pointless and uncaring

slog is neat
23 October, 2025
- tech
Why did golang get a new logging framework after ~10 years?

Existing “log” package
- It’s like printf but there’s a timestamp!
- I can call Fatal and Panic; I guess that’s nice in main.
```
log.Println("Hello, 世界")
log.Printf("I will see you in %s\n", time.Hour)
log.Fatal("Oh no it's all gone wrong")
```
▶ Go Playground
```
2009/11/10 23:00:00 Hello, 世界
2009/11/10 23:00:00 I will see you in 1h0m0s
2009/11/10 23:00:00 Oh no it's all gone wrong
```
- There is SetOutput; let’s come back to this later.
- I can get a default logger, but I can’t set a default logger.
What should libraries do?

This approach is pretty great if you’re in main.go, but what if you’re writing a package for reuse?
- Should libraries log at all?
  - Well, I can pass up errors (wrap them if you’re extra cool!).
  - What about that thing that seems wrong but I’m suppressing? I don’t want to break compatibilities, but I’m pretty sure this is wrong.
  - https://github.com/Azure/eno/pull/532
- How do I get the logger?
  - Use a package-level variable or function?
  - Pass it everywhere and save it as a member?
  - Shove it onto context?
  - I have seen all of these approaches.
- How do I log respectfully without over-logging?
- What if I want to customize my log format?
  - JSON or binary
  - Coloring or emoji ✅
What about X
- zap and logrus are pretty neat. They introduced level, structure, sinks, and a lot of the stuff I am going to talk about below.
- But if I’m writing a library, how do I know what interface to get/log to?
- I have seen several instances where logging adapters are regularly rewritten, usually losing something each time.
- Sometimes it’s nice to just have the standard library.
Enter Slog

Gopkg and Blog
- There are package- and type-level log functions.
  - Most call into func Log(ctx context.Context, level Level, msg string, args …any)
- There is a high-level interface you can implement behind (handlers).
- It includes text and JSON implementations.
- You can set the default! (Claim: just use the slog.Info/Error methods instead of passing loggers around most of the time.)
Levels

Debug/Info/Warn/Error
- Throw all your junk into debug! Main.go can set level and toss it.
- Indecisive? Warn away!
- Make up your own log levels and go crazy.
```
	slog.Debug("Hello, 世界")
	slog.Info("I will see you in ", "time", time.Hour)
	slog.Warn("It's getting hot in here")
	slog.Error("Nuts", "error", fmt.Errorf("next time gadget"))
	slog.Log(context.TODO(), slog.Level(10), "more important than an error"
```
▶ Go Playground
```
  2009/11/10 23:00:00 INFO I will see you in  time=1h0m0s
  2009/11/10 23:00:00 WARN It's getting hot in here
  2009/11/10 23:00:00 ERROR Nuts error="next time gadget"
  2009/11/10 23:00:00 ERROR+2 more important than an error
```
Structure
- String parsing stinks because we have Kusto and lots of other fancy databases!
- Request IDs, other tracing items, latencies as durations, and general metadata all live happily here.
- Attrs have lots of nice built-in types, and there are groups (I have not used groups).
- You can set attributes globally on the default or a particular log instance with With/WithAttrs.
```
slog.With("author", "pmiller")
	slog.Info("I will see you in ", "time", time.Hour, slog.Duration("strongtypedtime", time.Minute),
		slog.Float64("pi", math.Pi),
		slog.Group("plane", "make", "boeing", "model", "737"))
	attrs := []slog.Attr{
		slog.String("author", "jsonpaul"),
		slog.String("program", "playground"),
	}
	handler := slog.NewJSONHandler(os.Stdout, nil).WithAttrs(attrs)

	slog.SetDefault(slog.New(handler))
	slog.Info("I will see you in ", "time", time.Hour, slog.Duration("strongtypedtime", time.Minute),
		slog.Float64("pi", math.Pi),
		slog.Group("plane", "make", "boeing", "model", "737"))
```
▶ Go Playground
```
 2009/11/10 23:00:00 INFO I will see you in  time=1h0m0s strongtypedtime=1m0s pi=3.141592653589793 plane.make=boeing plane.model=737
{"time":"2009-11-10T23:00:00Z","level":"INFO","msg":"I will see you in ","author":"jsonpaul","program":"playground","time":3600000000000,"strongtypedtime":60000000000,"pi":3.141592653589793,"plane":{"make":"boeing","model":"737"}}
```
Putting things on context?
- Okay, what if I want to add attributes that everything below me in the call stack can use? ChatGPT
- I can stick them on the context, but then the code I am calling has to pull them out.
- There is an InfoContext and equivalent for each helper.
- I didn’t find a great solution for this unless you control the handler too.
- Then you can use https://github.com/PumpkinSeed/slog-context. Example blog
- Written by someone named PumpkinSeed, so it must be good. I kind of wanted a standard library way to do it, though.
Handlers!
- The fools gave us an interface: Handle(context.Context, Record) error
- Maybe you don’t want JSON; maybe you want BSON or Thrift or whatever compact nonsense.
- Maybe you want to forward records to Kusto, Blob storage, Kafka, or Service Bus.
- You can do whatever you want!
- Sometimes you might just want to do an io.Writer, but then you lose per-record control.
- My Azure append blob example.
  - [Started as a handler](https://github.com/paulgmiller/careme/blob/8621cce4cd5f240d881a0bf4e22cb26c281c8e8e/internal/logsink/appendblob.go
  - Hit an interesting deadlock
  - Went to a writer much simpler but regret it because now I might split log records
Awesome slog
- Lots more stuff here
- Want to log tests to console but put JSON somewhere else? slog-multi!
- Want to log directly to Teams? (WHY!) Here’s a forwarder. Maybe Datadog or Kafka make more sense.
- Miss Logrus? There’s an adapter
Comments
2026 hikes
21 October, 2025
- hiking
What are we doing next year
- knox creek/ Throp mountain
- Finish the wonderland
- Gem Lake
- Cascade pass
- Freemont lookout
- Storm king mountain
Comments
Lasthike
5 October, 2025
Comments
summer whistler
24 August, 2025
- family
- photos
Comments
wireguard kubernetes
18 August, 2025
- tech
What if you built your kubernetes cluster with wireguard/tailscale from the ground up.
- Api servers are always private except for wireguard port
  - Use a side car for each api server pod https://tailscale.com/kb/1185/kubernetes#sample-sidecar. Drop a tailnet secret in ccp.
  - like api server vnet integration pods show up in your tailnet/overlay even if they have no nodes. Can be hit directly with endpoint slices
  - No konnectivity proxy ingress and egress goes direct to api server pods
  - Could still have a internal load balancer for legacy nodes but ideally want none of that.
  - Rest of ccp does not need to be in tailnet but could be
- Head scale lets you join the cluster. (can we use certs to verify this public key is from a baremetal node? We can’t be the ones giving them a secret?)
- Nodes are locked down except for wireguard port (use the top of rack)
  - node wireguard nic get a /24 and we use a bridge to hand out pod ips. (could use cilium or soemthign else if you want net pols but don’t need it for ipam)
  - node ip should be just the first ip.
- Do you need boostrapping tokens? If you’re on the network is that enough? Alot of tls becomes uncessary
- Can tailscale provide the load balancer service? https://tailscale.com/kb/1223/funnel to get things to kubenet.
Can we prove this out by 1) sidecarring api servers on a standalone 2) Creating a BYO CNI cluster 3) Adding a Demonset to VM nodes that boostraps tailscale and gets a /24 + brige plugin 4) Create a script that adds us to tailscale then gets a bootstrap token and joins api server from some random home machine?
Comments