I have already written about how FediMeteo was born, and about how HAProxy helps reduce the number of requests that reach snac.

Seen from the outside, FediMeteo almost seems still. There is a static homepage, regenerated every hour. There are the city pages, with their forecasts. There are RSS feeds waiting to be fetched, JSON objects waiting to be requested, Fediverse instances refreshing data, subscribing, unsubscribing, retrieving profiles, and reading notes.

That is the visible part.

Behind it, however, FediMeteo is much more than a homepage, a few ActivityPub accounts, and a well-behaved reverse proxy. It is a chain of small pieces, in proper Unix style, each trying to do one thing and do it as well as possible.

That chain, although almost invisible from the outside, was not born already tidy. It changed, was rewritten, adapted to new countries, timezones, ambiguous city names, external service limits, and also to my own mistakes.

Some mistakes were small. Others were much less so.

Because FediMeteo is a human project and, as such, imperfect. Imperfect in the way humans are imperfect, which today almost seems unfashionable. I like that.

The first version of the bot was almost embarrassingly simple, and I was proud of that.

It took a city name as input, asked Nominatim for the coordinates through geopy, called the Open-Meteo API for the current weather and the next several days, and printed a markdown block with current conditions, the forecast for today, the next twelve hours, and the coming days. The text was in Italian. The cities were Italian. The timezone was Europe/Rome. There was nothing to calculate.

Around the script, a small sh wrapper read a list of cities and, for each one, ran the Python program and piped its output into snac note_unlisted. A cron job ran the wrapper every six hours. The output was loose markdown, which snac happily renders, and the integration was: standard output goes into standard input. Nothing fancier than that.

I like this kind of design. It is the part of the Unix philosophy that survives even when fashions change.

When I started adding other European countries, I did not need to change much. I separated the operational logic from the localized strings, moved the strings into one JSON file per country, and spread the cron entries so that not every country posted in the same minute. Each country had its own snac instance, in its own FreeBSD jail, with its own dataset. The bot, internally, was almost the same script as before.

This worked because Europe is, in essence, two or three timezones across most of the countries I cared about.

Then I added Germany, and Germany taught me my first lesson about names.

There are several places called Neustadt in Germany. There is a Frankfurt am Main, and a Frankfurt an der Oder, and they are not the same city. There is a Halle in Saxony-Anhalt and a Halle in North Rhine-Westphalia. Asking Nominatim for "Frankfurt, Germany" produced one of the two, consistently, but not always the one I wanted. Some German users wrote to me, politely, to point out that the forecast for "their" Frankfurt was, in fact, for the other one.

I started thinking about disambiguation, but only enough to fix the immediate cases. The bot still took a single city name. The ambiguous ones I worked around by editing the cities file and hoping for the best.

In hindsight, this was the seed of what would happen later.

The United States broke every assumption the bot had grown up with.

The first problem was the number of cities. I wanted reasonable coverage at state level, which meant identifying the main cities for each of the fifty states. The list ended up at more than 1200 entries. That alone is more cities than every other country in the project combined.

The second problem was timezones. The contiguous United States covers four of them, and Alaska and Hawaii bring the total to six. A "current weather at 12:00" line generated at the same instant for New York and for Los Angeles is technically the same instant, but the two cities are living different parts of the day, and the forecast for "today" is not even quite the same window. A bot that pretended every city was on the same clock would be wrong, sometimes embarrassingly so, every single day.

The third problem was the name thing again, only larger. There are dozens of Springfields. There is a Portland in Oregon and a Portland in Maine. The Germany workaround - editing the cities file by hand and hoping Nominatim picked the right city - was clearly not going to scale to a country where the same name is also a state.

I sat with this for a couple of days before admitting what I already knew.

The bot needed to be rewritten.

What made this hard was not the rewriting itself. It was the requirement to do it without breaking everything else.

By the time I decided to add the United States, the infrastructure around the bot had grown into something I trusted. Jails, snapshots, backup jobs, cron schedules, snac instances on production paths, the HAProxy layer, the homepage cron that aggregated follower counts, and a long list of cities being processed in series every six hours. None of that knew or cared about the bot's internal shape. All of it cared, very much, about the bot's external behavior: a city name and a country code go in, valid markdown comes out, and that markdown ends up in a timeline.

So the contract was clear, even if I had never written it down anywhere. The command-line interface, the output format, the exit codes, the way the wrapper script invoked it, the structure of the JSON country configs - all of it had to keep working. Italian had to keep working. German had to keep working. The cron job that ran every six hours had to keep producing the same shape of output, just with new countries added.

What I changed was almost everything below the surface.

The city argument grew an optional __state suffix, with a double underscore as separator:

python3 main.py springfield__illinois us
python3 main.py springfield__massachusetts us
python3 main.py new_york__new_york us

A city without the suffix continued to work exactly as before, which is what every European country needed. The country config gained a timezone field that could be a fixed string or the literal "auto"; when it was "auto", the bot used timezonefinder against the resolved coordinates to determine the right zone for that specific city. Internally I separated the weather provider behind an interface, so Open-Meteo could remain the primary while MET Norway and wttr.in sat behind as alternatives, with automatic fallback when the primary failed. Units became configurable per country: temperature, wind speed, precipitation. The United States needed Fahrenheit, miles per hour, and inches. Most of Europe wanted Celsius, kilometers per hour, and millimeters. The bot now does either, on a per-country basis, without caring which is which.

I am skipping a lot of small detail here, but the principle was always the same: every new degree of freedom had to be expressible as an optional field in the config or as an optional CLI flag. If a country did not set the new field, the old behavior continued, identical to before.

I tested this by running the new bot against the old country configs and comparing the output line by line. Where it differed, it was a bug in the new bot. Not in the test.

The first cycle after deploying the rewrite was, for every country except the United States, indistinguishable from the cycle before. That was the point.

This is the part of the story I dislike telling, which is precisely why I should tell it.

At some point during the development, while debugging an Open-Meteo response that did not look right, I added a print statement to the error path that dumped the full request URL whenever something went wrong. The full URL of the Open-Meteo customer endpoint includes the apikey query parameter. The print was meant for development. I forgot to remove it.

I deployed.

The next time Open-Meteo had an outage - and small ones happen, sometimes for several minutes at a time - the bot dutifully printed the failing request URL into the post body. For every city. For every cycle that ran during the outage. The wrapper script piped the output into snac note_unlisted without complaint. The posts went out, federated across the Fediverse, with my API key sitting in the text for anyone who cared to read.

Some users were kind enough to write me and tell me. Others were less kind, and made fun of me. Both groups were correct. This should not have happened.

I reported the incident to the Open-Meteo team, who were extremely understanding. They rotated the key immediately and gave me a fresh one. I removed the debug print, and then I did the slightly more useful thing, which was to add redaction at multiple layers - in the bot's output, in the daemon's logging, and in the debug helpers themselves. URL query parameters that look like API keys are masked. Environment variables and config keys named apikey or OPEN_METEO_APIKEY are redacted before any string reaches stdout or a log file. Even JSON-like fields that include open_meteo_apikey are scrubbed if they ever appear in something the program prints.

The lesson is not "be more careful." The lesson is that debug paths leak, sooner or later, so the secrets have to be unreachable from the debug paths in the first place. Now they are.

That afternoon, when I realised what was happening, I closed everything for a minute and looked out of the window. Then I started fixing.

Nominatim is a public service, and it is generous, but it is not infinite. Every city in the project needs coordinates, and at the start of the project every cycle would re-ask Nominatim for every city. Most of the time this worked. Sometimes it did not.

There was one cycle, before I added caching, when Nominatim simply did not respond for one of my queries. The geopy call timed out. The bot raised an exception. The wrapper script gave up on that city and moved on to the next one. A few users noticed that a particular city had not received its forecast that day, and asked what had happened.

I added a coordinate cache, and I am still grateful that I did.

The cache is intentionally boring. The first time the bot resolves a city, it writes the latitude and longitude into a small file under /tmp, named after the city, and the state when present. Every subsequent run reads the file. If the file exists, no Nominatim call is made. If the file is missing, the bot calls Nominatim and writes the file. After the first successful lookup, the cache becomes the source of truth for the coordinates of that city.

This is lighter on Nominatim, faster for every cycle, and much more resilient against transient failures. It is also nice for a reason I did not anticipate.

Nominatim is a geocoder, and like every geocoder it has opinions.

I live in Ferrara, so when I added Italy I made sure Ferrara was in the list, and I checked the first cycle to make sure everything looked right. The forecast came out fine. The temperature was reasonable. The icon matched the sky outside my window. I closed the laptop and forgot about it.

Then, one evening months later, I looked more carefully at the coordinates Nominatim had returned for "Ferrara, Italy", and I realised they did not point to the city. They pointed to a location closer to the centroid of the province, which is a much larger area and mostly countryside. The forecast had been, on average, for a field somewhere outside town, not for the city center.

I am not entirely sure why I had not noticed earlier. Probably because the weather in Ferrara and the weather in the fields outside Ferrara is, on most days, indistinguishable to anyone who is not paying attention. But this is the kind of detail I do not want to leave wrong, especially for my own city.

There are other places where geocoding lands slightly off. Sometimes it is a few kilometers, sometimes a different neighborhood, sometimes genuinely the wrong place.

Because the cache is just a file per city, the fix is also just a file per city. I open the cache file, replace the latitude and longitude with the correct values, save. The next cycle uses the corrected coordinates. No code change, no redeploy, no special tooling. I keep a small list of patched cities in a separate text file, so that if I ever rebuild the cache, I do not lose the manual corrections.

This is the kind of operational simplicity I like. A cache made of plain files costs almost nothing and quietly pays back every time a small problem appears.

For every report it generates, the bot also writes a simplified English text snapshot to /tmp/<city>.txt, or /tmp/<city>__<state>.txt when there is a state.

This is intentional, and it is not a debug artifact. I am not ready to say what I am doing with it yet, but it is part of a future direction for the project. Text is a useful intermediate format, and having a clean, language-neutral representation of every forecast sitting on disk costs almost nothing and might be worth a great deal later.

I prefer to let ideas mature in private before I commit to them in public. So I will leave it at this for the moment.

A full cycle for the United States takes hours.

It is not because the work is heavy. It is because I deliberately inserted a small sleep between cities, to give snac time to dispatch the previous post before the next one is generated. With more than 1200 cities in series, even a short pause adds up. I am not in a hurry. Forecasts that arrive a few minutes apart from each other are not a problem, and the bot was already a polite citizen elsewhere. A polite cycle is fine.

The problem with a slow cycle is not the duration. The problem is what happens to it.

In the original design, the cycle was launched by cron. Every six hours, cron called the wrapper script, the wrapper iterated through the cities file, and for each city it ran the bot and piped the output into snac. There was no scheduler in the project at all. Cron was the scheduler. The wrapper was just a loop.

Restarting snac was harmless. The wrapper would call snac note_unlisted per city, and if snac happened to be unavailable for a moment, that single call might fail, but the loop kept moving and snac was usually back within seconds. Snac itself was not what held the cycle together.

What held the cycle together was the wrapper process. And the wrapper process lived inside the jail.

If the FreeBSD jail was restarted while the wrapper was running, the loop stopped wherever it happened to be. The cron schedule did not care. Six hours later, the next cron tick started a new cycle from the first city, and the cities that had been about to be processed at the moment of the restart were simply skipped for that window. For the United States, this could mean several hundred cities going without an update.

There was a worse case, and it took me longer than it should have to recognise it. If the host was rebooting exactly in the minute when cron should have fired, cron simply did not fire. There was no daemon waiting to pick up the missed tick. The cycle never even started. Six hours of forecasts would be lost, in silence, with nothing in any log to suggest anything had gone wrong.

I lived with this for a long time. Reboots were rare, the impact was limited, and adding state was the kind of thing I always meant to do "next week."

What finally changed it was not a dramatic incident. It was the slow accumulation of small ones. A scheduled VPS reboot. A jail restart after an upgrade. Each one on its own was nothing. Together, they were a steady drip of missed cycles.

So I wrote a daemon.

The crontab entries for the bot went away. There is now a long-running process inside the jail, started at boot, and it does the scheduling itself. The schedule is a list of hours and a minute, read from a JSON config. The daemon wakes up once a minute, checks whether it is time to start a cycle, and either starts one or waits.

The interesting part is the state file.

As the daemon walks through the cities file, it writes its position to a small JSON file: which cities file it is processing, and the index of the next city to handle. The write happens at the boundary between one city and the next, because that is the only place where resuming makes sense. If the daemon is interrupted mid-city, that city is retried on resume; no half-finished post escapes.

When the daemon starts, it reads the state file. If it finds one matching the current cities file, it resumes from the saved index. If the cities file has changed since the state was written, the daemon starts fresh. The check is deliberately conservative: a renamed or modified cities file is treated as a different cycle, because the indices would otherwise be meaningless.

The result is the behavior I should have had from the start. If the host reboots while the United States cycle is running, the daemon comes back up with the jail, reads the state, and continues from where it left off. Every city still gets its update, just with a small gap corresponding to the reboot itself. The cycle finishes. The state file is reset. Life goes on.

And the worst case from the cron days is gone. The daemon does not need anyone to fire it. As long as the jail is running, the daemon is running, and the next scheduled cycle will happen when its hour comes, regardless of what was happening at any specific minute.

Of all the changes I have made to the project, this is the one I like most. It is not exciting work. It is the kind of thing that earns no applause because, when it works, it produces no visible event. But it removes a whole class of small daily annoyances, and it makes a slow process robust against the boring kind of failure: the kind nobody plans for, but that always eventually happens.

The current bot does considerably more than the original Italian script. It handles per-city timezones, three weather providers with automatic fallback, unit conversion for temperature, wind, and precipitation, optional air quality, pressure trend indicators when the provider supplies pressure data, a simplified English text snapshot for future use, a coordinate cache that can be patched by hand, secret redaction at multiple layers, a heartbeat that adapts to whichever HTTP client is installed on the host, and a scheduler-and-resume daemon that survives reboots.

But from the outside, almost nothing has changed.

The European country configs work the same way they always did. The wrapper scripts are unchanged. The snac integration is the same one-line pipe. The HAProxy layer in front does not know or care that the bot was rewritten. The homepage cron that counts followers and regenerates the static page works exactly as before.

The original Italian script does not exist as a file anymore, but it survives as a default. A country config with timezone set to Europe/Rome and no special options behaves, today, exactly as the first version of the bot would have. Everything else is opt-in.

I like this kind of work.

When I wrote about FediMeteo for the first time, I told the story from the beginning: the idea born almost by chance while checking the weather for a holiday, the memory of my grandfather, who for years had been my personal meteorologist, the decision to build something small and useful, and then the surprise of seeing people actually use it. What began as a personal experiment quickly became a small global service, still running with the same philosophy: FreeBSD, jails, simple scripts, snac, text, emoji, and a lot of small pieces doing their work quietly.

That article was mostly about the birth and growth of the project. This one is about one of the less romantic parts of the same story, although I have to admit that I find a certain beauty in it too: keeping the service light as it grows.

FediMeteo is still intentionally simple from the outside. A homepage, some numbers, a list of countries, and many ActivityPub accounts publishing weather forecasts. The posts are text and emoji. There is no JavaScript requirement to read the pages, no heavy frontend, no unnecessary media attached to every forecast, and no dynamic homepage recalculated at every visit just to show the same numbers. This is not accidental. It is the way I wanted the service to behave from the beginning.

But the more the service is used, the more the small details matter. A request that looks harmless when there are ten followers may become a repeated request when there are thousands of followers, remote instances, crawlers, previews, and other servers fetching the same public objects. In the Fediverse, the same small thing can be asked many times by many different places, each one with a perfectly legitimate reason. The backend doesn't care: it just needs to deal with the requests.

And in FediMeteo, the backend is snac.

I like snac very much precisely because it is small, clear, and efficient. It is not a giant application that tries to be everything. It does a focused job and does it well. But this also means that I want to respect its shape. I do not want to waste its threads on work that the reverse proxy can safely do. A snac thread serving the same public avatar again and again is not a tragedy, but it is still a waste. A snac thread answering the same public ActivityPub object several times in the same minute is doing real work, but often not necessary work.

This is the reason behind the HAProxy tuning I am currently using in front of FediMeteo.

It is not about making the configuration look clever. It is about keeping snac quiet.

A continuation of the same idea

I had already explored the same problem with snac and nginx in two previous posts: Improving snac Performance with Nginx Proxy Cache and Caching snac Proxied Media with Nginx. In both cases, the idea was that the reverse proxy should absorb repeated public requests instead of letting them consume snac resources.

This is especially important because snac uses a limited number of threads. I like that. Limits are healthy. They force us to understand what the service is doing, and they prevent a small program from pretending to be an infinite resource. But limits also make waste visible. If a few threads are busy serving files that could have been served from cache, those threads are not available for something more useful.

With FediMeteo the implementation is different because the reverse proxy is HAProxy, but the reasoning is the same. I have many small snac instances, each one in its own FreeBSD (Bastille) jail, and one public entry point that has to route, terminate TLS, compress, cache, and generally remove as much repetitive work as possible from the backends.

This is, in a way, the natural continuation of the original FediMeteo design. In the first article I wrote that I wanted to manage everything according to the Unix philosophy: small pieces working together. This is another piece of that same puzzle. HAProxy does the edge work. snac does the ActivityPub work. Scripts generate forecasts. cron launches updates. ZFS gives me snapshots. FreeBSD jails keep countries separated. Nothing is particularly heroic by itself, but the whole system becomes pleasant because each part has a clear responsibility.

Why there is almost no media

Before talking about HAProxy, it is worth mentioning one of the most important optimizations, which is not in the proxy configuration at all.

FediMeteo does not use media in its forecasts.

No images attached to the posts, no generated weather cards, no maps for each city, no decorative banners. The forecasts are text and emoji. This was a deliberate decision. Weather information does not become more useful just because it is put inside an image, and every media file used by the service would become something to store, serve, cache, federate, expire, back up, and occasionally debug.

Text and emoji are enough. They are accessible, light, readable in text browsers, friendly to timelines, and understandable even when someone does not know the local language perfectly. This was one of the original design principles of FediMeteo, and it also helps the infrastructure. Less media means less work, fewer cache entries, fewer repeated fetches, fewer surprises.

There is one exception: the avatar.

All FediMeteo accounts use the same avatar, and this is also intentional. I could have used a different avatar for each country, or for each city, or created something visually richer. It would have been nicer in some screenshots, perhaps. It would also have been operationally worse.

With one shared avatar, the reverse proxy has one very useful object to cache. It is public, identical for everyone, small, requested often, and therefore almost always hot in cache. HAProxy can serve it directly instead of asking each snac instance to return the same file. Since avatars are requested by remote instances, browsers, profile previews, and all sorts of federation-related fetches, this single decision removes a surprising amount of pointless backend traffic.

So the avatar is not only a visual identity. It is part of the architecture.

This is the kind of optimization I like most, because it starts before the software. It starts with deciding not to create a problem.

The homepage is static because it can be static

The main homepage follows the same logic.

It is a static HTML page generated from a template. Once per hour, a cron script updates the numbers and statistics. It counts the data I want to show, regenerates the page, and then the page remains static until the next run.

This is not because I cannot make a dynamic page. It is because I do not need one. Boring is good.

The homepage does not need to query all the country instances on every visit. It does not need a database request for each user who opens it. It does not need to ask snac anything in real time. The numbers are useful, but they do not need to be updated every second. Once per hour is enough, and it also fits the spirit of the whole project: do the work when it is needed, then serve the result cheaply.

I have seen too many small services become heavy because the first implementation was convenient rather than appropriate. A cron job and a template are not fashionable, but they are often exactly what a page like this needs.

Many countries, one entry point

FediMeteo is made of many country instances. Each one runs in its own jail and listens on its own internal address and port. From the outside, however, they all live under the same domain structure:

fedimeteo.com
www.fedimeteo.com
it.fedimeteo.com
uk.fedimeteo.com
jp.fedimeteo.com
us.fedimeteo.com
usa.fedimeteo.com
can.fedimeteo.com
canada.fedimeteo.com

And many more.

At the beginning, it is always tempting to write one ACL after another in the HAProxy frontend. It is quick, it is explicit, and for five hostnames it is perfectly fine. But FediMeteo did not remain at five hostnames. As countries and aliases grew, a long chain of ACLs would have turned the frontend into a list of names instead of a description of how the proxy behaves.

So I moved the hostname to backend mapping into a map file:

fedimeteo.com        backend_fedimeteo
www.fedimeteo.com    backend_fedimeteo
it.fedimeteo.com     backend_it
uk.fedimeteo.com     backend_uk
jp.fedimeteo.com     backend_jp
us.fedimeteo.com     backend_us
usa.fedimeteo.com    backend_us
can.fedimeteo.com    backend_ca
canada.fedimeteo.com backend_ca

The frontend then needs only one rule:

use_backend %[req.hdr(host),field(1,:),lower,map(/usr/local/etc/fedimeteo.map,backend_fedimeteo)]

This reads the Host header, removes the port if present, lowercases the result, and looks it up in /usr/local/etc/fedimeteo.map. If nothing matches, it falls back to the main FediMeteo backend.

I like this because it keeps the configuration honest. The frontend contains the policy. The map contains the data. Adding a country means adding an entry to the map and defining a backend. I do not need to make the frontend more complicated every time the service grows.

Backends as small compartments

The country backends are deliberately plain:

backend backend_it
    mode http
    http-reuse safe
    server srv1 10.0.0.2:8001 maxconn 30

backend backend_uk
    mode http
    http-reuse safe
    server srv1 10.0.0.7:8001 maxconn 30

backend backend_jp
    mode http
    http-reuse safe
    server srv1 10.0.0.32:8001 maxconn 30

One backend, one jail, one snac instance. This is exactly the same organizational principle as the rest of the project. If I need to reason about Italy, I look at the Italian jail. If I need to reason about the United Kingdom, I look at the UK jail. If one day I need to move a country elsewhere, the separation is already there.

The maxconn 30 value is not a magic number. It is a ceiling. I want each small backend to have a visible limit in front of it. If something starts hammering a country instance, I prefer the pressure to appear at the HAProxy layer instead of becoming unlimited concurrent work inside snac.

http-reuse safe lets HAProxy reuse backend connections where appropriate. This is another small reduction in unnecessary work. Opening connections repeatedly is not the biggest problem in the world, but avoiding it is still better, especially when many small services sit behind the same proxy.

The front door

The HTTPS frontend listens on IPv4 and IPv6 and offers both HTTP/2 and HTTP/1.1:

frontend https_in
    bind :::443 v4v6 ssl crt /usr/local/etc/certs/ alpn h2,http/1.1
    mode http
    option http-keep-alive

TLS defaults are set globally:

ssl-default-bind-ciphersuites TLS_AES_128_GCM_SHA256:TLS_AES_256_GCM_SHA384:TLS_CHACHA20_POLY1305_SHA256
ssl-default-bind-options no-sslv3 no-tlsv10 no-tlsv11 no-tls-tickets

Port 80 only redirects to HTTPS, except for Let's Encrypt challenges:

acl letsencrypt-acl path_beg /.well-known/acme-challenge/
http-request redirect scheme https code 301 unless letsencrypt-acl
use_backend letsencrypt-backend if letsencrypt-acl

In the HTTPS frontend I also set the usual forwarding headers:

http-request set-header X-Real-IP %[src]
http-request set-header X-Forwarded-Proto https

And I add HSTS:

http-response set-header Strict-Transport-Security "max-age=31536000; includeSubDomains; preload"

None of this is unusual, and that is fine. The interesting parts of an infrastructure are not always the parts that should be unusual.

Two caches, because the requests are different

The HAProxy configuration defines two caches:

cache mediacache
  total-max-size 128
  max-object-size 10000000
  max-age 3600
  process-vary on
  max-secondary-entries 12

cache jsoncache
  total-max-size 16
  max-object-size 1000000
  max-age 60
  process-vary on
  max-secondary-entries 12

I keep media and ActivityPub JSON separate because they are not the same kind of traffic.

The media cache is larger and has a longer maximum age. In FediMeteo, this mostly means the shared avatar and a few static-looking objects. Since there is intentionally almost no media, the important cached object is requested very often and remains warm.

The JSON cache is smaller and short-lived. It is there for public ActivityPub GET requests, not to store federation state forever. A 60 second cache is enough to collapse many repeated requests that arrive close together in time, without pretending that ActivityPub responses should be treated like immutable files.

This distinction is important. Caching is not one decision. It is a set of small decisions about what a response means, who can see it, how often it changes, and what happens if it is served again.

Recognizing media

For media, the ACL is based on file extensions:

acl is_media path_end -i .jpg .jpeg .png .gif .webp .svg .ico .mp4 .webm .mp3 .ogg .wav .flac .mov .avi .mkv .m4v

Then I store the result in a transaction variable:

http-request set-var(txn.is_media) bool(true) if is_media

The cache lookup is straightforward:

http-request cache-use mediacache if { var(txn.is_media) -m bool true }

And on the response side:

http-response set-header Cache-Control "max-age=3600, public" if { var(txn.is_media) -m bool true }
http-response del-header Set-Cookie if { var(txn.is_media) -m bool true }
http-response del-header Vary if { var(txn.is_media) -m bool true }
http-response cache-store mediacache if { var(txn.is_media) -m bool true }

The Cache-Control header makes the intent explicit. Set-Cookie is removed because a public media object should not carry session information. Vary is removed because I do not want the same avatar to fragment into many cache entries because of harmless header differences.

This is aggressive only if removed from its context. In this service, with this media policy, it is a reasonable choice. FediMeteo is not serving private media under these paths. It is mostly serving the same public avatar over and over.

For the same reason, I clean the request before it reaches the backend:

http-request del-header Authorization if { var(txn.is_media) -m bool true }
http-request del-header Cookie        if { var(txn.is_media) -m bool true }

I would not do this globally. I do it after deciding that the request is media. Scope is what makes these rules safe.

The result is exactly what I want: the shared avatar becomes an almost perfect cache object. Small, public, repeatedly requested, and served by HAProxy instead of snac.

ActivityPub JSON microcaching

The ActivityPub side starts from the Accept header:

acl is_ap_json   req.hdr(Accept),lower -m sub application/activity+json
acl is_ap_ldjson req.hdr(Accept),lower -m sub application/ld+json
acl is_outbox    path_end /outbox
acl is_get       method GET
acl has_auth     req.hdr(Authorization) -m found
acl has_cookie   req.hdr(Cookie) -m found

This part matters because ActivityPub uses content negotiation. The same path may return HTML to a browser and JSON to a remote instance. If the proxy pretends that a URL is always one thing, it will eventually cache the wrong representation.

So I only mark public ActivityPub GET requests as cacheable:

http-request set-var(txn.is_activitypub) bool(true) if is_get !is_outbox is_ap_json !has_auth !has_cookie
http-request set-var(txn.is_activitypub) bool(true) if is_get !is_outbox is_ap_ldjson !has_auth !has_cookie

There are several decisions here, all important.

It must be a GET, because I am not caching deliveries or anything that changes state. It must not be /outbox, because outbox collections are not the traffic I want to cache here. It must not have Authorization, and it must not have cookies, because authenticated or user-specific requests do not belong in a shared public cache.

Then the cache can be used and populated:

http-request cache-use jsoncache if { var(txn.is_activitypub) -m bool true }

http-response set-header Cache-Control "max-age=60, public" if { var(txn.is_activitypub) -m bool true }
http-response cache-store jsoncache if { var(txn.is_activitypub) -m bool true }

Sixty seconds is short, but useful. Federation often creates small clusters of identical requests. A remote server fetches an actor, another fetches the same actor, something asks for the same object, something retries. I do not need to cache these responses for hours. I only need HAProxy to answer the second and third identical request during the same small burst.

This is microcaching in the most practical sense. It reduces repeated work without changing the nature of the service.

Static media paths

There is also a rule for static paths:

acl is_short_path path_reg ^/[^/]+/s/
http-request cache-use mediacache if is_short_path

This comes from the same observation that led me to cache snac media with nginx. snac uses static media paths, and those paths often represent the kind of public, repeatable traffic that should not consume backend threads if the proxy can serve it. I call them "short", not because they are, but because the first time I saw them, I thought the 's' stood for "short", not "static". The name just stuck.

In FediMeteo this is less central than on a normal social instance, because I deliberately do not use media except for the avatar and basic static objects. Still, the rule fits the general policy: let HAProxy handle repeatable edge work, and let snac spend its threads where they are actually needed.

Vary, but not without limits

Both caches have:

process-vary on
max-secondary-entries 12

I want HAProxy to process Vary, because content negotiation is real, especially when ActivityPub is involved. But I also want variation to be bounded. If every slightly different header creates another cache entry, the cache becomes a complicated way to miss.

For media, I remove Vary before storing the response. A shared avatar does not need to vary by Accept. For ActivityPub JSON, I am more careful because the representation matters.

Again, the important thing is not the number itself. It is the decision to make variation explicit and limited.

Seeing whether it works

During rollout, I like to expose a very small diagnostic header:

http-response set-header X-Cache-Status HIT if !{ srv_id -m found }
http-response set-header X-Cache-Status MISS if { srv_id -m found }

This is intentionally simple. If HAProxy selected a backend server, I call it a miss. If no backend server was selected, the response came from cache, so I call it a hit. It is not a complete observability system, but it is enough to answer the first question I usually have after changing a cache rule.

Did this request reach snac?

A test can be as simple as:

curl -I https://it.fedimeteo.com/path/to/avatar.png
curl -I https://it.fedimeteo.com/path/to/avatar.png

The second request should be a hit.

For ActivityPub JSON, the test must use the right Accept header:

curl -I \
  -H 'Accept: application/activity+json' \
  https://it.fedimeteo.com/some/activitypub/object

And I also want to verify that cookies and authorization prevent public caching:

curl -I \
  -H 'Cookie: test=value' \
  -H 'Accept: application/activity+json' \
  https://it.fedimeteo.com/some/activitypub/object

curl -I \
  -H 'Authorization: Bearer fake' \
  -H 'Accept: application/activity+json' \
  https://it.fedimeteo.com/some/activitypub/object

A cache that works should be visible. A cache that is invisible can be correct, but it can also be silently wrong. I prefer to know.

Compression and operational paths

HAProxy also handles gzip compression:

filter compression
compression algo gzip
compression type text/css text/html text/javascript application/javascript text/plain text/xml application/json application/activity+json

This keeps another common responsibility at the edge. The country instances can stay focused on snac and the forecast data, while HAProxy deals with client-facing compression for HTML, JSON, and ActivityPub responses.

There is also a local Prometheus exporter:

frontend prometheus
  bind 127.0.0.1:8405
  mode http
  http-request use-service prometheus-exporter
  no log

And I keep internal operational paths, such as statistics and Grafana, handled before the hostname map. These are small details, but ordering matters. Special paths should be explicit and early. The hostname map is for FediMeteo routing, not for every internal tool I happen to expose behind the same proxy.

What this changes in practice

The nice thing about this configuration is that none of its parts is particularly surprising.

The map keeps hostname routing manageable. The backend definitions keep each country isolated and limited. The static homepage avoids dynamic work for something that changes once per hour. The shared avatar gives HAProxy one very hot media object to serve directly. The media cache keeps public files away from snac. The JSON microcache absorbs short ActivityPub bursts. Header cleanup prevents useless variation. Connection reuse avoids unnecessary backend connection churn.

But all of this is only a longer way of saying one thing:

fewer requests reach snac.

That is the metric I care about here.

Not because snac is slow. If anything, FediMeteo exists in its current form because snac is efficient enough to make this kind of project possible on a very small VPS. But precisely because the whole architecture is small and pleasant, I do not want to waste resources where there is no need.

This is also consistent with the rest of the project. Forecasts are serialized by scripts. Updates happen every six hours. The homepage is regenerated hourly. Countries live in separate jails. Snapshots and backups are handled outside the application. No single component tries to be the entire system.

HAProxy is just another small piece, but it sits in the right place to remove a lot of repeated work.

Caveats

This configuration is not a universal HAProxy recipe for ActivityPub services.

It matches FediMeteo as it is now: almost no media, one shared avatar, static homepage, public forecasts, many small snac instances, and ActivityPub traffic that can benefit from a short public cache when there are no cookies or authorization headers.

If I decide one day to use media in forecasts, the media cache rules will need to be reviewed. If I use different avatars for each city or country, the cache will still work, but I will lose the very nice property of one shared, always-hot avatar. If ActivityPub responses become actor-dependent, public JSON caching must be reconsidered. If one country grows a very different traffic pattern from the others, it may deserve a different limit or policy.

This is why I do not like presenting configurations as magic. A good configuration is a written form of the assumptions behind a service. When the assumptions change, the configuration must change too.

Conclusion

FediMeteo started as a small idea and became larger than I expected, but I still want it to feel small in the right ways. Small does not mean fragile. Small means understandable. It means that each part has a reason to exist, and that unnecessary work is removed before it becomes a problem.

The HAProxy layer follows this idea. It terminates TLS, routes hostnames through a map, reuses backend connections, serves the shared avatar from cache, microcaches public ActivityPub JSON, avoids authenticated and cookie-based traffic, and gives me a small diagnostic header to see what is happening.

There is no single brilliant directive here. There is only the usual work of matching infrastructure to reality.

FediMeteo publishes weather forecasts as text and emoji. The homepage is static HTML updated every hour. The accounts share the same avatar because it is enough, and because it is better for the cache. Each country has its own snac instance in its own FreeBSD jail. HAProxy stands in front of them and tries, quietly, not to bother them unless it has to.

I like this kind of infrastructure.

Not because it is invisible, but because when it works well, it leaves very little to say.

LibreNMS has been a faithful companion for years now. It quietly handles the monitoring of my servers, devices, and services without demanding much in return - exactly what you want from a tool whose job is to watch over everything else. It's a solid alternative to heavier solutions like Zabbix, and it gives you alerts, data, and graphs on virtually anything reachable over SNMP.

I usually install it on a host that is not reachable from the outside, then let it poll all the devices through a VPN: a single observation point, clean perimeter. The ability to create multiple dashboards - and to filter them by user - has also let me give clients a transparent window onto their own servers. Transparency, in my experience, is always the better long-term bet.

Together with Uptime-Kuma (and the good old Nagios/Munin pair), LibreNMS lives in a FreeBSD jail on my monitoring servers and just does its job.

This post walks through a plain installation of LibreNMS on FreeBSD: package-based, no reverse proxy, no HTTPS, no fancy hardening. The goal is to get to a working setup you can build on top of.

Assumptions

• FreeBSD 15.0-RELEASE, in a jail or on a dedicated VM/host
• nginx + php-fpm + MySQL 8.4
• LibreNMS installed from the official package — not via git clone

One note before we start: in this guide I use plain HTTP just to reach the first-time setup. If your LibreNMS instance won't stay confined to a private network or behind a VPN, configuring HTTPS is mandatory, not optional.

Installation

pkg install librenms mysql84-server python3 nginx

LibreNMS currently depends on PHP 8.4. If you want to speed PHP up, install OPcache too:

pkg install php84-opcache

MySQL

Two settings need to be in place before MySQL starts for the first time. After the first start they cannot be changed without reinitializing the data directory, so it's worth getting them right now.

cd /usr/local/etc/mysql
cp my.cnf.sample my.cnf

In the [mysqld] section, add:

innodb_file_per_table=1
lower_case_table_names=0

Now start MySQL:

service mysql-server enable
service mysql-server start

On a fresh FreeBSD install, the local root user can connect to MySQL without a password from the command line. Connect and create the database and user. I'm using password here as a placeholder - don't.

mysql

CREATE DATABASE librenms CHARACTER SET utf8mb4 COLLATE utf8mb4_unicode_ci;
CREATE USER 'librenms'@'localhost' IDENTIFIED BY 'password';
GRANT ALL PRIVILEGES ON librenms.* TO 'librenms'@'localhost';
exit

php-fpm

Edit /usr/local/etc/php-fpm.d/www.conf and adjust the listen directives:

listen = /var/run/php-fpm-librenms.sock
listen.owner = www
listen.group = www
listen.mode = 0660

Then create php.ini from the production sample:

cd /usr/local/etc
cp php.ini-production php.ini

And set the timezone in php.ini:

date.timezone = Europe/Rome

nginx

Since this jail (or host) is dedicated to LibreNMS, we can rewrite the server block in /usr/local/etc/nginx/nginx.conf directly:

server {
    listen      80;
    #server_name yourServerName
    root        /usr/local/www/librenms/html;
    index       index.php;

    charset utf-8;
    gzip on;
    gzip_types text/css application/javascript text/javascript application/x-javascript image/svg+xml text/plain text/xsd text/xsl text/xml image/x-icon;

    location / {
        try_files $uri $uri/ /index.php?$query_string;
    }

    location /api/v0 {
        try_files $uri $uri/ /api_v0.php?$query_string;
    }

    location ~ \.php$ {
        fastcgi_split_path_info ^(.+\.php)(/.*)$;
        set $path_info $fastcgi_path_info;
        try_files $fastcgi_script_name =404;
        include fastcgi_params;
        fastcgi_param SERVER_SOFTWARE "";
        fastcgi_param SCRIPT_FILENAME $document_root$fastcgi_script_name;
        fastcgi_param PATH_INFO $path_info;
        fastcgi_index index.php;
        fastcgi_pass unix:/var/run/php-fpm-librenms.sock;
        fastcgi_buffers 256 4k;
        fastcgi_intercept_errors on;
        fastcgi_read_timeout 14400;
    }

    location ~ /\.(?!well-known).* {
        deny all;
    }
}

Now start nginx and php-fpm:

service nginx enable
service nginx start

service php_fpm enable
service php_fpm start

LibreNMS configuration

Copy the default config:

cp /usr/local/www/librenms/config.php.default /usr/local/www/librenms/config.php

Because we installed from the package, this file already has the right commands and paths for FreeBSD - no need to hunt down mtr, fping, snmpwalk and friends one by one.

Create the directory for RRD graphs and set ownership:

mkdir -p /var/db/librenms/rrd
chown -R www:www /var/db/librenms
chmod 775 /var/db/librenms/rrd

Then the .env file:

cd /usr/local/www/librenms
cp .env.example .env
chown www .env

Edit .env and set at least:

• DB_DATABASE - librenms
• DB_USERNAME - librenms
• DB_PASSWORD - the one you actually used (not password, please)

Then add this line, which tells LibreNMS we still need to run the web installer:

INSTALL=true

A note on permissions. The official LibreNMS documentation suggests chown -R www:www over the entire application tree, but on FreeBSD the package already lays down sane ownership, with storage/ and bootstrap/cache/ writable by www. There's no reason to widen the rest of the codebase. If validate.php complains later about something write-related, the first place to check is:

ls -la /usr/local/www/librenms/storage /usr/local/www/librenms/bootstrap/cache

Now generate the app key as www, since the file is owned by www:

su -m www -c "php artisan key:generate"

And tighten .env:

chmod 600 .env

Refresh the configuration cache:

su -m www -c "lnms config:clear"
su -m www -c "lnms config:cache"

Web installer

Open http://host/install and follow the steps. The validation process may fail. Refreshing the cache picks up the values written to config.php during the install:

su -m www -c "lnms config:clear"
su -m www -c "lnms config:cache"

When the web installer is done, edit .env again and remove the INSTALL=true line if it's still there. Leaving it in place re-exposes the installer to anyone who can reach the URL.

Polling service

LibreNMS needs something to actually run the polls. On FreeBSD, the package ships an rc service that runs the LibreNMS dispatcher, so there's no need to manage cron entries by hand the way most Linux guides assume.

service librenms enable
service librenms start

Validate

cd /usr/local/www/librenms
su -m www -c './validate.php'

You may see a couple of complaints right after starting the service - usually scheduler-related and self-resolving within a few minutes. Re-run validate.php once the dispatcher has had time to settle. Anything still red after that is worth investigating.

Next steps

At this point you can log into the web interface and start adding devices, configuring SNMP, and building dashboards. For that, the official LibreNMS documentation is excellent, and there's no point in me paraphrasing it here.

Update: This post has been updated to include Docker benchmarks and a comparison of container overhead versus FreeBSD Jails and illumos Zones.

Note: Some operating systems (FreeBSD and Linux) support kernel TLS (kTLS) and the related SSL_sendfile path in nginx, which can improve HTTPS performance for static files. Since this feature is not available on all the systems included in the comparison (for example NetBSD, OpenBSD and illumos), the benchmarks were run with a common baseline configuration that does not rely on kTLS. The goal is to compare the systems under similar conditions rather than to measure OS specific optimizations.

I often get very specific infrastructure requests from clients. Most of the time it is some form of hosting. My job is usually to suggest and implement the setup that fits their goals, skills and long term plans.

If there are competent technicians on the other side, and they are willing to learn or already comfortable with Unix style systems, my first choices are usually one of the BSDs or an illumos distribution. If they need a control panel, or they already have a lot of experience with a particular stack that will clearly help them, I will happily use Linux and it usually delivers solid, reliable results.

Every now and then someone asks the question I like the least:

“But how does it perform compared to X or Y?”

I have never been a big fan of benchmarks. At best they capture a very specific workload on a very specific setup. They are almost never a perfect reflection of what will happen in the real world.

For example, I discovered that idle bhyve VMs seem to use fewer resources when the host is illumos than when the host is FreeBSD. It looks strange at first sight, but the illumos people are clearly working very hard on this, and the result is a very capable and efficient platform.

Despite my skepticism, from time to time I enjoy running some comparative tests. I already did it with Proxmox KVM versus FreeBSD bhyve, and I also compared Jails, Zones, bhyve and KVM on the same Intel N150 box. That led to the FreeBSD vs SmartOS article where I focused on CPU and memory performance on this small mini PC.

This time I wanted to do something simpler, but also closer to what I see every day: static web hosting.

Instead of synthetic CPU or I/O tests, I wanted to measure how different operating systems behave when they serve a small static site with nginx, both over HTTP and HTTPS.

This is not meant to be a super rigorous benchmark. I used the default nginx packages, almost default configuration, and did not tune any OS specific kernel settings. In my experience, careful tuning of kernel and network parameters can easily move numbers by several tens of percentage points. The problem is that very few people actually spend time chasing such optimizations. Much more often, once a limit is reached, someone yells “we need mooooar powaaaar” while the real fix would be to tune the existing stack a bit.

So the question I want to answer here is more modest and more practical:

With default nginx and a small static site, how much does the choice of host OS really matter on this Intel N150 mini PC?

Spoiler: less than people think, at least for plain HTTP. Things get more interesting once TLS enters the picture.

---

Disclaimer
These benchmarks are a snapshot of my specific hardware, network and configuration. They are useful to compare relative behavior on this setup. They are not a universal ranking of operating systems. Different CPUs, NICs, crypto extensions, kernel versions or nginx builds can completely change the picture.

---

Test setup

The hardware is the same Intel N150 mini PC I used in my previous tests: a small, low power box that still has enough cores to be interesting for lab and small production workloads.

On it, I installed several operating systems and environments, always on the bare metal, not nested inside each other. On each OS I installed nginx from the official packages.

Software under test

On the host:

SmartOS, with:
- a Debian 12 LX zone
- an Alpine Linux 3.22 LX zone
- a native SmartOS zone

FreeBSD 14.3-RELEASE:
- nginx running inside a native jail

OpenBSD 7.8:
- nginx on the host

NetBSD 10.1:
- nginx on the host

Debian 13.2:
- nginx on the host

Alpine Linux 3.22:
- nginx on the host
- Docker: Debian 13 container running on the Alpine host (ports mapped)

I also tried to include DragonFlyBSD, but the NIC in this box is not supported. Using a different NIC just for one OS would have made the comparison meaningless, so I excluded it.

nginx configuration

In all environments:

• nginx was installed from the system packages
• worker_processes was set to auto
• the web root contained the same static content

The important part is that I used exactly the same nginx.conf file for all operating systems and all combinations in this article. I copied the same configuration file verbatim to every host, jail and zone. The only changes were the IP address and file paths where needed, for example for the TLS certificate and key.

The static content was a default build of the example site generated by BSSG, my Bash static site generator. The web root was the same logical structure on every OS and container type.

There is no OS specific tuning in the configuration and no kernel level tweaks. This is very close to a “package install plus minimal config” situation.

TLS configuration

For HTTPS I used a very simple configuration, identical on every host.

Self signed certificate created with:

openssl req -x509 -newkey rsa:4096 -nodes -keyout server.key -out server.crt -days 365 -subj "/CN=localhost"  

Example nginx server block for HTTPS (simplified):

server {  
listen 443 ssl http2;  
listen [::]:443 ssl http2;  

server_name _;  

ssl_certificate /etc/nginx/ssl/server.crt;  
ssl_certificate_key /etc/nginx/ssl/server.key;  

root /var/www/html;  
index index.html index.htm;  

location / {  
try_files $uri $uri/ =404;  
}  
}  

The HTTP virtual host is also the same everywhere, with the root pointing to the BSSG example site.

Load generator

The tests were run from my workstation on the same LAN:

• client host: a mini PC machine connected at 2.5 Gbit/s
• switch: 2.5 Gbit/s
• test tool: wrk

For each target host I ran:

• wrk -t4 -c50 -d10s http://IP
• wrk -t4 -c10 -d10s http://IP
• wrk -t4 -c50 -d10s https://IP
• wrk -t4 -c10 -d10s https://IP

Each scenario was executed multiple times to reduce noise; the numbers below are medians (or very close to them) from the runs.

The contenders

To keep things readable, I will refer to each setup as follows:

• SmartOS Debian LX → SmartOS host, Debian 12 LX zone
• SmartOS Alpine LX → SmartOS host, Alpine 3.22 LX zone
• SmartOS Native → SmartOS host, native zone
• FreeBSD Jail → FreeBSD 14.3-RELEASE, nginx in a jail
• OpenBSD Host → OpenBSD 7.8, nginx on the host
• NetBSD Host → NetBSD 10.1, nginx on the host
• Debian Host → Debian 13.2, nginx on the host
• Alpine Host → Alpine 3.22, nginx on the host
• Docker Container → Alpine host, Debian 13 Docker container

Everything uses the same nginx configuration file and the same static site.

Static HTTP results

Let us start with plain HTTP, since this removes TLS from the picture and focuses on the kernel, network stack and nginx itself.

HTTP, 4 threads, 50 concurrent connections

Approximate median wrk results:

Environment	HTTP 50 connections
SmartOS Debian LX	~46.2 k
SmartOS Alpine LX	~49.2 k
SmartOS Native	~63.7 k
FreeBSD Jail	~63.9 k
OpenBSD Host	~64.1 k
NetBSD Host	~64.0 k
Debian Host	~63.8 k
Alpine Host	~63.9 k
Docker Container	~63.7 k

Two things stand out:

1. All the native or jail/container setups on the hosts that are not LX zones cluster around 63 to 64k requests per second.
2. The two SmartOS LX zones sit slightly lower, in the 46 to 49k range, which is still very respectable for this hardware.

In other words, as long as you are on the host or in something very close to it (FreeBSD jail, SmartOS native zone, NetBSD, OpenBSD, Linux on bare metal), static HTTP on nginx will happily max out around 64k requests per second with this small Intel N150 CPU.

The Debian and Alpine LX zones on SmartOS are a bit slower, but not dramatically so. They still deliver close to 50k requests per second and, in a real world scenario, you would probably saturate the network or the client long before hitting those numbers.

HTTP, 4 threads, 10 concurrent connections

With fewer concurrent connections, absolute throughput drops, but the relative picture is similar:

• SmartOS Native around 44k
• NetBSD and Alpine Host around 34 to 35k
• FreeBSD, Debian, OpenBSD around 31 to 33k
• The Docker Container sits slightly lower at ~30.2k req/s, showing a small overhead from the networking layer
• The SmartOS LX zones sit slightly below, around 35 to 37k req/s

The important conclusion is simple:

For plain HTTP static hosting, once nginx is installed and correctly configured, the choice between these operating systems makes very little difference on this hardware. Zones and jails add negligible overhead, LX zones add a small one.

If you are only serving static content over HTTP, your choice of OS should be driven by other factors: ecosystem, tooling, update strategy, your own expertise and preference.

Static HTTPS results

TLS is where things start to diverge more clearly and where CPU utilization becomes interesting.

HTTPS, 4 threads, 50 concurrent connections

Approximate medians:

Environment	HTTPS 50 connections	CPU notes at 50 HTTPS connections
SmartOS Debian LX	~51.4 k	CPU saturated
SmartOS Alpine LX	~40.4 k	CPU saturated
SmartOS Native	~52.8 k	CPU saturated
FreeBSD Jail	~62.9 k	around 60% CPU idle
OpenBSD Host	~39.7 k	CPU saturated
NetBSD Host	~40.4 k	CPU saturated
Debian Host	~62.8 k	about 20% CPU idle
Alpine Host	~62.4 k	small idle headroom, around 7% idle
Docker Container	~62.7 k	CPU saturated

These numbers tell a more nuanced story.

1. FreeBSD, Debian and Alpine on bare metal form a “fast TLS” group.
   All three sit around 62 to 63k requests per second with 50 concurrent HTTPS connections.

2. FreeBSD does this while using significantly less CPU.
   During the HTTPS tests with 50 connections, the FreeBSD host still had around 60% CPU idle. It is the platform that handled TLS load most comfortably in terms of CPU headroom.

3. Debian and Alpine are close in throughput, but push the CPU harder.
   Debian still had some idle time left, Alpine even less. In practice, all three are excellent here, but FreeBSD gives you more room before you hit the wall.

4. SmartOS, NetBSD and OpenBSD form a “good but heavier” TLS group.
   Their HTTPS throughput is in the 40 to 52k req/s range and they reach full CPU usage at 50 concurrent connections. OpenBSD and NetBSD stabilize around 39 to 40k req/s. SmartOS native and the Debian LX zone manage slightly better (around 51 to 53k) but still with the CPU pegged.

HTTPS, 4 threads, 10 concurrent connections

With lower concurrency:

• FreeBSD, Debian and Alpine still sit in roughly the 29 to 31k req/s range
• SmartOS Native and LX zones are in the mid to high 30k range
• The Docker Container drops slightly to ~27.8k req/s
• NetBSD and OpenBSD sit around 26 to 27k req/s

The relative pattern is the same: for this TLS workload, FreeBSD and modern Linux distributions on bare metal appear to make better use of the cryptographic capabilities of the CPU, delivering higher throughput or more headroom or both.

What TLS seems to highlight

The HTTPS tests point to something that is not about nginx itself, but about the TLS stack and how well it can exploit the hardware.

On this Intel N150, my feeling is:

• FreeBSD, with the userland and crypto stack I am running, is very efficient at TLS here. It delivers the highest throughput while keeping plenty of CPU in reserve.
• Debian and Alpine, with their recent kernels and libraries, are also strong performers, close to FreeBSD in throughput, but with less idle CPU.
• NetBSD, OpenBSD and SmartOS (native and LX) are still perfectly capable of serving a lot of HTTPS traffic, but they have to work harder to keep up and they hit 100% CPU much earlier.

This matches what I see in day to day operations: TLS performance is often less about “nginx vs something else” and more about the combination of:

• the TLS library version and configuration
• how well the OS uses the CPU crypto instructions
• kernel level details in the network and crypto paths

I suspect the differences here are mostly due to how each system combines its TLS stack (OpenSSL, LibreSSL and friends), its kernel and its hardware acceleration support. It would take a deeper dive into profiling and configuration knobs to attribute the gaps precisely.

In any case, on this specific mini PC, if I had to pick a platform to handle a large amount of HTTPS static traffic, FreeBSD, Debian and Alpine would be my first candidates, in that order.

Zones, jails, containers and Docker: overhead in practice

Another interesting part of the story is the overhead introduced by different isolation technologies.

From these tests and the previous virtualization article on the same N150 machine, the picture is consistent:

• FreeBSD jails behave almost like bare metal and are significantly more efficient than Docker.
  For both HTTP and HTTPS, running nginx in a jail on FreeBSD 14.3-RELEASE produces numbers practically identical to native hosts.
  The contrast with Docker is striking: while the Docker container required 100% CPU to reach peak for the HTTP and HTTPS throughput, the FreeBSD jail delivered the same speed with ~60% of the CPU sitting idle. In terms of performance cost per request, Jails are drastically cheaper.

• SmartOS native zones are also very close to the metal.
  Static HTTP performance reaches the same 64k req/s region and HTTPS is only slightly behind the "fast TLS" group, although with higher CPU usage.

• SmartOS LX zones introduce a noticeable but modest overhead.
  Both Debian and Alpine LX zones on SmartOS perform slightly worse than the native zone or FreeBSD jails. For static HTTP they are still very fast. For HTTPS the Debian LX zone remains competitive but costs more CPU, while the Alpine LX zone is slower.

• Docker on Linux performs efficiently but eats the margins. I ran an additional test using a Debian 13 Docker container running on the Alpine Linux host. At peak load (50 connections), the throughput was impressive and virtually identical to bare metal: ~63.7k req/s for HTTP and ~62.7k req/s for HTTPS. However, there is a clear cost. First, while the bare metal host maintained a small CPU buffer (~7% idle) during the HTTPS test, Docker saturated the CPU to 100%. Second, at lower concurrency (10 connections), the overhead became visible. The Docker container scored ~30.2k req/s for HTTP and ~27.8k req/s for HTTPS, slightly trailing the ~31-34k and ~29-31k range of the bare metal counterparts. The abstraction layers (NAT, bridging, namespaces) are extremely efficient, but they are not completely free.

This leads to a clear conclusion on efficiency: FreeBSD Jails provide the highest throughput with the lowest CPU cost. LX zones and Docker containers can match the speed (or come close), but they burn significantly more CPU cycles to do so.

What this means for real workloads

It is easy to get lost in tables and percentages, so let us go back to the initial question.

A client wants static hosting.
Does the choice between FreeBSD, SmartOS, NetBSD or Linux matter in terms of performance?

For plain HTTP on this hardware, with nginx and the same configuration:

• Not really.
  All the native hosts and FreeBSD jails deliver roughly the same maximum throughput, in the 63 to 64k req/s range. SmartOS LX zones are slightly slower but still strong.

For HTTPS:

• Yes, it starts to matter a bit more.
• FreeBSD stands out for how relaxed the CPU is under high TLS load.
• Debian and Alpine are very close in throughput, with more CPU used but still with some headroom.
• SmartOS, NetBSD and OpenBSD can still push a lot of HTTPS traffic, but they reach 100% CPU earlier and stabilize at lower request rates.

Does this mean you should always choose FreeBSD or Debian or Alpine for static HTTPS hosting?

Not necessarily.

In real deployments, the bottleneck is rarely the TLS performance of a single node serving a small static site. Network throughput, storage, logging, reverse proxies, CDNs and application layers all play a role.

However, knowing that FreeBSD and current Linux distributions can squeeze more out of a small CPU under TLS is useful when you are:

• sizing hardware for small VPS nodes that must serve many HTTPS requests
• planning to consolidate multiple services on a low power box
• deciding whether you can afford to keep some CPU aside for other tasks (cache, background jobs, monitoring, and so on)

As always, the right answer depends on the complete picture: your skills, your tooling, your backups, your monitoring, the rest of your stack, and your tolerance for troubleshooting when things go sideways.

Final thoughts

From these small tests, my main takeaways are:

1. Static HTTP is basically solved on all these platforms.
   On a modest Intel N150, every system tested can push around 64k static HTTP requests per second with nginx set to almost default settings. For many use cases, that is already more than enough.

2. TLS performance is where the OS and crypto stack start to matter.
   FreeBSD, Debian and Alpine squeeze more HTTPS requests out of the N150, and FreeBSD in particular does it with a surprising amount of idle CPU left. NetBSD, OpenBSD and SmartOS need more CPU to reach similar speeds and stabilize at lower throughput once the CPU is saturated.

3. Jails and native zones are essentially free, LX zones cost a bit more.
   FreeBSD jails and SmartOS native zones show very little overhead for this workload. SmartOS LX zones are still perfectly usable, but if you are chasing every last request per second you will see the cost of the translation layer.

4. Benchmarks are only part of the story.
   If your team knows OpenBSD inside out and has tooling, scripts and workflows built around it, you might happily accept using more CPU on TLS in exchange for security features, simplicity and familiarity. The same goes for NetBSD or SmartOS in environments where their specific strengths shine.

I will not choose an operating system for a client just because a benchmark looks nicer. These numbers are one of the many inputs I consider. What matters most is always the combination of reliability, security, maintainability and the human beings who will have to operate the
system at three in the morning when something goes wrong.

Still, it is nice to know that if you put a tiny Intel N150 in front of a static site and you pick FreeBSD or a modern Linux distribution for HTTPS, you are giving that little CPU a fair chance to shine.

Mastodon 4.5.0 is here, and with it come some interesting changes that, in my opinion, might encourage more people to consider it for self-hosting their Fediverse community.

While it may not be as lightweight and simple as other solutions (like snac or GoToSocial or many others), I believe it remains one of the best platforms for managing a medium-sized Fediverse community, thanks in part to the direct feedback that many admins have provided to the developers.

I have previously written about how to install Mastodon in a FreeBSD jail and how to modify its character and poll limits.

One of the most critical initial decisions (which can be changed later, but with extra work) is where to store your media files. Mastodon downloads and re-processes all media it encounters from other instances for three main reasons:

• Local Caching: Your users connect to your media server, reducing the load on the original instance.
• Security: Re-processing media helps to remove any potential "impurities" before they reach the user's device.
• Privacy: It prevents disclosing your users' IP addresses to other instances. A user will only connect to their own instance to fetch all data, including remote content.

At least initially, media files will be the largest part of your instance's storage footprint. It is therefore essential to plan where to store them and to add a regular cleanup script; otherwise, their growth will be exponential.

Mastodon supports uploading media to external S3-compatible solutions, and many admins use the usual commercial providers, paying for data uploads and transfers.

I am a firm believer in "Own Your Data", so I have always used my own self-hosted S3 servers. I initially started with Minio, but over time, I realized that, by design, it doesn't perform well with a multitude of small files (performance degrades). After running some tests, I decided to switch to SeaweedFS.

SeaweedFS "is a fast distributed storage system for blobs, objects, files, and data lake, for billions of files! Blob store has O(1) disk seek..." - this, combined with the fact that it is a mature and proven piece of software, was enough for me to give it a try. The result? Excellent. The I/O and CPU load on my media server dropped drastically, making SeaweedFS an incredibly suitable solution. Furthermore, some of its features (like the ability to run a filer.sync) allow for efficient and fast replication to other storage, another host, or... anything else.

SeaweedFS works perfectly with Mastodon, and I will explain the steps to get it into production.

I will install SeaweedFS in a dedicated jail and use a dedicated subdomain. This ensures that the media server can be moved to another host at any time without reconfiguring everything or changing domains. SeaweedFS has its own FreeBSD package, installable via pkg, or can be downloaded directly from the project's website.

In either case, I will describe a "test" setup - which can also be used in production without issues. However, I highly recommend diving deeper into the tool, as it is incredibly powerful and flexible and can solve many more problems than one might imagine.

Setting up the SeaweedFS Jail

First, let's create a dedicated jail with BastilleBSD:

bastille create media 14.3-RELEASE 10.0.0.66 bastille0

Now, let's enter the jail and install SeaweedFS (and tmux, which can be useful):

bastille console media
pkg install -y tmux seaweedfs

I suggest launching SeaweedFS in a tmux session so you can monitor its output. Later, you should configure an automatic startup method, such as using the included rc.d file or any other method you prefer.

Create a directory for the data and start SeaweedFS as the "seaweedfs" user:

mkdir -p /seaweedfs/data
chown -R seaweedfs /seaweedfs
su -m seaweedfs
cd /seaweedfs/
/usr/local/bin/weed server -dir /seaweedfs/data -s3

At this point, SeaweedFS will start and create everything it needs to function, including the S3 server.

Configuring Buckets and Users

Now, let's open the weed shell to create the necessary bucket and users:

weed shell
s3.bucket.create -name mastomedia

Still in the weed shell, create a user for Mastodon and grant read permissions for unauthenticated users (which is necessary to serve media to the world):

s3.configure -access_key=mastomedia -secret_key=CHANGEME -buckets=mastomedia -user=mastodon -actions=Read,Write,List,Tagging,Admin -apply
s3.configure -buckets=mastomedia -user=anonymous -actions=Read -apply
s3.configure -buckets=mastomedia -actions=Read -apply

Security Tip: For the -secret_key, avoid using a simple password. You can generate a strong, random key directly from your shell with a command like openssl rand -base64 32.

Done. SeaweedFS is now ready to receive (and serve) media. The next step is to set up a reverse proxy to serve everything over HTTPS. My preferred approach is to configure the system as if it were external, even if the services are in adjacent jails. This might use slightly more resources, but the time and trouble it saves in the future are well worth it.

Nginx Reverse Proxy Configuration

The reverse proxy can be configured something like this:

[...]

server {
   server_name  media.mastodon.example.com;

   ignore_invalid_headers off;
   client_max_body_size 0; # Allow large file uploads without Nginx limits

   location / {
      proxy_set_header Host $http_host;
      proxy_set_header X-Real-IP $remote_addr;
      proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
      proxy_set_header X-Forwarded-Proto $scheme;

      proxy_connect_timeout 300;
      proxy_http_version 1.1;
      proxy_set_header Connection "";
      chunked_transfer_encoding off;

      expires 1y;
      add_header Cache-Control public;

      add_header X-Cache-Status $upstream_cache_status;
      add_header X-Content-Type-Options nosniff;

      proxy_pass http://10.0.0.66:8333;
   }

# ... other server configurations like SSL ...

}

Mastodon Configuration

Now let's configure Mastodon. If you are running the setup wizard for the first time, here is a summary of the options:

[...]
Do you want to store uploaded files on the cloud? yes
Provider Minio
Minio endpoint URL: https://media.mastodon.example.com
Minio bucket name: mastomedia
Minio access key: mastomedia
Minio secret key: CHANGEME
Do you want to access the uploaded files from your own domain? Yes
Domain for uploaded files: media.mastodon.example.com

If Mastodon is already active, or once the setup is complete, the options in your .env.prod file should be modified to be consistent with what SeaweedFS expects:

S3_ENABLED=true
S3_PROTOCOL=https
S3_REGION=us-east-1
S3_ENDPOINT=https://media.mastodon.example.com
S3_HOSTNAME=media.mastodon.example.com
S3_BUCKET=mastomedia
AWS_ACCESS_KEY_ID=mastomedia
AWS_SECRET_ACCESS_KEY=CHANGEME
S3_FORCE_SINGLE_REQUEST=true
# remove the S3_ALIAS_HOST if it is set

IMPORTANT NOTE: If both services are in jails on the same host (i.e., SeaweedFS is on the same host as Mastodon), you should ensure that the Mastodon jail can reach the SeaweedFS jail through the reverse proxy and not via the external IP. To do this, add the following line to the /etc/hosts file of the Mastodon jail:

10.0.0.1        media.mastodon.example.com

In this example, the reverse proxy is at 10.0.0.1. If you are not using a separate reverse proxy but are exposing Nginx directly from the jail (as described in my Mastodon installation article), use the IP of the Mastodon jail itself instead (e.g., 10.0.0.42).

With this setup, Mastodon will be able to upload media to the SeaweedFS server and generate the correct links for other instances, public visitors, and users of your own instance.

Have fun with SeaweedFS!

Disclaimer
These benchmarks were performed on my specific hardware and tuned for the workloads I expect to run.
They should not be taken as absolute or universally applicable results.
Different CPUs, storage, networking setups, or workload profiles could produce very different outcomes.
What I’m sharing here is a faithful snapshot of my test environment and use case - a guidepost, not a final verdict.

Years ago, I installed a PCEngines APU at a client's site. It dutifully ran Proxmox with a few small VMs inside. It wasn't a speed demon, but it got the job done. Tasked with running in a closed, uncooled, and unsupervised server closet, it soldiered on for about seven years.

Then, while I was at BSDCan, I got the call. A series of power outages and surges had finally taken their toll, and the APU was dead. It was probably just the power supply, but given its age, we decided it was time for a replacement. I set up a remote bypass to keep them running, but I knew I'd need to install something more powerful soon.

I ordered a modern MiniPC-based on the low-power Intel Processor N150 platform, but with 16GB of RAM and more than enough performance to serve as a decent workstation. I have a similar one in my office running openSUSE Tumbleweed, and it works beautifully.

This time, however, I decided to replace Proxmox with a different virtualization system. This decision wasn't made in a vacuum. In the past, I've put bhyve head-to-head with Proxmox, and my findings were clear: bhyve on FreeBSD is an extremely efficient hypervisor, often outperforming KVM on Proxmox in my tests.

This positive experience is what made FreeBSD with bhyve a top contender. The other path was a KVM-style approach (which would require fewer changes to the VMs), where my options would be NetBSD or an illumos-based OS like SmartOS. Since I had the new hardware on hand, I decided to run some tests to see how these different technologies stacked up against each other, and against the bare metal itself.

The Lineup: What I Put on the Test Bench

My goal was to test every reasonable option on this Intel N150 hardware. The final lineup covered the entire spectrum:

• The Baseline:
  • FreeBSD 14.3-RELEASE Bare Metal: The ground truth for performance on this hardware.
• OS-Level Virtualization (Containers):
  • SmartOS Native Zone: The baseline native container on SmartOS.
  • SmartOS LX Zone: Running Ubuntu 24.04 and Alpine Linux.
  • FreeBSD Native Jail: The baseline native container on FreeBSD.
  • FreeBSD Jail with Linux: A jail running a Ubuntu 22.04 userland.
• Full Hardware Virtualization (HVM):
  • SmartOS bhyve Zone: A FreeBSD guest inside the bhyve hypervisor on a SmartOS host.
  • SmartOS KVM Zone: A FreeBSD guest inside the KVM hypervisor on a SmartOS host.
  • FreeBSD bhyve VM: A FreeBSD guest inside the bhyve hypervisor on a FreeBSD host.

The Benchmark: My sysbench Commands

To keep the comparison fair and simple, I used two core sysbench commands. To ensure consistency, I even compiled sysbench from scratch on the SmartOS native zone to match the versions and compile options on the other systems as closely as possible.

The commands I used in each environment were:

• For CPU performance: sysbench --test=cpu --cpu-max-prime=20000 run
• For memory performance: sysbench --test=memory run

First Look: CPU and Memory on the Intel N150

My initial tests on the Intel N150 hardware immediately revealed some interesting trends. The sysbench CPU results from any native FreeBSD environment (bare metal or jail) were on a completely different scale from the Linux and SmartOS guests, making a direct comparison meaningless.

However, by excluding the incompatible FreeBSD-native results, we get a very clear picture of the overhead between the various container technologies.

Valid CPU Performance Comparison (Single Thread, Intel N150)

Host OS	Container Tech	Guest OS	CPU Performance (Events/sec)
FreeBSD	Jail (OS-level)	Ubuntu 22.04	1108.18
SmartOS	LX Zone (OS-level)	Ubuntu 24.04	1107.13
SmartOS	Native Zone (OS-level)	SmartOS	1107.04
SmartOS	LX Zone (OS-level)	Alpine Linux	1022.81

The takeaway here was clear: for CPU work, the overhead from these containers is basically a rounding error. For CPU-bound tasks, neither SmartOS Zones nor FreeBSD Jails will be a bottleneck.

The memory results, which were consistent across all platforms, were far more revealing.

Overall Memory Performance Comparison (Intel Processor N150)

Host OS	Virtualization Tech	Guest OS	Memory Performance (Transfer Rate)
SmartOS	LX Zone (OS-level)	Ubuntu 24.04	4970.54 MiB/sec
SmartOS	Native Zone (OS-level)	SmartOS (Native)	4549.97 MiB/sec
FreeBSD	Jail (OS-level)	Ubuntu 22.04	4348.32 MiB/sec
FreeBSD	Bare Metal	FreeBSD (Native)	4005.08 MiB/sec
FreeBSD	Native Jail (OS-level)	FreeBSD (Native)	3990.13 MiB/sec
SmartOS	LX Zone (OS-level)	Alpine Linux	3803.72 MiB/sec
FreeBSD	bhyve VM (Full HVM)	FreeBSD	3636.01 MiB/sec
SmartOS	bhyve Zone (Full HVM)	FreeBSD	3020.15 MiB/sec
SmartOS	KVM Zone (Full HVM)	FreeBSD	205.18 MiB/sec

These initial numbers led to a few conclusions: a virtual layer could be a performance boost, the userland matters, and bhyve clearly outclassed the legacy KVM on SmartOS. However, one result was nagging at me: the performance gap between FreeBSD bare metal (4005.08 MiB/sec) and a native bhyve VM (3636.01 MiB/sec) was about 9%. This was a larger drop than I expected. It prompted a new question: was this overhead inherent to bhyve, or was it a quirk of the new N150 hardware?

Going deeper: Testing on an Intel i7-7500U

To see if more mature, better-supported hardware would tell a different story, I replicated the FreeBSD tests on an older Qotom Mini-PC powered by an Intel i7-7500U. The results were illuminating and dramatically changed the narrative.

CPU Performance Comparison (Intel i7-7500U)

Once again, the CPU tests produced strange results. The native FreeBSD environments all reported incredibly high numbers in the millions of events/sec, while the Ubuntu Linuxulator jail's result was on a completely different, incompatible scale. Frankly, given the massive discrepancy between FreeBSD-native and Linux-based environments, I'm unsure that the sysbench CPU figures can be considered totally reliable in absolute terms.

However, what is useful is comparing the native FreeBSD results against each other. This tells us about relative overhead.

Platform	CPU Performance (Events/sec)	Overhead vs. Bare Metal
FreeBSD Bare Metal	6,377,778	Baseline
FreeBSD Native Jail	6,379,271	~0.0%
FreeBSD bhyve VM	6,346,852	-0.48%

Even if we're skeptical of the absolute numbers, the relative comparison is crystal clear: the CPU overhead of bhyve is less than half a percent. This is the key takeaway.

Memory Performance Comparison (Intel i7-7500U)

The memory benchmarks, in contrast, were consistent and highly informative.

Platform	Memory Performance (Transfer Rate)	Overhead vs. Bare Metal
Ubuntu 22.04 Jail	4856.23 MiB/sec	+7.55%
FreeBSD Native Jail	4517.73 MiB/sec	+0.05%
FreeBSD Bare Metal	4515.24 MiB/sec	Baseline
FreeBSD bhyve VM	4491.60 MiB/sec	-0.52%

This is where the real story is. The memory performance of a bhyve VM was a mere 0.52% slower than bare metal. This is the kind of near-native performance one hopes for from a top-tier hypervisor and stands in stark contrast to the 9% drop seen on the newer N150.

Breaking Down the Results: What I Learned From Both Tests

This comprehensive two-platform analysis paints a much clearer picture.

1. Hardware Really Matters Performance is not an absolute. The difference between the two platforms was stark: on the mature i7-7500U, bhyve’s overhead was less than 1%, while on the newer, budget N150, it was a more significant 9%. This suggests the performance dip is likely due to missing optimizations for that specific CPU architecture, rather than a fundamental flaw in bhyve itself.

2. bhyve's True Potential is Near-Native Speed The i7 tests prove that bhyve is an exceptionally efficient hypervisor on well-supported hardware. The relative CPU overhead was a negligible -0.48%, and more importantly, the reliable memory benchmarks showed a performance drop of just 0.52% compared to bare metal. This is the gold standard for virtualization.

3. FreeBSD Jails are Feather-Light On both platforms, native FreeBSD jails demonstrated almost zero performance overhead. On the i7, both CPU and memory performance were virtually identical to bare metal (a 0.05% difference). The N150 CPU tests further showed that FreeBSD's container implementation is so efficient that running a Linux userland inside a jail delivered the best CPU scores of the entire lineup.

4. SmartOS Zones Are Also Extremely Efficient Just like Jails, SmartOS's native Zones proved to be remarkably lightweight. The N150 CPU tests confirm this, showing that native and LX zones have virtually identical, top-tier performance. On the memory front, the native Zone delivered performance over 13% faster than the FreeBSD bare-metal baseline, pointing to the high efficiency of the illumos kernel.

5. The Linux Userland Excels at Throughput A clear pattern emerged on both testbeds: the Ubuntu userland consistently delivered excellent benchmark results. On the CPU front, Ubuntu on both FreeBSD and SmartOS delivered the highest, and nearly identical, performance scores on the N150. For memory, the story was even more dramatic: the Ubuntu LX Zone on SmartOS was the top performer, beating bare-metal FreeBSD by nearly 25%, while the Ubuntu jail on the i7 also surpassed its host by over 7%.

Final Thoughts: The Verdict for My Client's New Server

So, what's the bottom line for my client's new MiniPC? This benchmarking journey has made the path forward much clearer.

At the beginning of this process, my main question was whether to stick with a KVM-based setup or make the switch to bhyve. The performance data answers that decisively. The legacy KVM on SmartOS showed a crippling performance penalty, making it a non-starter. Given that, the extra effort to migrate the existing VMs to a bhyve-compatible format is absolutely worth it. The performance gain is just too significant to ignore.

The final question, then, is which host OS to use for bhyve: SmartOS or FreeBSD? This is a much tougher call, as both platforms demonstrated incredible strengths.

SmartOS, powered by the illumos kernel, was a true surprise. It delivered astonishing performance on the target N150 hardware. Its key advantage is the raw speed of its containerization for both CPU and memory tasks. The Ubuntu LX Zone not only ran flawlessly but delivered top-tier CPU scores and outperformed the bare-metal FreeBSD baseline in memory by a massive 25% margin. This points to a highly efficient kernel and offers the tantalizing prospect of running ultra-fast Linux containers alongside performant bhyve VMs on the same host.

On the other hand, FreeBSD proved its mastery of bhyve virtualization. The tests on the i7 hardware showed its implementation to be the gold standard, offering virtually zero performance overhead for full hardware virtualization. Its native Jails are equally efficient, and its Linux compatibility layer is so effective that an Ubuntu jail delivered the fastest CPU performance of all containers tested on FreeBSD. For workloads that must live in a full VM, FreeBSD offers the most performant and native bhyve experience, with the reasonable expectation that its support for newer hardware like the N150 will only improve over time.

Ultimately, the choice comes down to the primary workload. It's a decision between the raw container speed and Linux flexibility of SmartOS versus the pure, uncompromising HVM performance of FreeBSD.

But one thing is certain: thanks to this deep dive, the path forward is much clearer, and it's paved by bhyve.

With the primary backup strategies and methodologies introduced, we've reached the point where we can get specific: the Backup Server configuration.

When choosing the type of backup server to use, I tend to favor specific setups: either I trust a professional backup service provider (like Colin Percival's Tarsnap), or I want full control over the disks where the backups will be hosted. In both cases, for the past twenty years, my operating system of choice for backup servers has been FreeBSD. With a few rare exceptions for clients with special requests, it covers all my needs. When I require Linux-based solutions, such as the Proxmox Backup Server, I create a VM and manage it within.

I typically use both IPv4 and IPv6. For IPv4, I "play" with NAT and port forwarding. For IPv6, I tend to assign a public IPv6 address to each jail or VM, which is then filtered by the physical server's firewall. Unfortunately, every provider, server, and setup has a different approach to IPv6, making it impossible to cover them all in this article. When a provider allows for routed setups, I use this approach: Make your own VPN: FreeBSD, WireGuard, IPv6, and ad-blocking included - assigning a /72 to the bridge for the jails and VMs.

In my opinion, FreeBSD is a perfect all-rounder for backups, thanks to its ability to completely partition services. You can separate backup services (or specific servers/clients) into different jails or even VMs. Furthermore, using ZFS greatly enhances both flexibility and the range of tools you can use.

The main distinction is usually between local backup servers (physically accessible, though not always attended, and in locations deemed secure) and remote ones, such as leased external servers. I personally use a combination of both. If the services I need to back up are external, in a datacenter, and need to be quickly restorable, I prefer to always have a copy on another server in a different datacenter with good outbound connectivity. This guarantees good bandwidth for restores, which isn't always available from a local connection to the outside world. However, an internal, nearby, and accessible backup server (even a Raspberry Pi or a mini PC) ensures physical access to the data. Whenever possible, I maintain both an external and an internal copy - and they are autonomous, meaning the internal copy is not a replica of the external one, but an additional, independent backup. This ensures that if a problem occurs with the external backup, it won't automatically propagate to the internal one. In any case, the backup must always be in a different datacenter from the one containing the production data. When the fire at the OVH datacenter in Strasbourg caused the entire complex to shut down, many people found themselves in trouble because their backups were in the same, now unreachable, location. I had a copy with another provider, in a different datacenter and country, as well as a local copy.

Despite it being "just" a backup server, I almost always use some form of disk redundancy. If I have two disks, I set up a mirror. With three or more, I use RaidZ1 or RaidZ2. This is because, in my view, backups are nearly as important as production data. The inability to recover data from a backup means it's lost forever. And it happens often, very often, that someone contacts me to recover a file (or a database, etc.) days or weeks after its accidental loss or deletion. Usually, pulling out a file from a two-month-old backup generates a mix of disbelief, admiration, but above all, a sense of security in the person requesting it. And that is what our work should instill in the people we collaborate with.

The backup server should be hardened. If possible, it should be protected and unreachable from the outside. My best backup servers are those accessible only via VPN, capable of pulling the data on their own. If they are on a LAN, it's even better if they are completely disconnected from the Internet.

For this very reason, backups must always be encrypted. Having a backup means having full access to the data, and the backup server is the prime target for being breached or stolen if the goal is to get your hands on that data. I've seen healthcare facilities' backup servers being targeted (in a rather trivial way, to be honest) by journalists looking for health details of important figures. It is therefore critical that the backup server be as secure as possible.

Based on the type of access, I use two types of encryption:

• If the server is local (especially if the ZFS pool is on external disks), I usually install FreeBSD on UFS in read-only mode, as I've described in a previous article, and encrypt the backup disks with GELI. This ensures that in the event of a "dirty" shutdown (more likely in unattended environments), I can reconnect to the host and then reactivate the ZFS pool. This approach makes it nearly impossible to retrieve even the pool's metadata if the disks are stolen, as GELI performs a full-device encryption. For example, an employee of a company I work with stole one of the secondary backup disks (which was located at a different, unmonitored company site) to steal information. He got nothing but a criminal complaint. With this approach, it's also not necessary to further encrypt the datasets, which avoids some issues (which I'll discuss later, in a future post).
• If the server is remote, in a datacenter, I usually use ZFS native encryption, encrypting the main backup dataset (and BastilleBSD's, if applicable). Consequently, all child datasets containing backups will also be encrypted. In this case as well, a password will be required after a reboot to unlock those datasets, ensuring that the data cannot be extracted if control of the disks is lost.

Here is an example of how to use GELI to encrypt an entire partition and then create a ZFS pool on it (in the example, the disk is da1 - do not follow these commands blindly, or you will erase all content on the da1 device!):

# WARNING: This destroys the existing partition table on disk da1
gpart destroy -F da1

# Create a new GPT partition table
gpart create -s gpt da1

# Add a freebsd-zfs partition that spans the entire disk
# The -a 1m flag ensures proper alignment
gpart add -t freebsd-zfs -a 1m da1

# Initialize GELI encryption on the new partition (da1p1)
# We use AES-XTS with 256-bit keys and a 4k sector size
# The -b flag means "boot," prompting for the passphrase at boot time
geli init -b -l 256 -s 4096 da1p1
# You will be prompted for a passphrase: choose a strong one and save it!

# Attach the encrypted partition. A new device /dev/da1p1.eli will be created.
# You will be prompted for the passphrase you just set
geli attach da1p1

# (Optional) Check the status of the encrypted device
geli status da1p1

# Create the ZFS pool "bckpool" on the encrypted device
# We enable zstd compression (an excellent compromise) and disable atime
zpool create -O compression=zstd -O atime=off bckpool da1p1.eli

In this setup, the reference pool for everything related to backups will be bckpool - and you'll need to keep this in mind for the next steps. Additionally, after every server reboot, you'll need to "unlock" the disk and import the pool:

# Enter the passphrase when prompted
geli attach da1p1

# Import the ZFS pool, which is now visible
zpool import bckpool

With this method, it's not necessary to encrypt the ZFS datasets, as the underlying disk (or, more precisely, the partition containing the ZFS pool) is already encrypted.

If, instead, you choose to encrypt the ZFS dataset (for example, if you install FreeBSD on the same disks that will hold the data and don't want to use a multi-partition approach), you should create a base encrypted dataset. Inside it, you can create the various backup datasets, VMs, and the BastilleBSD mountpoint. Due to property inheritance, they will all be encrypted as well.

To create an encrypted dataset, a command like this will suffice:

# Creates a new dataset with encryption enabled.
# keylocation=prompt will ask for a passphrase every time it's mounted.
# keyformat=passphrase specifies the key type.
zfs create -o encryption=on -o keylocation=prompt -o keyformat=passphrase zfspool/dataset

In this case, after every reboot, you will need to load the key and mount the dataset:

zfs load-key zfspool/dataset
zfs mount zfspool/dataset

Keep in mind the setup you choose, as many of the subsequent choices and commands will depend on it.

Base System Setup

I'll install BastilleBSD - a useful tool for separating services into jails. It will be helpful for isolating our backup services:

pkg install -y bastille

If you used ZFS for the root filesystem, you can proceed directly with the setup. Otherwise (i.e., ZFS on other disks), you'll need to edit the /usr/local/etc/bastille/bastille.conf file and specify the correct dataset on which to install the jails. Then run:

bastille setup

Once the automatic setup is complete, check the /etc/pf.conf file - it will be automatically configured to only accept SSH connections. Ensure the network interface is set correctly. When you activate pf, you will be kicked out of the server, but you can then reconnect.

service pf start

Let's bootstrap a FreeBSD release for the jails - this will be useful later.

bastille bootstrap 14.3-RELEASE update

Now, we create a local bridge. Jails and VMs can be attached to it, making them fully autonomous. Using VNET jails, for example, will allow the creation of VPNs or tun interfaces inside them, simplifying potential future setups (and increasing security by using a dedicated network stack).

Modify the /etc/rc.conf file and add:

# Add lo1 and bridge0 to the list of cloned interfaces
cloned_interfaces="lo1 bridge0"
# Assign an IP address and netmask to the bridge
ifconfig_bridge0="inet 192.168.0.1 netmask 255.255.255.0"
# Enable gateway functionality for routing
gateway_enable="yes"

Let's also modify /etc/pf.conf to allow the 192.168.0.0/24 subnet to access the Internet via NAT. We will skip packet filtering on bridge0 and enable NAT. This isn't the most secure setup, but it's sufficient to get started:

#...
# Skip PF processing on the internal bridge interface
set skip on bridge0
#...
# NAT traffic from our internal network to the outside world
nat on $ext_if from 192.168.0.0/24 to any -> ($ext_if:0)
#...

To ensure the new settings are correct, it's a good idea to test with a reboot.

Since I often configure vm-bhyve in my setups, I prefer to install it right away, creating the dataset that will contain the VMs and installation templates. Remember that zroot is only valid if you installed the entire system on ZFS; otherwise, you'll need to change it to your own dataset:

# Install required packages
pkg install vm-bhyve grub2-bhyve bhyve-firmware
# Create a dataset to store VMs
zfs create zroot/VMs
# Enable the vm service at boot
sysrc vm_enable="YES"
# Set the directory for VMs, using the ZFS dataset
sysrc vm_dir="zfs:zroot/VMs"
# Initialize vm-bhyve
vm init
# Copy the example templates
cp /usr/local/share/examples/vm-bhyve/* /zroot/VMs/.templates/

At this point, I usually enable the console via tmux. This means that when a VM is launched, it won't open a VNC port by default, but a tmux session connected to the VM's serial port. Let's install and configure tmux:

pkg install -y tmux
vm set console=tmux

Let's also attach the switch we created (bridge0) to vm-bhyve so we can use it:

vm switch create -t manual -b bridge0 public

Now, vm-bhyve is ready.

The basic infrastructure is complete. We now have:

• ZFS to ensure data integrity, which will also handle redundancy, etc.
• BastilleBSD to manage jails, useful for backing up Linux, NetBSD, OpenBSD, and non-ZFS FreeBSD machines.
• vm-bhyve to install specific systems (like Proxmox Backup Server).

Backup Strategies

I use various backup tools, too many to list in this article. So I'll make a broad distinction, describing how to use this server to achieve our goal: securing data.

• For FreeBSD servers with ZFS (hosts, VMs, jails, hypervisors, and their respective VMs), I use an extremely useful, efficient, and reliable tool: zfs-autobackup.
• For Linux servers (without ZFS), NetBSD, OpenBSD, etc. (any non-ZFS OS), I usually use BorgBackup. There are other fantastic tools like restic, Kopia, etc., but BorgBackup has never let me down and has served me well even on low-power devices and after incredibly complex disasters.
• For Proxmox servers (a solution I've used with satisfaction in production since 2013, although I'm recently migrating to FreeBSD/bhyve where possible), I use two possible alternatives (often both at the same time): if the storage is ZFS, I use the zfs-autobackup approach. In either case, the most practical solution is the Proxmox Backup Server. And the Proxmox Backup Server is one of the reasons I proposed installing vm-bhyve: running it in a VM and storing the data on the FreeBSD host gives you the best of both worlds. Some time ago, I tried running it in a FreeBSD jail (via Linuxulator), but it didn't work.

Backups using zfs-autobackup

zfs-autobackup is an extremely useful and effective tool. It allows for "pull" type backups, as well as having an intermediary host that connects to both the source and destination, which is useful if you don't want direct contact between the source and destination. I won't describe the latter setup, but the documentation is clear, and I have several of them in production, ensuring that the production server and its backup server cannot communicate with each other.

I usually create a dataset for each server and instruct zfs-autobackup to keep that server's backups in that dataset. The snapshots taken and transferred will all be from the same instant, so as not to create a time skew (some tools of this kind snapshot a dataset, then transfer it, which can result in minutes of difference between two different datasets from the same server).

I've described in detail how I perform this type of backup in a previous post, so I suggest reading that post for reference.

Let's install zfs-autobackup on the FreeBSD server:

pkg install py311-zfs-autobackup mbuffer

Backups for other servers using BorgBackup

When I don't have ZFS available or need to perform a file-based backup (all or partial), I use a different technique. BorgBackup backups are primarily "push" based, meaning the client will connect to the backup server. This is not optimal or the most secure approach, as the backup server should, in theory, be hardened. Even when protecting everything via VPN, the risk remains that a compromised server could connect to its backup server and alter or delete the backups. I have seen this happen in ransomware cases (especially in the Microsoft world), and so I try to be careful to minimize this type of problem, mainly through snapshots of the backup server (an operation that will be described later).

To ensure the highest possible security, I create a FreeBSD jail on the backup server for each server I need to back up. The advantage of this approach is the complete separation of all servers from each other. By using a regular user inside a jail, a compromised server that connects to its backup server would only be able to reach its own backups, as it would be confined to a user account and, even if it managed to escalate privileges, still be inside a jail.

Let's say, for example, we want to back up a server called "ServerA" (great imagination, I know). We create a dedicated jail on the backup server:

# Create a new VNET jail named "servera" attached to our bridge
bastille create -B servera 14.3-RELEASE 192.168.0.101/24 bridge0

BastilleBSD will automatically set the host's gateway for the jail. In our case, this is incorrect, so we need to modify it and set the jail's gateway to 192.168.0.1 in the /usr/local/bastille/jails/servera/root/etc/rc.conf file:

# ...
defaultrouter="192.168.0.1"
# ...

Restart the jail and connect to it:

bastille restart servera
bastille console servera

Now, inside the jail, we install borgbackup:

pkg install py311-borgbackup

BorgBackup doesn't run a daemon; it's launched by the remote server (which sends its data to the backup server), so it's important that the installed version is compatible with the one on the remote host.

Since we'll be using SSH, let's enable it:

service sshd enable
service sshd start

And create a non-privileged user for this purpose:

# The 'adduser' utility provides an interactive way to create a user.
root@servera:~ # adduser
Username: servera
Full name: Server A
Uid (Leave empty for default): 
Login group [servera]: 
Login group is servera. Invite servera into other groups? []: 
Login class [default]: 
Shell (sh csh tcsh nologin) [sh]: 
Home directory [/home/servera]: 
Home directory permissions (Leave empty for default): 
Use password-based authentication? [yes]: 
Use an empty password? (yes/no) [no]: 
Use a random password? (yes/no) [no]: yes
Lock out the account after creation? [no]: 
Username    : servera
Password    : <random>
Full Name   : Server A
Uid         : 1001
Class       : 
Groups      : servera 
Home        : /home/servera
Home Mode   : 
Shell       : /bin/sh
Locked      : no
OK? (yes/no) [yes]: yes
adduser: INFO: Successfully added (servera) to the user database.
adduser: INFO: Password for (servera) is: JIkdq8Ex

The user is created and can receive SSH connections. After setting everything up, I suggest disabling password-based login in the jail's SSH configuration, using only public key authentication.

As mentioned, the biggest risk of a "push" backup is that a compromised client could access the backup server and delete or encrypt the backup history, rendering it useless.

To drastically mitigate this risk, we can configure SSH to force the client to operate in a special Borg mode called append-only. In this mode, the SSH key used by the client will only have permission to create new archives, not to read or delete old ones. However, this approach could complicate some client-side operations (like mount, prune, etc.), forcing them to be done on the server. For this reason, I won't describe it in this setup, "limiting" myself to taking snapshots of the repositories. It can be a very good practice, so I recommend considering it.

Let's initialize the BorgBackup repository. In this example, for simplicity, I won't set up repository encryption. If the jails are on an encrypted dataset or GELI-encrypted disks, there will still be data encryption on the disks, but there will be no protection against someone who could physically access the server while the disks are mounted. As usual, security is like an onion: every layer helps. Personally, I suggest enabling and using it ALWAYS.

# Switch to the new user
su -l servera
# Initialize a new Borg repo named "servera" with no encryption (for this example)
borg init -e none servera

The jail is ready, but it's unreachable from the outside. There are two ways to make it accessible:

• Install a VPN system inside the jail itself. Using tools like Zerotier or Tailscale (which don't need to expose ports) will immediately create the conditions to connect to the jail, which will remain inaccessible from the outside. As the jail is a VNET jail, we're free to choose any of the supported VPN technologies.
• Expose a port on the backup server, i.e., on the host, to allow external connections. Many advise against this path as they consider it less secure. It is, but sometimes we don't have the luxury of installing whatever we want on the server we're backing up.

To expose the port, go back to the host and modify the /etc/pf.conf file, creating the rdr and pass rules to let packets in:

# ...
# Redirect incoming traffic on port 1122 to the jail's SSH port (22)
rdr on $ext_if inet proto tcp from any to any port = 1122 -> 192.168.0.101 port 22
# ...
# Allow incoming traffic on port 1122
pass in inet proto tcp from any to any port 1122 flags S/SA keep state

Reload the pf configuration:

service pf reload

The jail will now be reachable on the server's public IP, on port 1122. Obviously, this port number is for illustrative purposes, and I used from any, but for better security, you should replace any with the IP address of the server that will be connecting to perform the backup.

By repeating this process for each server and creating a separate jail for each, you can have isolated jails in separate datasets with their backups, potentially setting space limits using ZFS quotas.

It's important to remember that backing up a live filesystem (i.e., without a snapshot or dumps) has a very high probability of being impossible to restore completely. Databases hate this approach because files will change while being copied and tend to get corrupted. Of course, it depends on the nature of the data (a backup of a static website will have no issues, but a WordPress database probably will), but it's crucial to think about a technique to snapshot the filesystem before proceeding. For example, I have already written about how to create snapshots on FreeBSD with UFS in a previous article: FreeBSD tips and tricks: creating snapshots with UFS.

I will cover other operating systems in a future, dedicated post.

Proxmox Backup Server in a Dedicated VM

Starting with version 4.0 (which is still in beta at the time of this writing), Proxmox Backup Server (PBS) supports storing its data in an S3 bucket. This is excellent news as it decouples the server from the data. There are great open-source S3 implementations, like Minio or SeaweedFS, which allow for clustering, replication, etc. In this "simple" case, we will install Proxmox Backup Server in a small VM, while for the data, we'll install Minio in a native FreeBSD jail. The advantage is undeniable: the VM will only serve as an "intermediary", but the data will rest directly on the FreeBSD host's dataset, natively. It will also be possible to specify other external endpoints, other repositories, etc.

As a philosophy, I tend not to use external providers unless for specific needs, so installing Minio in a jail is a perfect solution to manage this situation.

Let's install PBS by downloading the ISO from their website (https://enterprise.proxmox.com/iso/) - at this moment, the version that supports this setup is 4.0 Beta.

The directory to download to is the vm-bhyve ISOs directory. It's not strictly necessary, but it's useful for not "losing" it somewhere. So, go to the directory and download it:

cd /zroot/VMs/.iso
fetch https://enterprise.proxmox.com/iso/proxmox-backup-server_4.0-BETA-1.iso

Now let's create a VM with vm-bhyve. We can start from the Debian template, but we'll make some modifications to optimize performance. In this example, I'm giving it 30 GB of disk space, 2 GB of RAM, and 2 cores.

If you want to store all backups inside the VM, you'll need to size the virtual disk correctly (or create and attach another one). In this case, I will focus on the "clean" VM that will store its data on a dedicated jail with Minio.

vm create -t debian -s 30G -m 2G -c 2 pbs

Once the empty VM is created, let's modify its options:

vm configure pbs

We will modify the VM to be UEFI and to use the NVME disk driver - bhyve performs significantly better on NVME than virtio, as previously tested:

loader="uefi"
cpu="2"
memory="2G"
network0_type="virtio-net"
network0_switch="public"
disk0_type="nvme"
disk0_name="disk0.img"

Fortunately, the Proxmox team has provided for the installation of the Backup Server on devices without a graphical interface, so the boot menu will allow installation via serial console. Let's launch the installation and connect to the virtual serial console:

cd /zroot/VMs/.iso
vm install pbs proxmox-backup-server_4.0-BETA-1.iso
vm console pbs

Select the installation via Terminal UI (serial console) and proceed normally as if it were a physical host, assigning an IPv4 address from the 192.168.0.x range (in this example, I'll use 192.168.0.3).

This way, the Proxmox Backup Server will run in a VM, with the ability to take snapshots before updates, etc., without any worries.

Once the installation is complete, PBS will reboot and listen on port 8007 of its IP. Again, as with the jails, we have two options: install a VPN system within the VM itself (thus exposing it automatically only on that VPN - generally a more secure operation) or expose port 8007 on the server's public IP.

In the latter case, add the relevant lines to the /etc/pf.conf file on the FreeBSD backup server:

# ...
# Redirect incoming traffic on port 8007 to the PBS VM's web interface
rdr on $ext_if inet proto tcp from any to any port = 8007 -> 192.168.0.3 port 8007
# ...
# Allow that traffic to pass
pass in inet proto tcp from any to any port 8007 flags S/SA keep state

Reload the pf configuration:

service pf reload

The PBS VM configuration is complete. If you chose to use the PBS's internal disk as a repository, no further operations are necessary (other than the normal repository creation, etc., within PBS).

In this case, however, we will use a different approach.

Creating a Minio Jail as a Data Repository for PBS

This approach, in my opinion, has a number of important advantages. The first is that Minio will run in a dedicated jail on the host, at full performance, and will store the data directly on the physical ZFS datapool, thus removing any other layer in between. This jail could potentially be moved to other hosts (by connecting PBS and the jail via VPN or public IP), made redundant thanks to all of Minio's features, etc. Another solution I am successfully testing (in other setups) is SeaweedFS.

Let's create a dedicated jail with Minio and put it on the bridge, so that PBS can access it on the LAN.

bastille create -B minio 14.3-RELEASE 192.168.0.11/24 bridge0

If not configured directly, BastilleBSD will use the host's gateway for the jail, which is incorrect in this case. So let's go modify it and restart the jail. Enter the jail with:

bastille console minio

And modify the /etc/rc.conf file to have the correct gateway (following the example addresses):

# ...
ifconfig_vnet0=" inet 192.168.0.11/24 "
defaultrouter="192.168.0.1"
# ...

Exit the jail and restart it:

bastille restart minio

Enter the jail and install Minio:

bastille console minio
pkg install -y minio

Minio is already able to start, but PBS, even on the LAN, wants an encrypted connection. Fortunately, there's a handy tool that can generate the certificates for us:

# Download the certgen tool
fetch https://github.com/minio/certgen/releases/latest/download/certgen-freebsd-amd64

# Make it executable and run it for our jail's IP
chmod a+rx certgen-freebsd-amd64
./certgen-freebsd-amd64  -host "192.168.0.11"

# Create the necessary directories and set permissions
mkdir -p /usr/local/etc/minio/certs
cp private.key public.crt /usr/local/etc/minio/certs/
chown -R minio:minio /usr/local/etc/minio/certs/

Let's view the certificate's fingerprint. Since it's self-signed, we'll need it for PBS later. For security reasons, PBS will ask for the fingerprint of non-directly verifiable certificates. Run the following command and take note of the result:

openssl x509 -in /usr/local/etc/minio/certs/public.crt -noout -fingerprint -sha256

At this point, enable and configure Minio in /etc/rc.conf. WARNING: The username and password (access key and secret) used in this example are insecure and for testing purposes only. It is strongly recommended to use different values:

# Enable Minio service
minio_enable="YES"
# Set the address for the Minio console
minio_console_address=":8751"
# Set the root user and password as environment variables
minio_env="MINIO_ROOT_USER=testaccess MINIO_ROOT_PASSWORD=testsecret"

Start Minio:

service minio start

If everything went correctly, Minio is now running (with its certificates) and ready to receive connections.

It's now time to create the bucket(s) that PBS will use. There are several ways to do this, but to test that everything is working and to configure PBS, I suggest connecting via an SSH tunnel.

# Create an SSH tunnel from your local machine to the backup server
# Port 8007 is forwarded to the PBS web UI
# Port 8751 is forwarded to the Minio console
ssh user@backupServerIP -L8007:192.168.0.3:8007 -L8751:192.168.0.11:8751

This way, we'll create a tunnel between the FreeBSD backup server and our workstation, mapping 127.0.0.1:8007 to 192.168.0.3:8007 (the PBS web interface) and 127.0.0.1:8751 to 192.168.0.11:8751 (the Minio console port).

Now, connect to https://127.0.0.1:8751, enter the credentials specified in /etc/rc.conf, and create a bucket.

Once the bucket is created, you can configure PBS to use it. Connect to PBS via https://127.0.0.1:8007 and go to S3 Endpoints. Set a name, use 192.168.0.11 as the IP and 9000 as the port, enter the access and secret keys, and the certificate fingerprint we generated earlier. Enable "Path Style" or it will not work.

Then go to Datastores and add it, as you would for any other S3 datastore, by specifying the created bucket and a local directory where the system will keep its cache.

If everything was set up correctly, PBS will create its structure in the bucket, and from that moment on, you can use it. Always keep in mind that this is still a "technology preview", so there may be issues, but from my tests, it is sufficiently reliable.

Taking Local Snapshots of Backups

One of the most common techniques used in ransomware attacks is to also delete or encrypt backups. They often use automated methods, relying on the fact that many (too many!) consider a "backup" to be a simple copy of files to a network share. However, it's not impossible that, in specific cases, they might compromise the machine and connect to the backup server. This is nearly impossible with a "pull" type backup (like the one managed by zfs-autobackup) but is still possible with the "push" approach, which involves using BorgBackup or similar tools.

This happened to one of my clients once - in that case, the problem originated internally, from an employee who wanted to cover up his mistake, inadvertently creating a disaster - but that will be material for another post.

Fortunately, the client had a system that solved the problem: thanks to ZFS, we can have local snapshots on the backup server, which are invisible and inaccessible to the production server. Since we have already installed zfs-autobackup, it's easy to use it for this purpose as well. I've already talked about this in a previous article and won't rewrite the steps here. Just consult that article, keeping in mind that in this case, it's not advisable to snapshot all the datasets on the backup server (the space would grow exponentially), but only those at risk. In the cases analyzed in this post, this applies only to the push part, as PBS will also be accessible only from the Proxmox servers and not from the VMs they contain. If, in this case too, you don't trust those who manage the Proxmox servers, just set up snapshots for the Minio jail as well.

Conclusion

This long post aims to analyze, in a general way, how I believe one can manage reasonably secure backups of their data. Obviously, there are many variables, additional precautions, possible optimizations, hardening, etc., that must be studied on a case-by-case basis. There are old rules, new rules, old and new philosophies. Recently, many people who have embraced the cloud have often stopped thinking about backups, only to realize it when something happens and the data has, indeed, vanished... into the clouds.

In this post, I have generically covered the setup of the backup server, and this demonstrates how FreeBSD, thanks to its features, can be considered an ideal platform for this type of task.

In the next articles in this series, I will examine the client side, i.e., how to structure them for a sufficiently reliable backup, and how to monitor everything - because I've seen this too: people resting easy because the backup was supposedly running every night, but in fact, the backup had been failing every night for more than 4 years.

Stay Tuned and stay...backupped!

New Article Published

I'm excited to announce that I have published a new, in-depth article on the BSD Cafe Journal: "WordPress on FreeBSD with BastilleBSD: A Secure Alternative to Linux/Docker".

This piece explores how to create a robust and secure WordPress installation on FreeBSD using BastilleBSD, leveraging the power and isolation of FreeBSD jails as a compelling alternative to the more common Linux and Docker stack.

Future Technical Content

I'm excited to announce that I'm expanding my writing to a new platform! From now on, some of my more technical, long-form articles and tutorials will be published on The BSD Cafe Journal, a fantastic hub for BSD-related content that I'm happy to now contribute to.

This new collaboration complements the work I do here. My personal blog will continue to be my home base, and you won't miss a thing! I'll still be posting my own articles and announcements right here, and I'll always include a direct link to any new content I publish elsewhere. This space will remain as active as ever.

Thank you for reading

Learn how to build a self-hosted CDN using FreeBSD jails, Nginx, Varnish, WireGuard, and PowerDNS to improve media delivery for your website without relying on external providers.