nginx ingress-nginx-admission --ignore-not-found <br><br>kubectl delete clusterrolebinding ingress-nginx ingress-nginx-admission --ignore-not-found <br><br>kubectl delete ingressclass nginx --ignore-not-found </pre><p>The webhook goes first for a reason. If you delete the namespace while the ValidatingWebhookConfiguration still exists, the webhook points at a Service that is gone, and every matching admission request in the cluster starts failing.</p><p>Deleting the ingress-nginx Service releases its ILB frontend. If the frontend lingers, something still references it, usually a load balancing rule, or the PLS you forgot to repoint. Azure will tell you exactly which:</p><pre>CannotDeleteLoadbalancerFrontendIPConfigsInUseByOtherResources ... <br><br>in use by ... /privateLinkServices/pls-app-01 </pre><p><strong>The short version</strong></p><p>Install the controller, mount the cert, build the Gateway, all of that is reversible and touches nothing live. The cutover is exactly two commands: repoint the PLS, update the DNS record. Everything before them is preparation, and everything after them is verification.</p><p>Delete ingress-nginx last, and only once you are sure.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=246984abfd58" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Configure an RC Transmitter for Advanced Drone Simulators on Linux]]></title>
            <link>https://medium.com/@datatomas/how-to-configure-an-rc-transmitter-for-advanced-drone-simulators-on-linux-4bad863e0aef?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/4bad863e0aef</guid>
            <category><![CDATA[drones]]></category>
            <category><![CDATA[simulator]]></category>
            <category><![CDATA[radiomaster]]></category>
            <category><![CDATA[fpv]]></category>
            <category><![CDATA[linux]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 21:48:18 GMT</pubDate>
            <atom:updated>2026-09-05T21:48:18.799Z</atom:updated>
            <content:encoded><![CDATA[<p>A practical, known-good profile tested with RadioMaster Pocket (EdgeTX) and Ukrainian Fight Drone Simulator (UFDS).</p><p><strong>The core idea: </strong>the simulator never sees “the left stick” or “SA.” It sees USB HID axes and buttons. Build the chain correctly: physical control → EdgeTX mix → output channel → USB axis/button → simulator binding.</p><p><strong>Understand the channel logic</strong></p><p><strong>Use a dedicated simulator model. </strong>Do not change the physical stick behavior. Instead, route each stick or switch to a channel, then tell EdgeTX how that channel should appear over USB. This is what solved both the half-range throttle and the missing-button problem in this setup.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9LeR7UetQb4lcesbg5v_dA.png" /><figcaption>Table 1: channels for throttle and rotation</figcaption></figure><p><strong>Why CH2 matters: </strong>UFDS treated throttle on the original CH3 as a half-range/trigger-style axis (roughly 0 to +100). Moving Thr to CH2 fixed the full-range throttle. CH5 + rotY restored Pitch after the remap.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sdnfeWeryxgm7YbBecy8wQ.png" /><figcaption><strong>Figure 1: Ignore the transmitter model shown by Steam or UFDS.</strong> The important abstraction is the USB HID interface exported by EdgeTX: axes for continuous controls and buttons for switches.</figcaption></figure><p><strong>Tested profile: RadioMaster Pocket + EdgeTX + Ukrainian Fight Drone Simulator (UFDS)</strong></p><p><strong>Radio model basics: </strong>Trainer OFF, RF modules OFF for simulator use, USB Joystick = Advanced, interface mode = Joystick. Connect through the data-capable USB port and select USB Joystick/HID when EdgeTX prompts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/530/1*17x84Yu0KbSQqqyOl-5GSQ.png" /><figcaption>Figure 2 — EdgeTX model setup: Trainer OFF and USB Joystick enabled.</figcaption></figure><h3>Verify the USB device</h3><pre>lsusb | grep -i 1209<br> # Working Pocket seen here:<br> # ID 1209:4f54 Generic Radiomaster Pocket Joystick</pre><p>For another transmitter, use its own VID:PID from <strong>lsusb</strong>.</p><h3>Steam settings that worked</h3><ul><li>Properties → Controller → Disable Steam Input for UFDS.</li><li>Properties → General → Launch Options:</li></ul><p>PROTON_ENABLE_HIDRAW=0x1209/0x4f54 SDL_JOYSTICK_HIDAPI=0 %command%</p><h3>Joystic Linux validation</h3><p>evtest is a simple Linux input-event diagnostic tool. It lets you verify that the kernel is actually receiving stick and switch activity before you troubleshoot Steam or the simulator.Install it on Debian/Ubuntu:</p><pre>sudo apt update<br>sudo apt install evtest<br><br>sudo evtest</pre><p>Select: EdgeTX Radiomaster Pocket Joystick<br>Sticks → EV_ABS<br>Switches → EV_KEY</p><p>If evtest shows the correct events, the transmitter and Linux input layer are working; any remaining issue is likely in Steam, Proton, or the simulator configuration.</p><p><strong>Important:</strong> after changing the Advanced USB Joystick descriptor, unplug/replug the Pocket before retesting. Restart the simulator as well.</p><p><strong>Useful references: </strong><a href="https://manual.edgetx.org/edgetx-how-to/configure-advanced-joystick-with-edgetx">EdgeTX Advanced Joystick</a> · <a href="https://manual.edgetx.org/edgetx-how-to/joystick-mapping-information-for-game-developers">EdgeTX joystick mapping</a> · <a href="https://steamcommunity.com/sharedfiles/filedetails/?id=3703666140">UFDS Linux / hidraw guide</a> · <a href="https://steamcommunity.com/sharedfiles/filedetails/?id=3627458818">Pocket UFDS setup</a></p><h3>3. Reproduce the working EdgeTX profile</h3><p>Before changing any channel mapping, get familiar with the RadioMaster Pocket controls used to navigate EdgeTX. The radio is operated mainly with the <strong>MDL button</strong>, <strong>SYS button</strong>, <strong>scroll wheel</strong>, and <strong>RTN/Page buttons</strong>, while the sticks and top switches become the actual simulator inputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KJ2-Yt3VgSuL5IDajOneZw.png" /><figcaption>Figure 3— RadioMaster Pocket control layout<br>Reference for the controls used throughout this guide. Press MDL to open the current model settings, use the page controls to move between configuration screens, and use the scroll wheel to highlight, edit, and confirm values. The top switches SA–SE and both sticks are the controls that will later be assigned to simulator channels.</figcaption></figure><p>For this setup, create or use a <strong>dedicated simulator model</strong> so the unusual UFDS channel order does not affect a real aircraft profile.</p><h3>Configure MIXES</h3><p>The <strong>MIXES</strong> page determines which physical stick or switch feeds each EdgeTX output channel. Think of it as the first half of the mapping:</p><p><strong>physical control → EdgeTX channel</strong></p><p>The USB Joystick page will handle the second half:</p><p><strong>EdgeTX channel → USB axis or button</strong></p><p>Open:</p><p><strong>MDL → MIXES — route physical controls to output channels</strong></p><p>In EdgeTX, the <strong>MIXES</strong> page decides which physical control feeds each output channel. For this simulator profile, the four stick functions are deliberately separated from the switch/button channels so they can later be exposed to Linux as distinct HID axes and buttons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Bq3W4lYcG3KtbiVK5-9hEg.png" /><figcaption><strong>Table 2. EdgeTX USB Joystick Configuration</strong><br> Working USB joystick channel mapping for the RadioMaster Pocket / EdgeTX profile, showing which channels are exposed as axes and which are exposed as buttons for simulator use.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7FLx-aHIFOVzQSPxqpyNfg.png" /><figcaption><strong>Table 3. Working EdgeTX Mixes Profile</strong><br> Working EdgeTX channel-to-source mix configuration used to map sticks and switches to the correct simulator functions, including the throttle fix and button assignments.</figcaption></figure><p>Channel-to-source mapping used by the tested simulator profile. The unusual CH2 throttle placement is intentional and solved the half-range throttle behavior observed in UFDS.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/615/1*EOK9-sWvGD9XEfqjGQG3mw.png" /><figcaption>Figure 4— Four stick channels exposed as four different axes.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/610/1*K3LejLJzbkwm0wHGfo9s2g.png" /><figcaption>Figure 5— Final Push/2POS/3POS button ranges.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/555/1*fv6GfG8m42nzQrcvBbsErg.png" /><figcaption>Figure 6— CH9 as a real Push button, not an axis.</figcaption></figure><p><strong>Button numbering: </strong>EdgeTX can number the HID buttons from 0 while the simulator may display them starting at 1.</p><h3>4. Calibrate in the simulator</h3><p><strong>The transmitter image in UFDS is generic. </strong>Do not try to match its physical layout. When a step asks for a control, move the Pocket control you want to bind. The game should identify an axis for sticks and a button for switches.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zx00YUa1vO51paV52Pp55A.png" /><figcaption>Figure 7— Arm/Disarm calibration. A toggle or push button is appropriate; the picture is not a model match.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LX6PUAMDGw1FcCIII4c41w.png" /><figcaption>Figure 8— After fixing the four flight axes, multiple actions still collapsed onto Axis 6. This was the signal to move switches to CH9+ and define them as B (Button) in Advanced USB mode.</figcaption></figure><p><strong>Do not “fix” this with Trainer mode: </strong>Trainer remains OFF. Steam Overlay is unrelated. The problem is the HID descriptor and channel mapping.</p><h3>5. Known-good UFDS result</h3><p>After reconnecting the radio and recalibrating, the working profile produced:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M2SqTrSIw-wOTXqKSLUxbg.png" /><figcaption><strong>Table 4. Known-good UFDS Control Mapping</strong><br> This configuration produced stable four-axis flight control and independent button bindings during testing. It is a practical baseline and can be refined later without changing the underlying HID layout.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BncfJzzuzte-ogfcVgOOTw.png" /><figcaption><strong>Table 5. Troubleshooting Checklist</strong><br>Quick-reference troubleshooting table for common RC transmitter simulator setup issues on Linux, including device detection, axis mapping, throttle range, switch binding, and reconnect/recalibration fixes.</figcaption></figure><p>Conclusion</p><p>Getting an RC transmitter working reliably with a Linux-based simulator is less about matching the controller shown on screen and more about understanding the complete input chain:</p><p><strong>physical control → EdgeTX MIX → output channel → USB HID axis/button → simulator binding</strong></p><p>Once that model is clear, configuration becomes much easier to troubleshoot. In this setup, moving throttle to <strong>CH2</strong>, separating the four flight controls onto distinct USB axes, and moving switches to <strong>CH9+ as HID buttons</strong> solved the main issues encountered with UFDS. Linux tools such as lsusb and evtest also make it possible to verify the transmitter independently of Steam or the simulator.</p><p>The exact channel numbers may differ with other radios or simulators, but the same approach applies: validate each layer independently, keep analog controls as axes, expose switches as buttons, and only calibrate the simulator once the operating system is receiving the expected events.</p><p>This RadioMaster Pocket profile is therefore not just a UFDS workaround , it is a useful template for configuring EdgeTX transmitters with more demanding simulator setups on Linux.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4bad863e0aef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Local-First Kubernetes for IoT: Owning Infrastructure in the Age of Rented Everything]]></title>
            <link>https://medium.com/@datatomas/local-first-kubernetes-for-iot-owning-infrastructure-in-the-age-of-rented-everything-b5889bee5db9?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/b5889bee5db9</guid>
            <category><![CDATA[k3]]></category>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[raspberry-pi]]></category>
            <category><![CDATA[kubernetes]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Sat, 25 Jul 2026 21:40:26 GMT</pubDate>
            <atom:updated>2026-07-28T16:22:28.712Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vbLMiE2zn_znVwxue5wwOw.png" /></figure><p>Most engineers rent everything now. Compute, storage, the database, the queue, the load balancer, the thing that watches the other things. It’s convenient, and as a cloud engineer I’ve watched large teams turn that convenience into real leverage, splitting Kubernetes into specialized roles so each person goes deep fast: someone administers the cluster, someone builds the apps that run on it, someone owns the networking and security posture. Teams move well this way.</p><p>But that same over-specialization has a cost. At big tech, it’s genuinely hard to become a Kubernetes *all-rounder* , you get good at your slice and the rest stays abstracted, someone else’s problem, someone else’s Terraform module or jira ticket. Convenient, right up until the bill, the outage on someone else’s schedule, or the quiet realization that you don’t actually understand the machine your business runs on. The essentials of standing up a cluster — the ones the platform team abstracted away for you years ago — are exactly the parts you never had a reason to learn.</p><p>That’s why I think every engineer or enthusiast who’s serious about Kubernetes should mount one of their own, on hardware, at least once. Not a managed offering, not `kind` on a laptop for an afternoon — a real cluster that has to survive a power blip and a bad SD card. This is where k3s earns its reputation: it’s a full, CNCF-conformant Kubernetes distribution that happens to run on machines small enough to hide in a drawer.</p><p>Call it the honest path to <strong>Kubestronaut</strong>. I’ll admit I don’t love that the official version of that title currently runs through a bundle of CNCF exams around $2000, and yes, I still want the jacket someday. But I don’t think a badge is what makes someone a Kubestronaut. Actually understanding Kubernetes down to the linux machine; cgroups, etcd, the CNI, the firewall rules that let nodes even find each other is. You get that by building the thing, not by answering multiple-choice questions about it or hanging 4 certificates on your wall, that only boosts your ego.</p><p>This is article tries to seduce you to take the other path: a small Kubernetes cluster on hardware you can hold in one hand, sitting in your house, that you own outright. Not because Kubernetes is simple, it isn’t, but because the problems it solves (deployment, restarts, rollouts, service discovery, “put this workload on the machine with the NPU”) are problems you will have the moment you run more than one container on more than one board. You can build a pile of shell scripts that reinvent half of it, badly, or you can run the boring, battle-tested thing on cheap silicon.</p><p>This is also where cloud engineering earns its keep in a different way. Working at cloud scale teaches you things you can’t learn any other way — what availability-zone resilience actually buys you, why a control plane needs quorum, what “the network is unreliable” means at 3 a.m. That knowledge usually costs someone millions of dollars in infrastructure to make real. You and I don’t have that budget, and we don’t need it: this cluster will never compete with a hyperscaler, and it was never supposed to. What it <strong>can </strong>do is automate real slice of a small business, or a real slice of a bigger one on hardware that fits in a shoebox and answers to nobody’s invoice. That’s where IoT and resourcefulness are still a legitimate edge: not at OpenAI’s scale, but at yours.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LNxMQCS2zubjlgM03ZnwqQ.png" /><figcaption><strong>Figure 1: 3 system nodes iot cluster set up</strong></figcaption></figure><h3>What you need</h3><p>The classic small cluster is three control-plane nodes running embedded etcd (three, not two or four so you can lose one and still have a majority), plus any number of agent nodes for workloads. The agents can be intermittent and there’s no need for homogeneity in them since they don’t hold cluster state and are specially designed to match whatever load you want to try it, so its designed to be hetereogeus.</p><p>The user/workload pool is where that diversity belongs, it’s the whole point of owning mixed hardware. A Raspberry Pi 5 running your API next to an Orange Pi 5 doing inference, each pool-tagged for what it’s actually good at, is a perfectly sensible cluster. Kubernetes was built to schedule across exactly that kind of mixed fleet. Just keep that mixing off the control plane — one consistent, boring ecosystem for etcd, and spend your adventurousness on the nodes that can afford to fail.</p><ul><li>Raspberry Pi : the better server. Mature OS, huge community, cgroups (with a twist we’ll show here) and networking “just work.” Ideal for control-plane / system nodes.<br>- Orange Pi (RK3588 class): the better *<strong>AI inference</strong>* box. More CPU and RAM per dollar and an on-board NPU, at the cost of a rougher OS image (more on that in the war stories). Great for the workload pool for simple ai/ml inference.</li><li>RAM: the honest version. k3s’s documented minimum is tiny: ~512 MB of RAM and a single core. You could, technically, run a node on a Raspberry Pi Zero 2 W. You should not. A real control-plane node spends 1.5–2 GB just on etcd, the API server, and the packaged system pods (CoreDNS, the ingress controller, metrics). Treat 4 GB as the practical floor for a node that runs actual workloads without living one memory spike away from the OOM killer. 8 GB if the board will also do inference. In AKS cluster i’ve run production portals that run on 8gb ram and 32 gib storage receiving hunderds of users, so 8gb ram is plenty for starters.</li><li>Storage: NVMe/SSD, or good eMMC for the orange pi. Never microSD for etcd. etcd fsyncs on every write; a slow card tanks the whole cluster’s latency and dies young. A simple user task that’s not meant to be read/write heavy could use a microSd though.</li></ul><p>This is a high-level guide with enough real detail to actually get you there.<br>We do it raw first by hand, so you understand every moving part — and then hand it to Ansible so it’s reproducible. At the end, the war stories: every way it broke for me, so it breaks less for you.</p><p><strong>Linux Knowledge:</strong></p><p><strong>Virtual networking : </strong>veth pairs, bridges, VXLAN. This is what your CNI<br>(Container Network Interface — flannel is k3s’s default) builds under the<br>hood. Every pod gets a veth pair: a virtual ethernet cable with one end<br>inside the pod’s network namespace, looking like a normal `eth0`, and the<br>other plugged into a bridge on the host. Pod-to-pod traffic on the same node is just switching across that bridge. Cross-node traffic goes out an overlay device — flannel’s is VXLAN, shipped over the `8472/udp` port you already opened in your firewall. See it for yourself:</p><pre>ip link show # cni0 (bridge), flannel.1 (VXLAN device), vethXXXX@ifN pairs<br>bridge link show # which veths are attached to cni0<br>ip -d link show flannel.1 # confirms it&#39;s VXLAN, vni 1, dstport 8472<br>ip route | grep flannel.1 # routes to other nodes&#39; pod subnets, via flannel.1</pre><p>flannel is the default, not the only option — CNI is a plugin spec, and<br>Calico and Cilium implement it differently:</p><ul><li><strong>flannel</strong> — the simple one: VXLAN overlay (or host-gw on a shared LAN), one job, get a packet from pod A to pod B across nodes. No NetworkPolicy<br> enforcement. The right default for a small cluster, and what the commands above show.<br>- Calico — the one with real pod firewalling. Its default mode skips the<br> bridge and VXLAN overlay, routing pods directly instead (often via BGP) — <br> faster on a flat LAN, and it only falls back to a tunnel device (`tunl0`<br> for IP-in-IP, `vxlan.calico` for VXLAN) when nodes aren’t on the same<br> network:</li></ul><pre>ip link show | grep cali # cali&lt;hash&gt;@ifN - one veth per pod, no bridge<br> ip route | grep &#39;proto bird&#39; # routes to other nodes&#39; pod blocks, no encapsulation</pre><ul><li><strong>Cilium </strong>— the eBPF one. Attaches programs at the kernel level instead of<br> leaning on iptables/nftables, can replace kube-proxy’s Service<br> load-balancing entirely, and can enforce policy up to L7 (allow `GET<br> /health`, block `POST /admin`). More power, and a great power comes with… (a lot more moving parts sorry spidey) :</li></ul><pre>ip -d link show cilium_vxlan # Cilium&#39;s own VXLAN device, in overlay mode<br> ip link show | grep cilium # cilium_host / cilium_net - its gateway into the eBPF datapath</pre><p><strong>conntrack (connection tracking)</strong>. You’re already using it without naming it — every ct state { established, related } accept line in your own nftables rules is conntrack. It matters doubly for k8s because kube-proxy’s default mode implements Services (the stable virtual IP that load-balances to pod IPs) as a pile of iptables/nftables DNAT rules riding on conntrack. A Service “randomly” dropping connections is very often a conntrack table problem, not a k8s problem. Here’s a few commands to better understand your own k8s networking.</p><p><strong>systemd unit states</strong>, not just “running or not.” You just learned this the hard way with zram: enabled: false isn’t the same as masked, and the difference is exactly what let swap resurrect itself. k3s itself is a systemd unit, journalctl -u k3s -f is your first debugging move for basically everything, and “why didn’t my config change take effect” is usually a missing systemctl daemon-reload or restart.</p><p><strong>DNS basics</strong>. CoreDNS is doing service discovery for the whole cluster. If you don’t understand /etc/resolv.conf, search domains, and how a name like myapp.default.svc.cluster.local resolves, “my pod can’t reach that service by name” turns into an unfalsifiable mystery instead of a two-command check (nslookup from inside the pod, kubectl get svc).</p><p><strong>PKI/TLS basics </strong>— certs, SANs, CA trust. You already hand-wrote tls-san entries into your k3s config. Every “x509: certificate is valid for X, not Y” error is this exact concept biting you — a cert issued for one IP/hostname being presented to a client that connected via another. Understanding a Subject Alternative Name isn’t optional once you’re running your own CA (which k3s does, silently, for you).</p><p>Lower priority but worth having: OverlayFS/union filesystems (how container images are layered, why containerd assembles a root fs the way it does), and process signals (SIGTERM vs SIGKILL, why a container without a real init can leave zombies — relevant the moment you tune terminationGracePeriodSeconds).</p><p><strong>Boards:</strong> Raspberry Pi or Orange Pi single-board computers. On the control plane, buy one ecosystem, not a zoo mixing vendors there means mixing OS images, quirks, and kernel surprises across the nodes running etcd, and you’ll pay for it in debugging exactly the part of the cluster you can least afford to destabilize.</p><blockquote>Pro tip: keep the control plane homogeneous memory and network.<br>etcd runs at the speed of its slowest member and the worst link between<br>them: it’s Raft consensus, so every write waits for a majority to acknowledge,<br>on a ~100 ms heartbeat with a ~1 s election timeout. A lone 4 GB board in a trio<br>of 8 GB boards is a weak link on the memory axis. A lone Wi-Fi board in a trio<br>of wired boards is a weak link on the network axis and network jitter hurts<br>etcd faster than a slow CPU does. The same goes for a “control-plane” node that<br>lives at a different site on a different network path, :It might add resilience but the extra latency and occasional packet loss reads to Raft as flakiness, and flaky etcd members cause elections, and elections you can feel.</blockquote><blockquote>I’ve run both configurations. Wi-Fi and cross-location nodes work fine in the<br><strong>user/workload pool </strong>I run mine that way on purpose. But every time I tried<br>it for a **system/control-plane** node, the cluster was measurably less stable<br>than the moment I put all three etcd members on wired ethernet, on the same<br>LAN, on matched hardware. Match the three: same power class, same wire, same<br>network. Save the diversity for the pool that can tolerate it.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/657/1*6UPzEkqZTcWWfdnIcICZ9Q.png" /><figcaption>Figure 2: control plane on 3 system pools, showing one system pool on orangepi, not perfect homogeneus but with same memore/capacity ration than the pi 5 adn wired</figcaption></figure><h4>The raw setup (do it by hand once)</h4><p>Automate later. First, understand the seven things that actually matter.</p><p>1) Same OS on every board. Flash the same Debian/Ubuntu ARM64 image. Create the same sudo user, install your SSH public key, and — this matters for automation later — give that user passwordless sudo. Ansible can’t answer an interactive password prompt.</p><p>2) Memory cgroups must be enabled. Kubernetes needs the kernel to expose the memory cgroup controller. On Raspberry Pi it’s off by default — append `cgroup_enable=memory cgroup_memory=1` to the boot `cmdline.txt` and reboot. Verify the right way: look for `memory` in<br>`/sys/fs/cgroup/cgroup.controllers`. (Do not test for<br>`/sys/fs/cgroup/memory.max` at the root — that file doesn’t exist even on a<br>perfectly healthy system, and it’ll send you chasing a ghost.)</p><p>3) Turn swap off. k3s and etcd want swap disabled. On most boards that’s<br>`swapoff -a` plus commenting the line out of `/etc/fstab`. On Orange Pi, swap is<br>provided by a zram service, not fstab — so it comes back on reboot unless<br>you disable the service itself.</p><p>4) Pick a zero-trust overlay network. Tailscale (or Netbird, Nebula,<br>headscale if you want to self-host the coordinator). This gives you encrypted SSH and API access to every node from anywhere, with no port-forwarding and no public exposure. Two things people get wrong:</p><p>- The overlay does not replace your host firewall. `sshd` still checks its<br> own rules regardless of which network the packet arrived on.<br>- Don’t run etcd over the overlay. etcd is chatty and latency-sensitive; run<br> its traffic over the wired LAN. Use the overlay for SSH and the human-facing API, not for cluster consensus.</p><p>5) Open the cluster ports on the firewall. Run a real firewall (`nftables`,<br>default `policy drop`), but you must let the nodes talk to each other on the<br>LAN, or they will never form a cluster:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1TunnlJDYY9HSCsUbSHhzQ.png" /><figcaption>▎ Figure 3: The nftables ports I had to open for k3s to talk to itself: 6443 (API server), 2379/2380 (etcd), 10250 (kubelet), and 8472/udp (flannel VXLAN). Pushed to every node identically through an Ansible playbook that validates the ruleset with nft -c -f before it’s allowed to replace the live firewall.</figcaption></figure><p>Scope these to your LAN subnet. Lock SSH to the overlay. This single step is the difference between “cluster” and “three lonely computers.”</p><p>6) Install k3s. First node initializes the embedded-etcd cluster; the others<br>join it with the same shared token:</p><p><strong>First node creates the cluster:</strong></p><pre>curl -sfL https://get.k3s.io | K3S_TOKEN=&quot;$SHARED_TOKEN&quot; \<br> INSTALL_K3S_EXEC=&quot;server - cluster-init - node-ip &lt;ip&gt; - advertise-address &lt;ip&gt;&quot; sh -</pre><p><strong>Every other server joins it:</strong></p><pre>curl -sfL https://get.k3s.io | K3S_TOKEN=&quot;$SHARED_TOKEN&quot; \<br> INSTALL_K3S_EXEC=&quot;server - server https://&lt;first-node-ip&gt;:6443 - node-ip &lt;ip&gt;&quot; sh -</pre><p>7) Verify. `kubectl get nodes -o wide` should list every server as<br>`control-plane,etcd` and `Ready`. If a node shows only itself, or is missing the `etcd` role, stop — it built its own cluster (see the war stories).</p><h3><strong>Then hand it to Ansible</strong></h3><p>Here’s the twist the “<a href="http://mac&/#39;s%20Tech%20Blog%20RSS%20Feed%20Dear%20friend,%20you%20have%20built%20a%20Kubernetes%202024-11-23%20%20This%20post%20will%20make%20more%20sense%20if%20you%20first%20read%20Dear%20Sir,%20You%20Have%20Built%20a%20Compiler.%20%20Dear%20friend,%20%20I%20am%20afraid%20to%20inform%20you%20that%20you%20have%20built%20a%20Kubernetes.%20I%20know%20you%20wanted%20to%20&quot;choose%20boring%20tech&quot;%20to%20just%20run%20some%20containers.%20You%20said%20that%20&quot;Kubernetes%20is%20overkill&quot;%20and%20&quot;it&#39;s%20just%20way%20too%20complex%20for%20a%20simple%20task&quot;%20and%20yet,%20six%20months%20later,%20you%20have%20a%20pile%20of%20shell%20scripts%20that%20do%20not%20work%E2%80%94breaking%20every%20time%20there&#39;s%20a%20slight%20shift%20in%20the%20winds%20of%20production.%20%20Surely,%20switching%20to%20Docker%20Compose%20will%20be%20the%20end%20of%20your%20woes;%20at%20least%20that%20way,%20someone%20else%20maintains%20a%20standard%20config%20file%20format%20for%20you%20to%20use.%20But%20wait,%20Compose%20is%20still%20not%20a%20holistic%20solution.%20&quot;Do%20I%20really%20need%20a%20separate%20solution%20for%20deployment,%20rolling%20updates,%20rollbacks,%20and%20scaling?&quot;%20you%20ask%20yourself?%20Surely%20not.%20Your%20app%20is%20so%20simple;%20a%20backend,%20a%20reverse%20proxy,%20postgres,%20and%20a%20job%20runner.%20So%20you%20march%20on,%20and%20add%20another%20few%20sections%20to%20your%20deploy.sh%20script,%20certain,%20that%20this%20will%20be%20the%20last%20of%20what%20you%20need%20to%20do%20to%20maintain%20this%20pile%20of%20hacks.%20%20Ah,%20but%20wait!%20Inevitably,%20you%20find%20a%20reason%20to%20expand%20to%20a%20second%20server.%20While%20it&#39;s%20true%20a%20single%20server%20can%20go%20a%20long%20way%20many%20things%20can%20force%20this%20decision,%20such%20as%20the%20need%20for%20special%20hardware,%20high%20availability,%20or%20the%20speed%20of%20light.%20Tired,%20you%20parameterize%20your%20deploy%20script%20and%20configure%20firewall%20rules,%20distracted%20from%20the%20crucial%20features%20you%20should%20be%20working%20on%20and%20shipping.%20One%20of%20your%20team%20members%20suggests%20connecting%20the%20servers%20with%20Tailscale:%20an%20overlay%20network%20with%20service%20discovery.%20After%20that%20you%20will%20know,%20yes%20know,%20that%20there%20is%20no%20tougher%20complexity%20hurdle%20to%20clear%20than%20the%20networking.%20%20Except%20if%20you%20quit%20or%20go%20on%20vacation,%20who%20will%20maintain%20this%20custom%20pile%20of%20shell%20scripts?%20The%20untested%20tarball%20handling?%20The%20inscrutable%20iptables%20rules?%20Who%20will%20know%20about%20those%20undocumented%20sysctl%20edits%20you%20made%20on%20the%20VM?%20So%20you%20add%20everything%20to%20Ansible,%20so%20that%20you%20may%20treat%20your%20VM%20as%20immutable%20and%20version-controlled.%20Certain,%20of%20course,%20that%20because%20you%E2%80%99re%20not%20using%20Kubernetes,%20it%20is%20going%20to%20be%20way%20easier%20to%20maintain%20than%20a%20Kubernetes%20cluster.%20What%20glorious%20engineering,%20you%20say%20to%20yourself.%20%20In%20the%20last%20leg%20of%20your%20journey%20to%20avoid%20building%20a%20Kubernetes,%20your%20manager%20tells%20you%20that%20your%20app%20needs%20to%20programmatically%20spawn%20other%20containers.%20Spawning%20containers,%20of%20course,%20requires%20you%20to%20mount%20the%20Docker%20socket%20in%20your%20web%20app,%20which%20is%20wildly%20insecure.%20Not%20my%20problem,%20your%20manager%20says.%20So%20you%20write%20a%20separate%20service%20that%20exposes%20a%20safe%20subset%20of%20the%20Docker%20API%20to%20your%20web%20app.%20Done%20at%20last,%20you%20say%20to%20yourself,%20without%20having%20to%20build%20a%20Kubernetes.%20%20A%20standard%20config%20format,%20a%20deployment%20method,%20an%20overlay%20network,%20service%20discovery,%20immutable%20nodes,%20and%20an%20API%20server.%20Dear%20friend,%20you%20have%20built%20a%20Kubernetes.%20%20Addressed%20to,%20%20Those%20who%20wanted%20to%20avoid%20Kubernetes.%20%20PS:%20I%20don&#39;t%20mean%20to%20imply%20that%20you%20can%20never%20roll%20your%20own%20deployment%20method%20that%20fits%20your%20needs%20better%20than%20Kubernetes,%20or%20that%20nothing%20will%20ever%20be%20better%20than%20Kubernetes.%20I%20just%20want%20to%20caution%20you,%20my%20friend,%20to%20make%20sure%20you%20understand%20the%20problems%20Kubernetes%20solves%20before%20dismissing%20it%20as%20overly-complex.%20Home%20%20This%20work%20is%20licensed%20under%20CC%20BY-SA%204.0">you accidentally built a Kubernetes</a>” crowd already knows: the moment you want this to be repeatable — reimage a dead SD card, add a fourth node, recover from a mistake — you need the whole thing in code. The swap tweak, the cgroup edit, the firewall ruleset, the Tailscale enrollment, the k3s install, the node labels and taints. Left in your head and your shell history,they rot.</p><p>So you put it in Ansible. Not because Ansible is fun, but because “immutable, version-controlled nodes” is a real requirement and this is how you get it. The division of labor that keeps you sane:</p><p>- Ansible owns the OS: prep, swap, cgroups, Tailscale, the firewall, the k3s<br> install, and node roles.<br>- A GitOps tool (Argo CD, Flux) owns the applications once the cluster<br> exists. Don’t let both manage the same object.</p><p>The playbook does exactly the seven raw steps above, idempotently, across every board with one command. You do the raw version once, to learn; you run the Ansible version forever, to live.</p><p>Actually using it: node pools</p><p>Here’s where owning heterogeneous hardware pays off. You have a beefy board with an NPU, a couple of plain servers, maybe a gaming PC that’s only on at night. You don’t want your web app landing on the inference box, and you definitely don’t want random workloads on your control-plane nodes. The tool for this is pools, and it’s just two primitives:</p><p>- Labels describe a node (“this one has the NPU”).<br>- Taints repel workloads unless they explicitly tolerate them (so only<br> things meant for that node land there).</p><p>Create an inference pool, label the board, then taint it so nothing lands<br>there by accident:</p><pre>kubectl label node orangepi-infer-01 pool=inference hardware=rk3588-npu<br>kubectl taint node orangepi-infer-01 pool=inference:NoSchedule</pre><p>Send a workload to it — the pod opts in with a `nodeSelector` (I want that<br>pool) and a matching `toleration` (I accept the taint):</p><pre>spec:<br> nodeSelector:<br> pool: inference<br> tolerations:<br> - key: pool<br> operator: Equal<br> value: inference<br> effect: NoSchedule</pre><p>That’s the whole pattern. Three pools cover most home clusters:</p><p>Figure 3: k3s ports I discovered needed opening within my nft tables, which I also automated with ansible</p><p>Because it’s all labels, scheduling reads like plain English: “run this on a<br>node in the inference pool.” And because Ansible sets the labels and taints at install time, a reimaged node rejoins with its role intact — you never<br>hand-label a fresh board.</p><blockquote>Pro tip: Drop the `sudo` from `kubectl`. k3s writes its admin config to<br>`/etc/rancher/k3s/k3s.yaml` as root-only, so out of the box every command needs<br>`sudo kubectl`. Don’t “fix” that by making the file world-readable — it holds<br>cluster-admin credentials. Instead, copy it once to your user’s default<br>kubeconfig path:</blockquote><pre>mkdir -p ~/.kube<br>sudo cp /etc/rancher/k3s/k3s.yaml ~/.kube/config<br>sudo chown &quot;$USER:$USER&quot; ~/.kube/config</pre><p>`kubectl` reads `~/.kube/config` automatically — no `sudo`, no `KUBECONFIG` export, and the credentials stay in a file only you can read. (In the Ansible version this is a task, so every node comes up ready to use.)</p><p>A few day-two commands worth muscle memory:</p><pre>kubectl get nodes -o wide # who&#39;s in the cluster, and where<br>kubectl get pods -A -o wide # every pod, and which node it&#39;s on<br>kubectl describe node orangepi-infer-01 # labels, taints, capacity, pressure<br>kubectl top nodes # live CPU/RAM (needs metrics-server)</pre><p><strong>Where it breaks (so it breaks less for you)</strong></p><p>Every one of these cost me time. Consider it paid forward.</p><ul><li>The node that formed its own cluster (split-brain). A joining node given an empty or mismatched token doesn’t error it silently falls back to its own single-node SQLite cluster. Tell: it’s `control-plane` with no `etcd` role and sees only itself. My root cause was fetching the token by SSH-ing node-to-node while the firewall (correctly) blocked LAN SSH — so the fetch timed out, the token came back empty, and k3s cheerfully bootstrapped a fresh cluster. Fetch the token from an admin machine, never between nodes.</li></ul><pre>  export K3S_CLUSTER_TOKEN=&quot;$(ssh ares@&lt;Admin pc/first-node-ip&gt; sudo cat /var/lib/rancher/k3s/server/token)&quot;<br>  echo &quot;token_len=${#K3S_CLUSTER_TOKEN}&quot;   # must be non-zero before you go further<br>  ansible-playbook -i inventory/hosts.yml site.yml --limit &lt;the-node-you&#39;re-joining&gt;</pre><p>- <strong>The firewall that drops your cluste (nftables).</strong> A newly added control-plane node with `policy drop` and no LAN allow-rules can reach the others but the others can’t reach it on `2380`. etcd logs `authentication handshake failed: context deadline exceeded` and loops forever. A new node needs its firewall applied before it can join etcd — the ports, not the SSH lockdown, are what matter.</p><p>-<strong> cgroups “not showing.”</strong> I tested for `/sys/fs/cgroup/memory.max` at the root, saw nothing, and panicked. That file doesn’t exist at the root even when cgroups are fine. Check `/sys/fs/cgroup/cgroup.controllers` for `memory` instead. (And on Raspberry Pi, remember the `cmdline.txt` + reboot.)</p><p>- <strong>The broken package mirror</strong>. Orange Pi images ship an extra `docker-ce`<br> repository pointed at a regional mirror that occasionally serves a truncated<br> index file. `apt update` then fails as a whole — blocking base packages that<br> live in a completely healthy repo. Two fixes: disable the junk repo (k3s uses<br> containerd, not Docker), and make your automation’s cache-refresh<br> best-effort so one flaky optional repo can’t wedge the whole run.</p><p>-<strong> SSH keys that “work” but don’t</strong>. Orange Pi images set<br> `PubkeyAuthentication no`. `ssh-copy-id` reports success and drops your key in `authorized_keys`, but sshd never offers it so key auth silently fails and Ansible (which can’t type a password) is dead in the water. Fix `sshd_config` first, on Orange Pi, every time.</p><ul><li><strong>“relay” instead of “direct.” </strong>A wired node showing a relayed overlay path<br> usually means its own firewall is blocking the UDP hole-punch. It still works; it’s just slower and routed through someone else’s server , mildly ironic for a project about owning your infrastructure.</li><li><strong>Swap that resurrects itself. </strong>You disabled swap, rebooted, and it’s back<br> because on Orange Pi it’s a zram service, not an `/etc/fstab` entry. Mask<br> the service, not the file.</li><li><strong>Nodes with names you don’t recognize</strong>. A manual install registers the node under its OS hostname (`raspberrypi`, `orangepi5pro`). Set the node name explicitly so your cluster reads like your inventory, not like a factory<br> default.</li><li><strong>Cost</strong>, honestly. Street prices as of mid-2026 for exactly the control plane<br>in Figure 2 — two Raspberry Pi 5 (8 GB) plus an Orange Pi 5 Plus (8 GB) as<br>the third system node, and an Orange Pi 5 Pro (4 GB) as the user-pool agent:</li><li>| Board | Price (mid-2026) |<br>| Raspberry Pi 5, 8 GB (×2) | ~$199 each |<br>| Orange Pi 5 Plus, 8 GB | ~$220 |<br>| Orange Pi 5 Pro, 4 GB | ~$176 |<br>| Total | ~$794 |</li></ul><p>Add power supplies, cases, and NVMe/eMMC storage and budget another<br>$150–250 across four boards. Worth flagging: these prices are inflated by<br>the 2026 AI-driven DRAM shortage — the same Pi 5 8 GB board was roughly $80 at launch. If memory prices have come back down by the time you’re reading this, expect the whole cluster closer to $400–450 all-in. Either way, it’s a fraction of a month of the managed equivalent, and it’s yours forever.</p><p>Sure, the capex-versus-opex crowd will point out that hardware depreciates<br>and eventually hits end-of-life, and that over a long enough horizon renting<br>can come out cheaper once you count the refresh cycle. Fair — at some<br>horizon that’s true. I won’t even claim owning the hardware is what forces<br>you to ship: a metered cloud bill applies the same pressure, just shaped<br>differently. What local actually wins on is more concrete. There’s no<br>shared-host tax — no noisy neighbor stealing your CPU cycles, no markup<br>baked in for someone else’s infrastructure. There’s privacy a rented box<br>can’t match by definition: the data never leaves your house. And there’s<br>raw performance-per-dollar a Raspberry Pi 5 with an NVMe drive will<br>comfortably outrun the cheap burstable VM tier cloud providers steer you<br>toward at a similar price.</p><p><strong>The payoff to walk the Mordor path (Self hosted over Managed)</strong></p><p>You end up with a three-node Kubernetes cluster that fits in a shoebox, costs nothing to run, answers to no one’s pricing page, and — because it’s all in Ansible. comes back from the dead with one command. That last part matters more than it looks: Ansible’s whole power is its simplicity. No agents on the boards, no daemons to babysit, no controller to rent, just SSH and YAML you can actually read. When a node dies, you don’t open a support ticket; you re-run a playbook. More importantly, you now understand the machine: not as a black box you rent, but as something you built and can reason about.</p><p>The lesson from “<a href="https://www.macchaffee.com/blog/2024/you-have-built-a-kubernetes/">Dear friend, you have built a Kubernetes</a>” runs both ways. Don’t reach for Kubernetes to run a single container. But don’t dismiss it as “too complex” before you understand the problems it solves, because if you’re running real workloads across real machines, you’ll solve those problems one way or another. Better to solve them with the boring tool that a thousand people maintain, on hardware that’s yours.</p><p>Owning the compute buys you independence: room to actually look under the hood instead of trusting someone else’s abstraction. In a future increasingly shaped by a handful of AI monopolies, I think that’s becoming less a hobbyist’s indulgence and more one of the few real paths to freedom left. Rent nothing you don’t have to.</p><p>The cyberpunk future we were promised was chrome and neon. The one arriving is quieter: compute consolidated into a handful of towers, prices going over the roof, intelligence rented by the token, and everyone’s data flowing through systems they do not own. One company even took its name from Tolkien’s seeing stones, but it is only one example of a larger shift: scattered data gathered into centralized systems that can answer questions about anyone. That machinery does not have to be evil. It only has to exist, be powerful, and be available mainly for rent. The dystopia was never going to be a robot; it was going to be a query box that everyone else’s data flows through. <br> <br> So consider this an invitation to build something else. Self-host what you can. Keep sensitive data on machines you control, reuse DDR4-era hardware instead of discarding it, and bridge the gap to newer systems through upgrades, optimization, and clever engineering — not by waiting for a vendor to decide what remains useful. Run open-weight models on hardware nobody respects. Every board you repair, every cluster you operate, and every service you bring back <br> with a playbook is a small refusal to let compute — and the intelligence built on top of it be monopolized. That is what a freer future looks like: not one provider that solves all, but ordinary people owning enough of the stack to make their own choices.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b5889bee5db9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Legacy Ingress to Gateway API: Why Kubernetes Routing Is Evolving]]></title>
            <link>https://medium.com/@datatomas/from-legacy-ingress-to-gateway-api-why-kubernetes-routing-is-evolving-4a014d041b06?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a014d041b06</guid>
            <category><![CDATA[ingresscontroller]]></category>
            <category><![CDATA[nginx]]></category>
            <category><![CDATA[api-gateway]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[traefik]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Sun, 14 Jun 2026 04:16:25 GMT</pubDate>
            <atom:updated>2026-06-14T04:54:41.531Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9w_4nCDSYBv9_eMz2R5gSQ.png" /></figure><p>When teams migrate from classic Kubernetes Ingress to Gateway API, the first reaction is usually: “Why did one object become multiple objects?” I have seen that question come up a lot during larger Kubernetes migrations. In one migration I lead across 30+ clusters, this was one of the biggest mental shifts for application teams. With the old Ingress model, the story was simple: an Ingress pointed to a Service, and the ingress controller handled the rest. That simplicity made it easy to adopt, but it also meant a lot of responsibility ended up concentrated in one place.</p><p>I recently rebuilt this flow in a small Minikube <a href="https://github.com/datatomas/minikube-affinity-lab">lab</a> so I could test the legacy Ingress pattern, NGINX Gateway Fabric, Traefik Gateway API, cert-manager, and Cloudflare DNS-01 side by side.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mnlBhZ_d9g5fwvLLCkrMww.png" /><figcaption>Figure 1: Legacy Ingress Controller Pattern</figcaption></figure><p>In the legacy model, a typical flow looked like this: a user hit a domain, DNS pointed traffic toward the cluster, an ingress controller like NGINX Ingress or <a href="https://www.linkedin.com/company/traefik/">Traefik Labs</a> Traefik received the request, and the Ingress object routed it to the right Service. TLS was usually handled by referencing a Kubernetes secret, often created manually or by cert-manager. For many apps, that model worked well. It was easy to explain and easy to debug when everything lived in one namespace and one team owned most of the stack.</p><blockquote><em>That became very real for a lot of Kubernetes teams during the ingress-nginx “IngressNightmare” vulnerability. It was a reminder that ingress controllers are not just simple traffic routers. They are powerful control-plane-adjacent components, often watching resources across namespaces and translating YAML into live edge configuration. When something goes wrong there, the blast radius can be much larger than a single app route.</em></blockquote><p>The problem starts to show up at scale. In large environments, the Ingress object often becomes a mix of application routing, TLS behavior, controller-specific annotations, security assumptions, and platform policy. Application teams want to expose services quickly, while platform teams need to control how traffic enters the cluster. Over time, the simple model becomes harder to govern. The YAML stays small, but the operational meaning behind it gets bigger.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ImW4dj6TDXmocvBzASyr3g.png" /><figcaption>Figure 2: The New Way: Gateway -&gt; HTTPRoute -&gt; Service</figcaption></figure><p>Gateway API changes that model by separating responsibilities. Instead of thinking Ingress -&gt; Service, the new model is closer to Gateway -&gt; HTTPRoute -&gt; Service. At first this feels like extra YAML, but the point is not just more objects. The point is clearer ownership. A platform team can own the Gateway, which represents the shared entry point into the cluster. An application team can own the HTTPRoute, which describes how its app wants traffic routed. The controller can be NGINX Gateway Fabric, Traefik, Envoy Gateway, Istio, or another implementation. Shoutout to <a href="https://www.linkedin.com/company/traefik/">Traefik Labs</a> here too. Testing the same Gateway API flow with Traefik made the namespace and route attachment behavior much more obvious. The 404 was annoying, but it was also a good reminder that different controllers can expose different assumptions in your manifests. Not to mention their built in integration with <a href="https://www.linkedin.com/company/k3sio/">K3s</a> is what made me fall in love of tha k8s distribution.</p><p>The same idea applies to supporting components like CRDs, certificates, and controller installs. In this lab I kept the Gateway API CRDs separate from the controller installation, then used Helm to install either NGINX Gateway Fabric or Traefik. <strong>That makes the shared API layer independent from the implementation. </strong>For TLS, I used cert-manager with <a href="https://www.linkedin.com/company/lets-encrypt/">Let’s Encrypt</a> Let’s Encrypt and Cloudflare DNS-01 instead of manually uploading a PFX into a Kubernetes secret. Also, shoutout to Let’s Encrypt for making publicly trusted certificates free and automatable. That changed the default expectation for TLS in a big way. Manual PFX secrets still work, and cloud options like Key Vault can also be a strong choice, especially in enterprise environments, but they come with their own operational overhead, renewal processes, permissions, and cost considerations. For a Kubernetes-native workflow, cert-manager keeps the certificate lifecycle close to the cluster while still making the final result just a normal TLS secret that Gateway or Ingress can use.</p><p>You can check an implementation of Nginx Gateway Fabrik in the <a href="https://github.com/datatomas/minikube-affinity-lab">git</a> repository as well.</p><p>That separation becomes especially important when namespaces are involved. This is where I recently got a very real reminder in a Minikube lab. I was testing the same Streamlit app behind NGINX Gateway Fabric and then Traefik Gateway API. The certificate was valid, the GatewayClass existed, the Gateway existed, the HTTPRoute existed, and Traefik was reachable. Still, every request returned 404 page not found. The problem was not DNS, TLS, or Minikube networking. The problem was where the route lived and what it was allowed to reference.</p><p>For example, this route references a Service in another namespace:</p><pre><br>apiVersion: gateway.networking.k8s.io/v1<br>kind: HTTPRoute<br>metadata:<br>  name: streamlit-route<br>  namespace: gateway-system<br>spec:<br>  parentRefs:<br>  - name: traefik-gateway<br>  hostnames:<br>  - app.example.com<br>  rules:<br>  - backendRefs:<br>    - name: streamlit-service<br>      namespace: app<br>      port: 80<br></pre><p>In Gateway API, an HTTPRoute does not automatically attach to any Gateway it names. If the route is in a different namespace from the Gateway, the Gateway listener must allow routes from that namespace. And if the route points to a backend Service in another namespace, the backend namespace must explicitly allow that reference with a ReferenceGrant. In my case, the working shape was: the HTTPRoute lived with the Traefik Gateway, the backend Service lived in the app namespace, and a ReferenceGrant allowed the route to reference that service.</p><p>Gateway API introduces two separate permission checks that do not exist in the classic Ingress model.First, the Gateway must allow the route to attach to it</p><pre>apiVersion: gateway.networking.k8s.io/v1beta1 <br>kind: ReferenceGrant <br>metadata: name: allow-gateway-route namespace: app <br>spec: from: <br>  - group: gateway.networking.k8s.io <br>    kind: HTTPRoute <br>    namespace: gateway-system<br>     to: <br>     - group: &quot;&quot; kind: Service</pre><p>This tells the Gateway that HTTPRoutes from other namespaces may attach to it.</p><p>Next, an HTTPRoute can reference that Gateway from another namespace:</p><pre>apiVersion: gateway.networking.k8s.io/v1 kind: HTTPRoute <br>metadata: <br>  name: streamlit-route <br>  namespace: app <br>  spec: <br>    parentRefs: <br>    - name: traefik-gateway <br>      namespace: gateway-system <br>    rules:  <br>      - backendRefs:<br>        - name: streamlit-service port: 80</pre><p>If the route references a Service in the same namespace, no additional permissions are required.</p><p>However, if the route references a Service in another namespace:</p><pre>backendRefs: <br>- name: streamlit-service <br>  namespace: backend <br>  port: 80</pre><p>the target namespace must explicitly allow that reference:</p><pre>apiVersion: gateway.networking.k8s.io/v1beta1 <br>kind: ReferenceGrant metadata: <br>name: allow-gateway-route <br>namespace: backend <br>spec: <br>  from: <br>    - group: gateway.networking.k8s.io <br>    kind: <br>    HTTPRoute <br>    namespace: app <br>  to: <br>  - group: &quot;&quot; <br>  kind: Service</pre><p>That is the part I think many teams will trip over during migrations. Gateway API is not just “Ingress v2.” It is a more explicit model for shared infrastructure. Classic Ingress made many boundaries feel invisible. Gateway API makes those boundaries part of the API. That is a good thing for platform engineering, but it requires teams to understand the relationship between GatewayClass, Gateway, HTTPRoute, Service, and ReferenceGrant.</p><p>In other words, Gateway API performs two distinct checks:</p><ol><li>Can this HTTPRoute attach to that Gateway? (allowedRoutes)</li><li>Can this HTTPRoute reference that Service? (ReferenceGrant)</li></ol><p>This separation is one of the biggest conceptual differences from the classic Ingress model and is what enables safer multi-team and multi-namespace deployments.</p><p>The practical takeaway is simple: if everything lives in one namespace, Gateway API feels straightforward. If platform and app resources live in different namespaces, you need to think about two permissions: can this route attach to that Gateway, and can this route reference that backend Service? Once that clicks, the extra objects stop feeling like noise and start feeling like guardrails.</p><p>It also helps to remember the separation of responsibilities in the Gateway API model. A Gateway API implementation such as NGINX Gateway Fabric, Traefik, Envoy Gateway, or Istio provides the controller. A GatewayClass defines which implementation will manage Gateways. A Gateway represents the entry point into the cluster, while HTTPRoute objects define how traffic should be routed to applications. Understanding that relationship makes it easier to see Gateway API not as &quot;Ingress with more YAML,&quot; but as a more explicit and portable model for managing traffic, ownership, and trust boundaries in Kubernetes.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a014d041b06" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Azure DevOps Deployment Governance: ManualValidation@1 vs Environments]]></title>
            <link>https://medium.com/azure-made-easy/azure-devops-deployment-governance-manualvalidation-1-vs-environments-456bce43500d?source=rss-e0afe4ea32bf------2</link>
            <guid isPermaLink="false">https://medium.com/p/456bce43500d</guid>
            <category><![CDATA[environment]]></category>
            <category><![CDATA[azure]]></category>
            <category><![CDATA[manualvalidation]]></category>
            <category><![CDATA[azure-devops]]></category>
            <dc:creator><![CDATA[Tomas Suarez]]></dc:creator>
            <pubDate>Mon, 18 May 2026 16:50:36 GMT</pubDate>
            <atom:updated>2026-05-18T16:50:36.387Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YHy7yq2CezeNPhKZSJmoGA.png" /><figcaption>Figure 1: approvals and check home page from within environments</figcaption></figure><p>Azure DevOps gives you two common ways to pause or control deployments: ManualValidation@1 and Pipeline <a href="https://learn.microsoft.com/en-us/azure/devops/pipelines/process/environments?view=azure-devops">Environments</a>.</p><p><a href="https://learn.microsoft.com/en-us/azure/devops/pipelines/tasks/reference/manual-validation-v1?view=azure-pipelines">ManualValidation</a>@1 is the quick and casual option. It lives inside the YAML pipeline, pauses the run, and waits for someone to resume or reject it. It works well for simple checks, demos, what-if reviews, or internal validation steps.</p><p>But for production-level governance, Pipeline Environments are the cleaner model.</p><p>With Environments, you can create deployment targets like iac-prb, iac-cal, and iac-prd. Each environment can have its own approvals, approver groups, branch controls, business hours, and exclusive locks.</p><p>Is it more annoying to configure? Yes. Naturally, enterprise safety comes wrapped in extra clicks.</p><p>But it is also safer because the approval rules live outside the YAML. A pipeline author can reference the environment, but they cannot simply remove the approval logic from the pipeline file.</p><p>For example:</p><p>- iac-prd can require the IAC Deployment Approvers group.<br>- portal-prd can require the Portal Deployment Approvers group.<br>- networking-prd can require a different technical group.<br>- production environments can block deployments from unapproved branches.</p><p>That separation matters.</p><p>YAML should define the deployment flow. <br>The Environment should define whether the deployment is allowed.</p><p>For real production deployments, I would use Azure DevOps Environments as the governance boundary, and keep ManualValidation@1 only for lightweight human checkpoints such as reviewing a what-if output before continuing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c6FDPIayVZxDxrcHmCBGUQ.png" /><figcaption>Figure 2: the image shows how Azure DevOps <strong>Exclusive Lock</strong> handles concurrent deployments to the same Environment.</figcaption></figure><ul><li><strong>RunLatest</strong> means only the newest waiting deployment continues. This is useful for app releases where the latest version replaces older queued runs.</li><li><strong>Sequential</strong> means every queued deployment runs one by one. This is safer for IaC, RBAC, networking, and policy changes where skipping a run could skip an important infrastructure update.</li></ul><p>At the end of the day, ManualValidation@1 is a pipeline checkpoint.</p><p>Azure DevOps Environments are a governance boundary.</p><p>That difference matters.</p><p>If the deployment is casual, temporary, or only needs a quick human review, ManualValidation@1 is enough.</p><p>If the deployment touches production, shared infrastructure, RBAC, networking, policies, or anything that can ruin someone’s afternoon, use Environments with approvals and checks.</p><p>Yes, it means creating environments, groups, and approver mappings.</p><p><strong>Yes, it is more setup</strong>. But that is the tradeoff: less flexibility inside YAML, more control outside of it.</p><p>And for <strong>production</strong>, that is exactly the point.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=456bce43500d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/azure-made-easy/azure-devops-deployment-governance-manualvalidation-1-vs-environments-456bce43500d">Azure DevOps Deployment Governance: ManualValidation@1 vs Environments</a> was originally published in <a href="https://medium.com/azure-made-easy">Azure Made Easy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>