Chameth.com - posts like docker-automatic-nginx-proxy, finding-an-awkward-bug-with-claude-code but not migrating-from-github-to-forgejo, understanding-docker-volume-mounts, why-you-should-be-using-https

The Case of the Unchanging Config

2026-04-09T00:00:00Z

Last week I was attempting to make it so I could share pictures on IRC directly from my client. This sounds simple, but it involves a bouncer¹ that proxies the request to a standalone image hosting service that I had to modify to be compatible. At one point my testing loop was:

Make a change to the hosting service
Commit it
Tag a new release
Wait for it to build
Update the version I’m running on my server
Reconfigure the bouncer to pass the new parameter or change the URL or whatever
Try uploading a photo from my phone
Realise I’ve overlooked something and go back to step one

I could have set up a local copy of everything and tested it sensibly. I should have set up a local copy of everything and tested it sensibly. But it seemed like such a trivial change, and setting up the whole environment seemed like such a pain. After the third or so iteration of failure I was pretty annoyed with myself, computers, and basically everything.

My biggest annoyance was that my bouncer would not pick up the new URL from the config when I changed it. It’s meant to reload the config when it receives a SIGHUP, and it claimed to in the logs, but I could clearly see it was still hitting the old URL. Restarting the bouncer to update the config is a pain, as it disconnects me from all the IRC networks, and has to reconnect to them all, reauthenticate, etc. It also mildly spams everyone who shares a channel with me².

When I finally got everything working I had a look at the bouncer source, and thought I’d spotted the issue. I raised a bug report, ending in this remark:

It looks like the config is reloaded properly, but the handler for uploads is created once at startup and has its own copy of the uploader, so effectively snapshots the config to whatever it is at startup:

fileUploadHandler := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
    cfg := srv.Config()
    h := fileupload.Handler{
        Uploader:    cfg.FileUploader,
        DB:          db,
        Auth:        cfg.Auth,
        HTTPOrigins: cfg.HTTPOrigins,
    }
    h.ServeHTTP(w, r)
})

I’m not quite sure what I thought I saw there, but that’s not right. I want to call it a hallucination, but that feels like a very overloaded word these days. I guess I was just seeing what I wanted to see, and that was a good excuse to stop investigating. Still, even if I misidentified the cause, the bug was still present, right? … Right?

The Science of Deduction

When the developer got back and said he couldn’t reproduce my issue, I went back to my install and immediately reproduced it. With a calmer head, I figured it was probably something with my particular set up. First thing to check: am I actually running the version I think I am?

Unfortunately the bouncer doesn’t actually seem to expose the version anywhere that I can see. It’s not in the logs, it doesn’t have a -version flag, and none of the IRC-based status commands seem to include it. But I know it’s a Go app, and I know Go embeds the version information. go version -m <binary> will dump it all out, but the binary is inside a Docker image, and the Docker image is one of my nice, minimal, artisanal ones so doesn’t ship a go binary. No problem, docker compose cp bouncer:/bnc ./bnc yoinks the binary out of the container, and then dumping the version shows that, yes, I am running the version I thought I was. Hmm.

The next tool I reached for in my toolbox was strace. Maybe it’s not actually reading the file for some reason? I immediately executed strace -p <pid> -e trace=openat,open,read,pread64 -f without having to look any part of that up. Yep. Definitely. Then I edited the config, HUP’d the bouncer, and saw that it was… reading the config file. As it’s meant to. By default strace truncates strings to 32 bytes, so I couldn’t actually see the line I’d changed. Some more definitely-not-RTFMing later, and rerunning it with an extra -s 65536 let me see the full config. Surprise! The config hadn’t changed!

To confirm my findings, I used docker compose cp again, this time yoinking the config file from inside the container. The inside config file was definitely different to the outside config file. What? My hypothesis at this point was “something something Docker nonsense”. I mount the config as read-only, and was wondering if that meant that Docker was doing something weird instead of just bind mounting it. A quick trip to /proc/<pid>/mounts showed that it was, in fact, not doing anything weird, and was just bind mounting it.

If the file is bind mounted, then surely it’s the same file? I ran stat on the file on the host, noted the inode number, then pondered how to actually run stat on the file inside the container, given the aforementioned awkwardly minimal image. The solution was easy: access it via /proc/<pid>/root/. I could’ve saved myself a bunch of docker compose cp if I’d thought about that earlier. Oh well. The inode of that file was different. What?

The answer was DNS. It’s always DNS. Oh, sorry, force of habit. I meant the answer was vim. I was editing the config in vim, and when it saves files, by default, it writes the new content to a temporary file and does an atomic rename. That’s normally a good thing: it prevents corruption if the write fails midway through³. In this case, though, that means the new file has a new inode. Bind mounting a file binds to the inode, so the container just keeps clutching onto the original config from when it was started, blissfully unaware that the party has relocated down the street.

This problem is likely to happen whenever you bind mount a file into a container. When I mentioned this to a friend, he immediately responded “oh yeah, never do that”, and went on to describe the horrible hacks he’s had to add to Ansible to sidestep the issue. The nicer solution is to just bind mount an entire directory if you can, as then it doesn’t matter what happens to the files within it. I really like having the config files sat alongside the Docker compose files, though; having to create a directory just to work around some bind mount weirdness upsets me.

Now I knew what the problem was, I found there was an issue raised against Docker twelve years ago⁴. The first response was “that’s expected”, and I fully agree with the author: “respectfully, that might be expected by you, but it was not expected by me”!⁵ As a result of the issue they added a nice note to the docs:

Note: Many tools used to edit files including vi and sed --in-place may result in an inode change. Since Docker v1.1.0, this will produce an error such as “sed: cannot rename ./sedKdJ9Dy: Device or resource busy”. In the case where you want to edit the mounted file, it is often easiest to instead mount the parent directory.

But in the intervening twelve years, both the note and the functionality described have gone missing. Ho-hum.

I found it interesting how I’ve only just hit this problem, given how long I’ve used Docker. But I realised that almost everything I run I’m happy to just restart. Cattle, not pets, and so on. My IRC bouncer is one of the few exceptions to that. The only other thing I regularly hot reloaded was Centauri, my reverse proxy, but that had a whole config directory mounted because it was shared between containers, so nicely sidestepped the foot-gun.

So lesson learnt: check for weird bind mount issues before raising issues about config hot reloading. It’ll join the esteemed company of “maybe the drive is full and causing completely unrelated problems?”, “perhaps everything is dog slow because the kernel ran out of entropy?”, and “did systemd sneakily take over that functionality while you weren’t looking?” in the troubleshooting checklist.

An IRC bouncer is basically an always-on proxy. It connects to the IRC networks for you, then your clients connect to your bouncer. The bouncer can then send incoming messages to all your different clients, cache them when you’re offline and replay them later, and lots of other nice things people take for granted in their chat apps these days. ↩︎
Join/part/quit spam is part of IRC, and clients have ways of handling it, but I still feel bad about doing it excessively. ↩︎
You can disable this in vim by setting backupcopy=yes. Good luck redoing all the debugging if you ever accidentally remove that from your vimrc, though! ↩︎
Which is weird, because Docker can’t possibly be that old. That would make me much older than I’m prepared to accept. ↩︎
If you read the thread they didn’t actually mean to sound so dismissive, but it’s still pretty funny. “Yes, it’s expected that the foot-gun causes your foot to hurt. Duh. What did you expect?” ↩︎

An interesting Tailscale + Docker gotcha

2026-01-29T00:00:00Z

As I’ve written about before, I use Tailscale for a lot of things. I thought I had it set up in a reasonably secure manner, but I recently noticed a problem.

I use Tailscale’s ACLs to limit what each node can access, based on the tags I apply to it. So an app node can’t access anything via Tailscale, while an integration or server node can access things tagged with either app or integration. This is expressed pretty simply in the Tailscale ACL JSON:

{
	"tagOwners": {
		// Servers that can be SSH'd into
		"tag:server": [],
		// Applications that are exposed on tailscale but never connect out
		"tag:app": [],
		// Things which talk to other services over tailscale (connecting to apps etc)
		"tag:integration": [],
	},
	"grants": [
		{
			// Users can access everything
			"src": ["autogroup:member"],
			"dst": ["*"],
			"ip":  ["*"],
		},
		{
			// Servers and integrations can access integrations and apps
			"src": ["tag:server", "tag:integration"],
			"dst": ["tag:integration", "tag:app"],
			"ip":  ["*"],
		},
	],
}

Over the past week I’ve been setting up a private Forgejo instance behind Tailscale, complete with an actions runner that runs things using a docker-in-docker container. I didn’t want the runner knowing anything about Tailscale, so I had it configured to speak to Forgejo direct over HTTP (forgejo:3000) instead of using full Tailscale HTTPS URL that I use when accessing it (http://git.example-net.ts.net/)¹.

Everything was going fine, until I forgot to do that translation… and it worked. My Forgejo action runners could access anything on my tailnet. I run several things on Tailscale that just have authentication turned off, on the basis that only authorised devices can access them. Things like the admin interface for this website. I definitely didn’t intend for any workflow I run on my git server to have access to edit my website!

This wasn’t just limited to Forgejo, either. Any docker container I was running could access the tailnet. In hindsight it’s fairly obvious why: the host is running Tailscale, connected as a node tagged with my server tag. That creates a tailscale0 interface, and automagically sets up iptables rules to route Tailscale traffic over the interface. Docker also automagically sets up iptables rules to bridge traffic, and apparently these two sets of rules interact in such a way that traffic from Docker containers is allowed to route via the tailscale0 interface.

I say it’s fairly obvious in hindsight — there’s no reason why Docker would special case any particular host interface after all — but it still feels pretty surprising. Because both bits of software inject their own iptables rules, I never really had a good mental model for how they interact. The host Tailscale node was a completely separate building block to Docker. It would be a pain to use either of them if they didn’t do these rules, but it’s also one of the reasons I don’t really like “magical” things².

There are lots of ways to fix this, but none of them feel particularly great. You can configure both Tailscale and Docker to not automatically fiddle with iptables and handle the rules yourself, but I really hate dealing with iptables³. As a stopgap I did hold my nose and add some iptables rules to drop traffic to the tailscale0 interface if it originated from the IP ranges that Docker was configured to use:

iptables -I DOCKER-USER -s 192.168.0.0/16 -o tailscale0 -j DROP
iptables -I DOCKER-USER -s 172.17.0.0/12 -o tailscale0 -j DROP

The DOCKER-USER chain is a nice little escape hatch; it comes before the main auto-generated DOCKER chain, and Docker leaves the rules in it alone.

So I added these rules, and felt pretty good about myself, and then… everything started breaking in weird ways. After some debugging I realised the problem was DNS⁴. The server’s DNS resolver is 100.100.100.100, a special Tailscale address. This is so that it can resolve tailnet hostnames via MagicDNS (and so I don’t have to configure my custom DNS servers manually on each device; Tailscale does it for me). My new iptables rules inadvertently dropped all the DNS packets coming from docker containers. D’oh.

Obviously the solution here is to double down and add MORE iptables rules:

iptables -I DOCKER-USER -s 192.168.0.0/16 -d 100.100.100.100 -o tailscale0 -j ACCEPT
iptables -I DOCKER-USER -s 172.17.0.0/12 -d 100.100.100.100 -o tailscale0 -j ACCEPT

This got DNS working again, but it all feels a bit gross. Not least because to actually make these persist you need to use iptables-save, and then you also get all the rules that both Docker and Tailscale have inserted. I came across a script to filter out the Docker ones, but… yuck⁵.

I realised a cleaner idea might just be to lock down what server tagged nodes could access within Tailscale itself. That way I can avoid touching iptables at all. It’s handy to be able to curl services from the host when debugging, but it’s not really necessary. So I removed that access, and… stuff broke again. I’m using my Forgejo instance as a registry for some of the docker images I run, so the docker daemon needs to be able to reach it. I ended up making a new tag for infrastructure, which can be accessed from server devices. This does still allow all the docker containers to reach Forgejo, but I already have it set up with appropriate access controls and public/private repository splits. Forgejo is a service designed to run publicly, so this seems a reasonable trade-off for convenience. I used the tests feature of Tailscale’s ACL config to make sure I’d got the rules right:

{
	"tests": [
		{
			// Servers can only access infrastructure
			"src":   "tag:server",
			"proto": "tcp",
			"allow": [
				"tag:infrastructure:8080",
			],
			"deny": [
				"100.84.16.43:8080",
				"me@example.com:8080",
				"tag:server:8080",
				"tag:app:8080",
				"tag:integration:8080",
			],
		},
	],
}

The whole situation still feels a bit messy. If I ever get around to switching to nftables I might loop back and manually craft some rules for routing traffic, instead of leaving Tailscale and Docker to do their own thing.

I discovered later on that logging into the container registry at forgejo:3000 actually issued a redirect to https://git.example-net.ts.net/ so this was all basically for naught… ↩︎
I didn’t use Tailscale’s MagicDNS for a long time just because the word “magic” put me off. Only when I eventually got around to learning how it worked, and seeing that it wasn’t really that magical under the hood, did I change my mind. ↩︎
nftables seems far better in lots of ways, but I can’t really be bothered migrating. Maybe next time I reimage the server for whatever reason… ↩︎
It’s always DNS… ↩︎
Again, nftables would almost certainly help here. It actually has (gasp) configuration files. But again, I really didn’t want to spend the time migrating. ↩︎

Finding an awkward bug with Claude Code

2025-06-30T00:00:00Z

The many times Claude thought it was right…

I recently encountered a bug in one of my projects that I couldn’t immediately figure out. It was an issue in Centauri, my reverse proxy. After its config was updated, I noticed¹ it stopped serving responses. Looking at the logs, I could see it was obtaining new certificates from Let’s Encrypt for a couple of domains, but I’d designed it so that wouldn’t block requests (or so I thought). After a few minutes of confusion, everything went back to working.

It felt like an issue with locking, but a quick glance didn’t show me anything obvious. I decided to throw the problem at Claude Code and see how it did. With some guidance² it managed to identify the cause, but it was a bit of a journey to get there. I’m going to share the prompts and responses, as it was an interesting process.

The initial prompt

I thought the issue might be quite obvious, and I’d just overlooked it. Centauri isn’t that complex a project, after all. So I gave Claude a fairly minimal prompt to set it off:

> Hello! Please examine this codebase. I have noticed an issue that when a
  certificate is being renewed, the application seems to not serve any other
  requests. I suspect there is a misplaced lock somewhere.
 
  Can you see what's happening? Think very hard.

Claude Code has special hardcoded triggers to engage “thinking” mode, which makes it take more time to think about alternatives. The keywords are: “think”, “think hard”, “think harder”, and “ultrathink”. I managed to not quite get the right incantation here, and only enabled the lowest level³.

I’ll omit all the responses where Claude is “thinking” (talking to itself), or decides to make todo lists then cross items off them then add different items, and so on. Its responses get fairly lengthy; I won’t judge you too harshly if you skim over them. Its first response was blessedly brief:

● I found the issue! Looking at the certificate manager code, I can see the
  problem:

… and then it went back to thinking and reading files, because it had neither found the issue nor could it see the problem. Ho-hum. Some time later it came back with:

● I found the issue! The problem is in certificate/manager.go:53-54:
  
  m.store.LockCertificate(subject, altNames)
  defer m.store.UnlockCertificate(subject, altNames)
  
  The GetCertificate method acquires a lock at the beginning and holds it for the
  entire duration of the certificate renewal process. This includes:
  
  1. Contacting the ACME server
  2. Solving DNS-01 challenges (can take minutes for DNS propagation)
  3. Obtaining the certificate
  4. Getting OCSP staple
  5. Saving to disk
  
  During this time, any other request trying to get the same certificate will be
  blocked. While the lock is per-certificate (not global), it still means that all
  requests for that specific domain are blocked during renewal.
  
  Let me check if there are other places where this might cause blocking:

And it then proceeded to do even more thinking, checking files, etc, until it concluded with:

● I found the root cause! The issue is actually worse than just a
  per-certificate lock. Here's what's happening:
  
  The main blocking issue is in proxy/manager.go:133-144:
   
  func (m *Manager) CheckCertificates() {
      m.lock.RLock()  // ← Acquires read lock on the entire manager
      for i := range m.routes {
          route := m.routes[i]
          // ...
          m.updateCert(route)  // ← This can take MINUTES for ACME renewal!
      }
      m.lock.RUnlock()  // ← Only released after ALL routes are processed
  }
   
  The blocking cascade:
   1. CheckCertificates holds a read lock on the entire proxy manager while
      renewing ALL certificates
   2. Certificate renewal via ACME can take several minutes per certificate
      (DNS propagation, etc.)
   3. Every incoming request calls routeFor (proxy/manager.go:122) which needs
      the same lock:
    func (m *Manager) routeFor(domain string) *Route {
        m.lock.RLock()  // ← BLOCKED until CheckCertificates completes!
        match := m.domains[strings.ToLower(domain)]
        m.lock.RUnlock()
        // ...
    }
  
  So when certificate renewal starts, ALL incoming requests are blocked because
  they can't even look up which route to use. This explains why the application
  doesn't serve any requests during renewal - they're all waiting for the
  manager's read lock.
  
  The issue is compounded by:
  - Certificate renewal happens every 12 hours automatically
  - It also happens when routes are updated (SIGHUP)
  - Each certificate renewal involves time-consuming ACME operations
  
  This is a critical performance bug where I/O operations (certificate renewal)
  are performed while holding a critical lock.

This doesn’t actually follow on from the earlier response, it’s completely flipped onto a different lock being the issue. Amusingly, it’s nearly right here, but it’s highlighting this code for completely the wrong reason.

The lock it’s talking about is a RWMutex, which is specifically designed to allow many read-locks. Everything it’s pointing out are read locks, so they cannot block one another as it described. The response Claude has given sounds very confident, seems appropriately technical, but is in fact complete rubbish. This is one of the big issue with all LLMs at the minute, and it makes me despair at the idea of people “vibe coding” things when they don’t know enough to push back against nonsense like this.

Course correction

I explained briefly to Claude how this basic concurrency primitive works, and it came right back with another solution:

> But that's a RW Mutex. You can hold many read locks concurrently.

● You're absolutely right! I apologize for the confusion. With a RWMutex,
  multiple readers can hold the lock concurrently. Let me trace through this
  more carefully.
  
● Task(Analyze updateCert method)
  ⎿ Done (8 tool uses · 20.0k tokens · 1m 11.9s)
  
● Excellent analysis! You're right about RWMutex allowing concurrent readers.
  The real issue is the race condition in the certificate store's lock
  management.

  Looking at certificate/store.go:82-91:

  func (j *JsonStore) lockFor(subjectName string, altNames []string) *sync.Mutex {
      key := strings.Join(append([]string{subjectName}, altNames...), ";")

      if mu, ok := j.locks[key]; ok {
          return mu
      } else {
          mu = &sync.Mutex{}
          j.locks[key] = mu  // ← Concurrent map write!
          return mu
      }
  }

  This is a classic concurrent map access bug. When multiple requests hit the
  same domain during certificate operations:

  1. Multiple goroutines call lockFor() simultaneously
  2. They read/write to j.locks map without synchronization
  3. Go runtime can:
    - Panic with "concurrent map writes"
    - Corrupt the map structure
    - Block indefinitely if map internals get corrupted

  This explains the blocking behavior - it's not the RWMutex, but rather a race
  condition causing either panics or deadlocks when the map gets corrupted
  during concurrent certificate operations.

  The fix would be to either:
  1. Add a mutex to protect j.locks map access
  2. Use sync.Map instead of a regular map
  3. Pre-initialize all locks at startup

I do appreciate when me stating a basic fact is called out as “excellent analysis”. Definitely doesn’t make me feel like I’m having smoke blown up my ass.

I think the issue it’s describing here is actually a problem⁴, albeit one that’s unlikely to be hit. But it’s definitely not the one we’re looking for. Claude even describes the possible consequences of the bug, and they don’t match what happened at all.

Maybe logs will help?

While Claude was chugging away, I’d been looking into the issue as well. I had logs from when it happened, and figured Claude may be able to glean something from them too. The first thing I’d noticed is that I wasn’t logging when a process finished, so I’d already added that to the code base, hence the note so Claude doesn’t get itself confused:

> That may be an issue, but it's not the one I was looking for. I've found logs
  from the time it happened. Note that a few logs have been added to the
  codebase since (so we can't tell from the logs when ACME or OCSP processes
  *finished*, just when they started). Here are the logs:
  
  {"time":"22:40:53.187887976Z","level":"INFO","msg":"Received signal, updating routes...","signal":1}
  {"time":"22:40:53.19947532Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:40:53.199507942Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:41:09.048465968Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:41:09.048482309Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:41:17.419299795Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:41:17.419325744Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:41:23.238027386Z","level":"INFO","msg":"Received signal, updating routes...","signal":1}
  {"time":"22:41:29.994330646Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:41:29.994353329Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:41:43.658794811Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:41:43.658819548Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:41:57.635325513Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:41:57.635347515Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:42:10.52439779Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:42:10.524417107Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:42:23.159430249Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:42:23.159452171Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:42:36.347407326Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:42:36.347433295Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:42:52.544628775Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:42:52.544653993Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:43:25.409579107Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:43:25.409602842Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:43:43.819479448Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:43:43.819502041Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:43:54.358957563Z","level":"INFO","msg":"Renewing certificate", ...}
  {"time":"22:43:54.358979355Z","level":"INFO","msg":"Starting ACME process to obtain certificate", ...}
  {"time":"22:44:06.180278128Z","level":"INFO","msg":"Renewing certificate", <big list of alt names>}
  {"time":"22:44:06.180302585Z","level":"INFO","msg":"Starting ACME process to obtain certificate", <big list of alt names>}
  {"time":"22:45:03.830331922Z","level":"WARN","msg":"Failed to connect to upstream","error":"context canceled", ...}
  {"time":"22:45:03.830528134Z","level":"WARN","msg":"Failed to connect to upstream","error":"context canceled", ...}
  {"time":"22:45:03.832388931Z","level":"WARN","msg":"Failed to connect to upstream","error":"context canceled", ...}
  {"time":"22:46:10.225854865Z","level":"WARN","msg":"Failed to connect to upstream","error":"context canceled", ...}
  {"time":"22:46:10.488432434Z","level":"WARN","msg":"Failed to connect to upstream","error":"context canceled", ...}
  {"time":"22:46:10.488439397Z","level":"WARN","msg":"Failed to connect to upstream","error":"context canceled", ...}
  {"time":"22:46:30.617614536Z","level":"WARN","msg":"Failed to connect to upstream","error":"context canceled", ...}

I’ve cut down the logs a bit to remove the specific domains, simplify the timestamps, and remove a bunch of other cruft. The key points are there were two SIGHUPs received, and the last certificate before everything started breaking covered 6 different hostnames.

Claude then went through a phase of periodically shouting “Eureka!” then deciding it hadn’t actually found anything yet:

● Now I see it! The key insight is in the timing and the "context canceled"
  errors. Let me trace through what's happening:
  
● Ah, now I see what's happening! Looking at the logs more carefully:

  1. 22:40:53 - First SIGHUP received, routes updated
  2. 22:40:53 to 22:44:06 - Multiple certificates start renewing (14 certificates!)
  3. 22:41:23 - Another SIGHUP received while renewals are still in progress
  4. 22:45:03 onwards - "context canceled" errors start appearing

  The issue is that SetRoutes is being called AGAIN (at 22:41:23) while the
  previous CheckCertificates() goroutine is still running and renewing
  certificates. Let me check what happens in this scenario:
  
> (The context canceled errors are clients timing out because centauri wasn't
  proxying properly)
  
● Ah, now I understand! The "context canceled" errors are from clients timing
  out because Centauri wasn't proxying their requests. This confirms the
  blocking issue.

  Let me check how incoming requests actually get their certificates:
  
● Now I found the real issue! Looking at the timeline and the code:

  1. 22:40:53 - SIGHUP triggers SetRoutes() which:
    - Calls loadCertificate() for each route
    - Sets route.certificate to the existing certificate
    - Starts go m.CheckCertificates() in background
  2. 22:41:23 - Another SIGHUP while certificates are still renewing
    - SetRoutes() is called AGAIN
    - Creates NEW Route objects with route.certificate = nil initially
    - Replaces the old routes in m.routes
    - The old Route objects (with valid certificates) are replaced!

  The critical issue is in proxy/manager.go:42-65:

  func (m *Manager) SetRoutes(newRoutes []*Route, fallback *Route) error {
      // ... validation ...

      for i := range newRoutes {
          route := newRoutes[i]  // These are NEW Route objects!
          // ...
          m.loadCertificate(route)  // Try to load certificate
      }

      m.lock.Lock()
      m.routes = newRoutes  // REPLACES all routes!
      m.lock.Unlock()
  }

  When the second SIGHUP arrives:
  1. New Route objects are created
  2. For domains currently being renewed, GetExistingCertificate might fail or
     return expired certs
  3. These routes get route.certificate = nil and CertificateMissing status
  4. CertificateForClient returns nil for these domains
  5. TLS handshake fails → clients can't connect → "context canceled" errors

  The fix would be to either:
  1. Prevent concurrent CheckCertificates runs
  2. Copy certificates from existing routes before replacing them
  3. Make certificate renewal truly independent of route management

Midway through it had started searching for “context canceled” errors in the codebase, which doesn’t make much sense. That’s what happens when a client gives up waiting, so the corresponding backend request is terminated. You won’t actually find “context canceled” anywhere in Centauri, because it’s a feature of the Go standard library. I gave it a little nudge to try to stop it rabbit holing.

As far as I can tell, what it’s talking about in this response isn’t an issue. I guess it’s possible that there’s a race condition whereby a route would be replaced and end up with an old version of a certificate that was just being renewed, but it wouldn’t cause an issue in practice; it’d just get updated next time the certificates were refreshed.

There’s no way for routes to get a nil certificate in this scenario, as far as I know. But I don’t have the unfounded confidence of an LLM, so I simply said:

> I don't follow why the certificate would be nil?

Attempt the third

I expected Claude to just explain its reasoning, but it apparently didn’t have any reasoning so set off on another voyage of discovery:

● You're right to question that. Let me trace through this more carefully.

● Good point! Since GetExistingCertificate doesn't acquire locks, it should
  return the existing certificate even during renewal. Let me dig deeper:
  
● Now I see a potential race condition! Here's what could happen:

● I found it! There's a critical race condition that explains the blocking:

  The Race Condition

  1. Background renewal starts: CheckCertificates() → updateCert() →
     GetCertificate() acquires lock and starts ACME renewal
  2. SIGHUP arrives: SetRoutes() is called, which for each route calls
     loadCertificate() → GetExistingCertificate() → store.GetCertificate()
  3. The race:
    - store.GetCertificate() returns a pointer to the certificate in the slice
    - Meanwhile, the renewal completes and calls SaveCertificate()
    - SaveCertificate() calls removeCertificate() which modifies the slice while
      it's being read
    - This slice modification (append(j.certificates[:i], j.certificates[i+1:]...))
      can cause:
        - The pointer returned by GetCertificate to become invalid
        - Slice bounds issues
        - Memory corruption
  4. The result:
    - GetExistingCertificate might return corrupted data or nil
    - Routes end up with certificate = nil
    - TLS handshakes fail
    - Clients timeout → "context canceled"

  The Timeline Confirms This

  - 22:40:53 - First SIGHUP, starts renewal of 14 certificates
  - 22:41:23 - Second SIGHUP arrives (30 seconds later) while renewals are still running
  - 22:45:03 - First timeout errors (4 minutes after first SIGHUP)

  The 4-minute delay matches ACME renewal time. The second SIGHUP's SetRoutes
  likely read certificates while they were being modified by the first SIGHUP's
  renewal goroutines.

  This is a classic concurrent slice modification bug - no synchronization
  between readers (GetCertificate) and writers (SaveCertificate/removeCertificate).

Once again we have a very confident and technical answer that does not explain the symptoms we’re looking for. It’s also mostly rubbish again. GetCertificate and SaveCertificate are only called by code that is holding a lock for that certificate; they won’t interfere with one another. There may, again, be a small concurrency issue if different certificates are handled in parallel, but it’s definitely not the cause of our current problems.⁵

Working through these explanations and debunking them was annoying me at this point. I decided to give it once last chance, and used the “ultrathink” shibboleth:

> So there may be a concurrent modification issue, but that doesn't explain the
  symptoms. If a TLS handshake failed it wouldn't result in the clients timing
  out. Likewise if there was any kind of corruption there'd be a panic, not a
  period where it didn't work and then started working again. Engage ultrathink.

Fourth time lucky

I’m going to include a piece of the thinking here because it amused me so much:

✻ Thinking…

  Oh! I need to think about this differently. Let me trace through what happens
  when a request comes in:

Apparently it takes “ultrathink” to actually work through the problem logically? Amusingly it did actually get there:

● You're absolutely right. Let me think more carefully about this...

● EXACTLY! That's the issue. Looking at the timeline:

  1. 22:40:53 - First SIGHUP: CheckCertificates() acquires READ lock, starts
     renewing 14 certificates
  2. 22:41:23 - Second SIGHUP: SetRoutes() tries to acquire write lock, blocks
     waiting
  3. Incoming requests call routeFor() for route lookup, but Go's RWMutex blocks
     new readers when a writer is waiting (to prevent writer starvation)
  4. 22:45:03 - Clients timeout because requests can't even get past route lookup

  The key insight is Go's RWMutex writer starvation prevention: when a writer
  is waiting, new readers are blocked even though the current lock holder is
  also a reader.

  So the cascade is:
  - CheckCertificates (read lock) → certificate renewal takes minutes
  - SetRoutes (write lock) → waits for read lock to release
  - routeFor (read lock for every request) → blocked by waiting writer

  Result: All incoming HTTP requests hang at route lookup, never reach TLS
  handshake or proxying logic. Clients timeout waiting for any response
  → "context canceled".

  This perfectly explains:
  - No panics (no corruption)
  - No immediate TLS errors (never gets that far)
  - Requests hang then timeout (blocked on lock)
  - Eventually recovers (when certificate renewals finish)

  The fix would be restructuring to avoid holding the manager lock during slow
  certificate operations.

We’re back to the RWMutex again, but this time it’s dropping knowledge on me not the other way around: if there’s a writer waiting, then new readers are blocked. This does indeed perfectly explain the symptoms, and was something I’d overlooked. The write lock is only held briefly to update some struct members to prevent it from blocking the readers, but this behaviour coupled with the long-running read lock ends up doing that anyway.

The fix

After all this work it’s very tempting just to ask Claude to fix it, but I had almost no confidence it’d be able to maintain any of the concurrency requirements given how much it flailed around with bad explanations. But in the interests of science, I asked it to try. Here’s what it came up with:

   // CheckCertificates checks and updates the certificates required for registered routes.
   // It should be called periodically to renew certificates and obtain new OCSP staples.
   func (m *Manager) CheckCertificates() {
+    // Create a snapshot of routes to avoid holding the lock during slow certificate operations
     m.lock.RLock()
+    routes := make([]*Route, len(m.routes))
+    copy(routes, m.routes)
+    provider := m.provider
+    m.lock.RUnlock()

+    // Process certificate renewals without holding the manager lock
-    for i := range m.routes {
+    for i := range routes {
-      route := m.routes[i]
+      route := routes[i]
 
       if provider == nil {
         route.certificateStatus = CertificateNotRequired
       } else {
         m.updateCert(route)
       }
     }
-    m.lock.RUnlock()
   }

   // updateCert updates the certificate for the given route.

That’s about what I expect. It definitely fixes the exact issue described, but it’s a bit… sloppy. Routes are now being updated when the lock isn’t being held, which is a change in semantics. It also doesn’t address the underlying issue that requests can be blocked if there’s a pending writer, and pending writes can be blocked by readers. A better approach would be to allow these operations to happen without blocking, maybe by using an atomic pointer to an immutable slice. I guess I’ve got some coding to do!

As for Claude: while the journey was frustrating, I think trying to figure this out by myself would have been even more so. It’s one of those things where everything looks right to me, and I’d have to go down a fairly long chain of checking my assumptions to find where I’d gone wrong.

There’s definitely a lot of room for improvement in coding LLMs: being able to estimate their own confidence level instead of stating everything as absolute truth would go a long way. I imagine there’s also a lot of benefit of having one LLM vet the responses of another; that’s sort-of how thinking mode works, but it doesn’t seem to go far enough to stop it outputting rubbish. It’ll be interesting to see how they develop in the future, anyway.

read as “I got spammed by a bunch of monitoring notifications” ↩︎
read as “Me repeatedly telling it that it was talking rubbish” ↩︎
If only we had some kind of language model technology that could understand this kind of thing. ↩︎
I sense some fun times with Go’s race detector in my near future ↩︎
The most likely outcome from these concurrency issues is Centauri panicking, which isn’t great but would only interrupt ongoing connections before it restarts and starts serving traffic again. Much less problematic than the “block for five minutes” behaviour I’m trying to fix. ↩︎

Coming around on LLMs

2025-05-28T00:00:00Z

Claude says hi.

For a long time I’ve been a sceptic of LLMs and how they’re being used and marketed. I tried ChatGPT when it first launched, and was totally underwhelmed. Don’t get me wrong: I find the technology damn impressive, but I just couldn’t see any use for it.

Recently I’ve seen more and more comments along the lines of “people who criticise LLMs haven’t used the latest models”, and a good number of developers that I respect have said they use coding models in some capacity. So it seemed like it was time to give them another shake.

The first decision to make was which model to try. OpenAI are no longer the only player in the game, every tech company of a certain size is now also somehow an AI company¹. I looked a bit at some benchmarks, and then mostly ignored them and went with the only company that I didn’t outright hate: Anthropic, and their model Claude.

Initial impressions

The latest Claude models do feel a lot more “capable” than the earlier ChatGPT versions I remember, but they also still have a lot of the same problems. At their heart, they’re still text-prediction models, and still seem to be trained to predict text that will please the user rather than be factually accurate or useful.

One of the improvements that people kept mentioning was the ability for models to access the web directly. I’ve seen it kick in a bit naturally, and asked for it explicitly sometimes, and it’s… nothing special? Are people just really bad at searching the web? Maybe they should try Kagi? I feel like about 80% of the time I could have found the information just as quickly myself, about 10% of the time it searched and then hallucinated an answer, and the remaining 10% it found something quicker than I otherwise would have. Those aren’t great results, especially when you consider how much cheaper and simpler just searching the web is.

There is another way to trigger web searches in Claude: using “research mode”. When you enable it, it splits off into different models: a “lead researcher” to come up with a plan, and then some minions that execute it. It ends up doing hundreds of web queries in the span of a few seconds. Suddenly I understand why things like Anubis need to exist. I’ve not used the “please launch a DoS attack” button since.

What did impress me, though, was its ability to churn out reasonable-ish code. It can hack together a bash script as well as I can, and do it far faster than I’d be able to. Sometimes they even work. That made me wonder what it would be like doing actual coding with it. Anthropic have a CLI tool called claude-code, so I paid them lots of money and gave it a spin.

Coding with Claude

The first thing I notice about claude-code is that I really like the interface. It’s basically an input box in a terminal. It doesn’t force me to use a certain IDE or do things in a certain way. By default it asks before making any changes, showing you a side-by-side diff of what it’s doing and allowing you to provide feedback. If I had to design a way to interact with a coding agent from scratch, I can’t think of many things I’d improve.

A simple example of a claude-code session

The power in claude-code versus just using the web UI is that it can use tools. It can query git, run grep commands, even use sed if it wants to change something in lots of files at once. It’s very good at figuring out its way around a codebase, even without any explicit instructions. You can see in the screenshot that with a little prompting it managed to get the permalink to this post; I didn’t tell it where the posts were stored, or how to work out the latest, I just told it when it was wrong. If I’d run /init before it would have probably picked up on the fact that my posts have custom permalinks, and noted it in CLAUDE.md.

But how good is it at actually writing code? It’s like having a keen but not particularly thorough Junior Engineer at your beck and call. If you give it a clearly defined task and guidance on how to implement it (and maybe some feedback as it suggests changes), it’s more than capable of doing it. If you don’t give it enough guidance it tends to go more off the rails. I tried having it generate a simple application from scratch with minimal technical guidance and no review of what it was doing, and it made such a mess of it I decided it was quicker to throw it away and start again by hand.

The man behind the curtain

Even with sufficient guidance, at times it’s really obvious that it’s an LLM generating pleasing-token-strings and not something that genuinely understands what it’s doing. It will spit out code like this:

if err != nil {
    if err == sql.ErrNoRows {
       return nil, err
    }
    return nil, err
}

Why’s that check for sql.ErrNoRows there? It’s entirely pointless. There was no instruction to check for it, none of the existing code checked for it, but I assume it comes up quite a bit in the training data. But it didn’t understand why, so it put the check in, and returned the exact same thing as if it hadn’t.

It also occasionally tries to “cheat” or solve problems the wrong way. It sometimes feels a bit like you’re asking for wishes from a Monkey Paw. “Stop the unit tests failing”, you’ll say; Claude will respond with a request to delete the failing test. The man behind the curtain isn’t particularly well hidden, and the training data and pattern matching often shows through.

This lack of understanding also makes it very hard to get Claude to use comments in a sensible way. It absolutely loves doing nonsense like this:

// render form
function renderForm() {
    // ...
}

Even with explicit instructions to avoid useless comments, or comments that only explain what the code does (not why it does it), or other prompts. Again, I’m chalking it up to a mixture of training data that does that, and lack of any actual understanding about why a human developer may want a comment to exist. Of course, there are plenty of flesh-and-blood devs out there that also don’t comment effectively, so maybe I shouldn’t give Claude too much of a hard time on this.

… and yet …

So with all those problems, it sounds like it’s just not worth it, right? Well, not quite. I feel like programming mostly consists of two distinct tasks: thinking how to implement things; actually implementing them; and then debugging why you’re off-by-one somewhere. Letting an LLM do the thinking is a big no-go for me: it’s evidently not good at it, and frankly that’s one of the things I most like about programming. Having let it loose on small projects, I shudder to think what the codebases of all the “vibe coded” projects that are popping up are like.

For the second stage, though, it’s actually quite nice. If I’ve thought through how I want something to be implemented, I can feed those steps to Claude and have it churn out the otherwise not-too-interesting code. This requires far less time and concentration on my part than writing code, to the extent that I can be thinking about the next feature, or doing something else at the same time. I don’t feel like I’m going to end up deskilling myself this way, as I’ve already formed the idea of what I want to code, I’m just using the LLM to spit it out faster than I can type it.

A lot of the most boring bits of coding like implementing CRUD-y operations can be summarised as “Look at this file/function. Do the same thing but slightly differently elsewhere”. Claude is great at this, especially when explicitly prompted like that. You’re basically playing to its pattern-matching strengths, rather than asking it to come up with anything novel. Most of my prompts tend to be prefixed with “Look at @some_file.go and @other_file.go.” to cue up the patterns I want it to use, and I find this works well.

As for debugging, it’s a mixed bag. It’s sometimes amazingly insightful, and sometimes just runs around in circles trying the wrong things over and over. It’s worth asking the question, but I definitely wouldn’t rely on it over my own abilities.

The future

I’m probably going to carry on using claude-code, at least for personal projects. I’ve got so much done that I just wouldn’t have been bothered to do if I was doing it all by hand. I very much enjoy being in the more “diffuse thinking” mindset, planning how things are going to work, rather than being stuck in the mines digging out SQL queries. After all, who wouldn’t want an over-eager assistant to work on all their hobby projects?

Work is a slightly different matter: a private project or even an open source project that disclaims any liability is different to something I’m being paid to deliver, and bear responsibility for fixing if it’s not done correctly. I’m not saying I won’t use it at all, but if I do it’ll be much more constrained than I would in personal projects.

As for non-code usages: I’m not sold. My sceptic hat is still firmly in place. I don’t think chat is a particularly good interface for many things, and hallucinations are still a big problem despite what people say. I hate the tide of AI slop that’s taking over the Internet, and how LLM-powered chat agents are being forced into every random product. You’re definitely not going to be seeing any AI-generated blog posts from me!

One topic I’ve not gone into here is the ethical concerns about using LLMs. They obviously exist, and I do have thoughts, but that’s a topic for another day.

I’m surprised it’s not gone more mainstream: why’s there no Tesco Value LLM model, yet? ↩︎

Docker reverse proxying, redux

2022-12-06T00:00:00Z

Six years ago, I described my system for configuring a reverse proxy for docker containers. It involved six containers including a key-value store and a webserver. Nothing in that system has persisted to this day. Don’t get me wrong – it worked – but there were a lot of rough edges and areas for improvement.

Microservices and their limitations

My goal was to follow the UNIX philosophy of “do one thing and do it well”. Unfortunately, that doesn’t really work when applied to network services that have to interact with one and other. UNIX tools are built upon a common file system and simple data passed over STDIN. Microservices don’t have that shared foundation. You could make one: companies that use microservices in anger often have a team that deals with the “developer experience” of creating and using microservices. But as a solo developer that’s not something I wanted to spend my time doing.

This became very apparent when trying to debug issues. In the UNIX world, if your series of commands piped together does something strange you can simply echo the data at various stages. Not so much when that data is flying around on a network, going into and out of things like etcd. Trying to figure out why a certificate hadn’t been acquired was a case of searching through logs from four containers, none of which had particularly good logging. There are many ways to get insight into what’s happening with microservices but, again, that’s not something I wanted to do myself.

Over time, and with experience in other projects, I came to realise that microservices only really make sense if you’re unable to deploy a monolith. For tech companies this naturally happens when different teams contribute to the same service: splitting it out into smaller services that are wholly owned by individual teams makes sense. For solo developers, that never happens. You can still gain the other benefits of microservices – such as code separation and having clearly defined APIs – by sticking to certain coding standards.

Proxy inconveniences

As well as being unhappy with the microservice nature of the solution, I wasn’t pleased with nginx. If you requested an unknown domain, nginx would use the first server block in its config to serve a response, instead of sending an “unrecognised name” alert as I wanted. It was a minor issue, but it irked me.

So from nginx I switched to haproxy. It has a strict-sni option when configuring TLS connections which makes it behave properly. It also performs a lot better for this type of workload than nginx. All was well for a while, but then I started getting alerts that requests were occasionally failing. I couldn’t reproduce the issue, but my nightly jobs to build and push containers managed to hit it nearly every night, causing them to fail.

After some investigation, I found that the haproxy developers had refactored the header parsing code, and neglected to properly reset flags when multiple requests were sent over the same connection. There was a patch, but it wasn’t released. No problem, I thought, I’ll just cherry-pick it onto the last release… Except that haproxy use Git in the most convoluted manner I’ve ever seen – they have one repository per release. This makes it harder to patch, but it also made me question whether I trusted them to ship stable software: there were no tests for the header parsing code (which is both fundamental and finicky, the perfect target for tests), the source code management was weird, and they didn’t seem in any rush to patch this bug.

Not long after that issue, Greg managed to encounter another bug where haproxy returned a 500 error whenever the upstream server replied with a particular, perfectly valid, header. The die was cast – it was time to move to something else.

Not Invented Here syndrome

Looking for a new solution, there were many more options than back in 2016. I’m still convinced, however, that anything exposed to the Internet should not have access to run docker containers. It’s the modern equivalent of running a CGI script as root. That single requirement eliminates most off-the-shelf solutions. What do you do when nothing quite meets your specific requirements? You make something yourself! My new solution has two components: Dotege and Centauri.

Dotege is a replacement for the microservices that monitored containers and obtained certificates. It’s fundamentally a templating engine - whenever the containers change, it evaluates a template and saves the result to disk. The template has access to details about the containers, their labels, ports, and so on. Dotege can also obtain certificates from Let’s Encrypt, and raise a signal against another process whenever the template or certificates change. I used this to generate the configuration and certificates used by haproxy for a while, and more recently changed the template so that it works for Centauri.

Centauri is my own reverse proxy. It’s configured using a simple text file and can also obtain certificates from an ACME provider. It doesn’t serve static content, has no knowledge about docker, and avoids the other bells and whistles that adorn most reverse proxies. It also has good test coverage to ensure that I don’t, say, accidentally break header parsing.

As a software engineer I enjoy writing software, but I also enjoy running simple, easy to understand software. That’s what I’ve achieved here: it’s very easy to identify where the problem is if anything goes wrong, both are small Go programs rather than vast sprawling C monstrosities, and their interaction is primarily through a file written to disk that can be inspected or edited as needed.