log/slog: JSONHandler should deduplicate keys

[veqryn]

What version of Go are you using (go version)?

8edcddd

Does this issue reproduce with the latest release?

yes

What did you do?

	logger := slog.New(slog.NewJSONHandler(os.Stdout))
	logger.LogAttrs(
		context.Background(),
		slog.LevelWarn,
		"fancy message",
		slog.Bool("msg", false),
		slog.String("my attr", "something"),
	)

What did you expect to see?

JSONHandler should be fixed by deduplicating those special keys (time, message, level, and source).
I use "source" as a key in a lot of my code right now, so I'd run into this problem.

Either

{
	"time": "2023-04-01T04:27:46.0959443-06:00",
	"level": "WARN",
	"msg": "fancy message",
	"my attr": "something"
}

Or

{
	"time": "2023-04-01T04:27:46.0959443-06:00",
	"level": "WARN",
	"msg": "fancy message",
	"msg.msg": false,
	"my attr": "something"
}

Or

{
	"time": "2023-04-01T04:27:46.0959443-06:00",
	"level": "WARN",
	"msg": "fancy message",
	"fields": {
		"msg": false,
		"my attr": "something"
	}
}

Or

{
	"time": "2023-04-01T04:27:46.0959443-06:00",
	"level": "WARN",
	"msg": "fancy message",
	"msg#01": false,
	"my attr": "something"
}

What did you see instead?

Invalid json. It would break our log aggregation. Default/Builtin handlers should never create invalid output.

{
	"time": "2023-04-01T04:27:46.0959443-06:00",
	"level": "WARN",
	"msg": "fancy message",
	"msg": false,
	"my attr": "something"
}

@jba

[dsnet]

Whatever solution is chosen here, I would argue that it should work even if the user explicitly provides duplicate names:

logger.LogAttrs(
	context.Background(),
	slog.LevelWarn,
	"fancy message",
	slog.String("duplicate", "alpha"),
	slog.String("duplicate", "bravo"),
	slog.String("duplicate", "charlie"),
	slog.String("duplicate", "delta"),
	slog.String("duplicate", "echo"),
)

One possible solution is to use a "#%d" scheme similar to how testing.T.Run deduplicates conflicting test names.
For example:

{
	"time": "2023-04-01T04:27:46.0959443-06:00",
	"level": "WARN",
	"msg": "fancy message",
	"duplicate": "alpha",
	"duplicate#01": "bravo",
	"duplicate#02": "charlie",
	"duplicate#03": "delta",
	"duplicate#04": "echo",
}

[seankhliao]

I think the current situation is fine, duplicate keys aren't invalid, just less interoperable.
Changing the key name to deduplicate it seems worse.

As for the original issue, I'll note that option 3 was why WithGroup was added (though this doesn't affect deduplication in general).

[dsnet]

duplicate keys aren't invalid, just less interoperable.

Technically you are correct. Under RFC 8259, section 4, duplicate names are essentially treated as undefined behavior.

In a possible future major version of "encoding/json", we will likely make duplicate keys an explicit error (which is still complaint behavior with RFC 8259) as this lack of checking has been exploited, resulting in severe security vulnerabilities (e.g., CVE-2017-12635). It's unlikely that duplicate names will have such drastic security implications in logging, but strict checking of duplicate names is likely the overall direction that JSON will gradually move towards.

As an aside, Tim Bray, the author of RFC 8259 actually recommends against naïve compliance with RFC 8259, but rather that implementations adhere to RFC 7493 which makes strict decisions on issues that RFC 8259 leaves as undefined.

[jba]

I think it's important to understand whether encoding/json v2 with its currently proposed semantics is likely to be the accepted version. We should probably align slog with those semantics, even if slog appears in Go before encoding/json v2. We're already doing that for formatting NaNs, infinities and, probably, durations (#59345).

For this issue, if we think no dups will be the default for encoding/json v2, then I wouldn't be opposed to making it the default in slog and adding an AllowDuplicateNames option. And we should do the same for TextHandler for consistency.

[dsnet]

whether encoding/json v2 with its currently proposed semantics is likely to be the accepted version

That I can't really answer unfortunately.

We haven't even started the GitHub discussion on anything related to v2 "encoding/json" and may not be ready to do so for some months. A significant piece of work that remains is figuring out the inter-compatibility story between v1 and a hypothetical v2. That's perhaps the hardest part and requires quite a lot of head-scratching.

That said, if there is a v2 "encoding/json", I feel like the security arguments for rejecting duplicate names by default is pretty compelling.

---

The (very) tentative plan for v2 is that v1 "encoding/json" will actually be implemented in terms of v2 with the appropriate number of options specified that makes it behave semantically identically to what v1 "json" does today. Essentially, with the appropriate set of options, you can configure the semantics to be any hybrid mix that is in-between the default behavior of v1 and the default behavior of v2.

Since "log/slog" already depends on the v1 semantics of "encoding/json", you could imagine doing nothing to address this issue for now. I don't think we should block adoption of "slog" on a hypothetical v2 "json". Nor should we overly massage the output of "slog" to match a hypothetical v2 "json".

Supposing v2 "json" one day lands, you could imagine that v2 "json" be integrated into "slog" by exposing something like:

package slog

type HandlerOptions struct {
    // JSONOptions is a list of Option
    // that configures how JSON is marshaled.
    JSONOptions []json.Option
}

To log each message, "slog" would do something like:

// Start with DefaultV1Options since "slog" has historically
// marshaled with mostly v1 semantics by default.
opts := []json.Option{json.DefaultV1Options}
// We have historically specified Encoder.SetEscapeHTML(false).
// By passing a nil escape function, we achieve the same behavior.
opts = append(opts, json.EscapeRune(nil))
// Append any user-specified options,
// which will override any previously set options.
opts = append(opts, h.opts.JSONOptions...)
return json.Marshal(v, opts...)

(You'll note that the options API for "json" above is different than what currently exists in "go-json-experiment/json". I mentioned elsewhere in #59339 that we're considering switching to variadic options as we see it as one of the ways to resolve the v1-to-v2 compatibility tensions).
(Also ignore the performance implications of appending a bunch of options right before calling json.Marshal; we can find ways to to optimize this. Let's focus on semantics.)

Thus, for someone to use v2 "json" with "slog", they would do something like:

slog.HandlerOptions{
    JSONOptions: []json.Option{json.DefaultV2Options},
}.NewJSONHandler(...)

I should also note that DefaultV1Options and DefaultV2Options represent a composite set of options. For example DefaultV1Options includes AllowDuplicateNames(true) while DefaultV2Options includes AllowDuplicateNames(false).

You could also simplify the above by providing NewJSONv2Handler that implicitly inserts json.DefaultV2Options before user-specified options.

One could imagine that "slog" inspects HandlerOptions.JSONOptions and checks AllowDuplicateNames:

• If true, it does what it currently does (i.e., just emits duplicate object members)
• If false, it does automatic name-deduplication similar to what I proposed in log/slog: JSONHandler should deduplicate keys #59365 (comment).

[jba]

Thanks, @dsnet, that makes a lot of sense.

[08d2]

It would be very unexpected if repeated calls to Log with the same key "duplicate" were to produce output like {"duplicate": "alpha", "duplicate#01": "bravo", "duplicate#02": "charlie"}. Among many other reasons, this behavior would make it impossible to reliably assert that a log event emitted with some key K would actually be rendered with that key K in the ultimate output format.

Structured logging defines a concept of a log event which is conceptually above any specific encoding like JSON or logfmt or et cetera. That concept of a log event is well-defined by convention, and that definition doesn't accommodate repeated keys.

[dsnet]

It would be very unexpected if repeated calls to Log with the same key "duplicate" were to produce output like {"duplicate": "alpha", "duplicate#01": "bravo", "duplicate#02": "charlie"}.

Possibly, but there are no good answers when it comes to duplicate names.

It would be nice if the API made it statically impossible to have duplicates,
but I don't see how that can be done without taking in a map[string]any as the input,
which would not be performant.

Given the possibility of runtime duplicates, your options are:

1. Emit the object with all duplicate fields
2. Emit the object, but only keep one of the duplicate fields (e.g., the first occurrence)
3. Rename the duplicate fields to avoid a conflict
4. Report an error
   a. and still do either 1, 2, or 3
   b. or simply reject the message and not log it.

With option 1, you risk the log parser later not being able to read the message entirely or accurately. Duplicate names result in undefined behavior where the parser might ignore duplicates, only use the first one, only use the last one, or use some merged form of all duplicates.

With option 2, you avoid any ambiguities when reading, but are subtly dropping information when writing.

With option 3, you avoid any ambiguities as to the meaning when reading, but are changing the meaning when writing. At least the meaning when reading is consistent across all parsers.

Option 4a is decent as it allows the caller to at least realize that they're possibly holding the API wrong, but I suspect most people will not check the error of logging calls.

Of all the bad options, option 4a with either 1 or 3 is probably the best approach.
Options 2 and 4b discard information, which seems antithetical to the point of a logging library.

[08d2]

Option 2 is pretty clearly the correct choice.

Options 2 and 4b discard information, which seems antithetical to the point of a logging library.

I'm not sure why anyone would expect a structured logging library to never discard information. Emitting multiple structured log events with the same key is pretty clearly going to resolve to a single "winner" for that key, most likely the last event, but really any event would be acceptably valid.

[dsnet]

Option 2 is pretty clearly the correct choice.

That's a matter of opinion. Just as how you find option 3 an odd position, I find option 2 an odd position. Every option has pros and cons.

I'm not sure why anyone would expect a structured logging library to never discard information.

Any value passed to LogAttrs by the programmer carries fairly clear intent that it be logged. Arbitrarily choosing some field to keep, while discarding all others goes against the intent. Thus, my expectation is for a logging library to try hard to never discard information.

[08d2]

Values passed to a structured logging library necessarily include a key which is, by definition, understood as unique within the target log event. So, no: it's definitely not the case that calling a structured logging method like LogAttrs can be assumed to produce any specific log output. Structured logging is, hopefully obviously, not equivalent to log.Printf, or fmt.Printf, or anything like that.

[veqryn]

@dsnet For option 4, how would an error be reported? A separate log line at warning level, or some out-of-band channel or something?

Personally I would have probably expected option 2, where the latest value wins (except that the Message, Time, Level, and Source fields would always win over any Attribute values).

My preferences, would probably be:
option 2 or 4a2
option 3 or 4a3

option 1 would break a lot of log aggregation systems
option 4b would seem like the worst option, as you lose the log line completely, making any error even harder to diagnose

[dsnet]

Values passed to a structured logging library necessarily include a key which is, by definition, understood as unique within the target log event

By technical definition, I agree. However, in practice, duplicates often occur due to human error. If I make a mistake, I would rather still have all duplicates logged (whether with option 1 or 3).

As an aside, we should definitely create a vet check that detects accidental duplicate names.

@dsnet For option 4, how would an error be reported? A separate log line at warning level, or some out-of-band channel or something?

I was thinking of an explicit error returned by slog.Handler.Handle. I guess a separate log line is a variation of option 4. However, a go vet check might be a better way to statically detect most of these cases.

[08d2]

Values passed to a structured logging library necessarily include a key which is, by definition, understood as unique within the target log event

By technical definition, I agree. However, in practice, duplicates often occur due to human error. If I make a mistake, I would rather still have all duplicates logged (whether with option 1 or 3).

I'm not sure how to evaluate whether or not a particular mistake made by consumers of an API is common enough to warrant changing the API to accommodate that mistake. Especially so if that change would violate the intuitive expectations of the API, as is the case here.

[AndrewHarrisSPU]

It would be nice if the API made it statically impossible to have duplicates, but I don't see how that can be done without taking in a map[string]any as the input, which would not be performant.

I agree, I think a map is at least very close to the ideal data structure here. Regarding performance, the current implementation of x/exp/slog/internal/buffer allocates output buffers as a sync.Pool resource.

type Buffer []byte

It's possible to also pool a map[string]bool for keys, and a []string for groups

type Buffer struct{
	buf []byte
	keys map[string]bool
	group []string
}

On freeing the Buffer, like the buf slice, the keys and group fields can be emptied without releasing backing memory. I used something similar experimenting with string interpolation of attributes into log messages which was fun and way more useless in practice than I anticipated, but I did benchmark and this Buffer approach seemed to:

• not allocate
• increase the time of logging proportional to the number of attributes (loosely and non-scientifically, each Attr was 20% more expensive - use 5 Attrs and pay for 6)
• increase the size of pooled objects by a constant factor - the additional fields are small, but so is the buf slice.

---

A contrasting approach: I think a Deduplicate wrapper for ReplaceAttr would be possible without any change to slog, although tricky because deduplication after ReplaceAttr seems significant. Notably, with no change to slog, I don't think this allows for your option 4 of reporting an error, which I thought was intriguing. Something like:

func Deduplicate(func([]string, Attr)(Attr)) func([]string, Attr) Attr

(I really don't have any opinion on what's necessary here - I don't see much need for deduplication personally, prior art in structured logging in Go doesn't deduplicate AFAIK, and yet the the arguments for deduplication seem pretty reasonable)

[veqryn]

Values passed to a structured logging library necessarily include a key which is, by definition, understood as unique within the target log event

By technical definition, I agree. However, in practice, duplicates often occur due to human error. If I make a mistake, I would rather still have all duplicates logged (whether with option 1 or 3).

As an aside, we should definitely create a vet check that detects accidental duplicate names.

Would that be able to detect duplicates in sub-packages though? I feel like the most common occurrence of duplicates will be adding attributes and telemetry before passing the logger down to libraries or other sub-packages.

@dsnet For option 4, how would an error be reported? A separate log line at warning level, or some out-of-band channel or something?

I was thinking of an explicit error returned by slog.Handler.Handle. I guess a separate log line is a variation of option 4. However, a go vet check might be a better way to statically detect most of these cases.

Right now, errors coming from Handler are discarded by the most common slog methods (ie: Info, WarnCtx, etc). I certainly wouldn't want to change the signature of Info, WarnCtx, etc, to include returning the error. I suppose it is possible to use a middleware type handler that captures and does something with errors, but that wouldn't be something people would use by default.
In any case, I would still want the log line outputted before any error returned.

[dsnet]

Would that be able to detect duplicates in sub-packages though?

Depends on the complexity of call-tracking. This is a similar problem as vet checks for fmt.Sprintf, which already has some degree of call tracking. I don't know if it descends through package boundaries, but it is able to cross function call boundaries.

[zephyrtronium]

One option might be to rename duplicated keys to !BADKEY(duplicateKey) as a sort of combination of @dsnet's options 3 and 4a while carrying some symmetry with the existing error-while-logging mechanism. Then again, there is also some asymmetry if it is only JSONHandler that checks duplicates.

[dsnet]

Another argument against option 2 (which discards data) is security.

If untrusted input is able affect what names are present, then that can be used as a mechanism to hide malicious activity, knowing that it will not be logged. IMO, it seems better for readers to possibly choke on the malformed messages, than for the malicious activity to not be properly logged. An human investigator looking at the malformed message can see the duplicate shenanigan at play.

[08d2]

There are many additional (and unavoidable) security concerns in a structured logging API beyond this one, often related to how values are rendered. It is always an error to pass untrusted input to a structured logger directly.

[dsnet]

@08d2

There are many additional (and unavoidable) security concerns in a structured logging API beyond this one

Like what? The other security issue I can imagine is a DOS attack where you're tricked into logging something massive, but that's a distinctly different class of security issue from being able to hide activity or perform remote-code-execution.

@AndrewHarrisSPU

I think a map is at least very close to the ideal data structure here. Regarding performance, the current implementation of x/exp/slog/internal/buffer allocates output buffers as a sync.Pool resource.

I'm not sure I understand. Suppose we used a Go map such that you did something like:

slog.Info(map[string]any{
    "msg":   "my wonderful message",
    "other": "something else",
})

Yes, you get the benefit of static compilation preventing duplicate map keys (although it can't check the implicit "time" and "level" keys). However, I'm missing the part where the pool helps out.

When slog.Info calls the underlying slog.Handler.Handle, it goes through an interface method call, and all escape analysis goes out the window, so the input map is treated as having escaped (and thus heap-allocated). To ensure the input remains stack allocated, you would have to copy the input map to one of the pooled maps and then call Handler.Handle with it. It's not clear to me that the act of copying the map is going to be fast enough.

[AndrewHarrisSPU]

@dsnet

When slog.Info calls the underlying slog.Handler.Handle, it goes through an interface method call, and all escape analysis goes out the window, so the input map is treated as having escaped (and thus heap-allocated). To ensure the input remains stack allocated, you would have to copy the input map to one of the pooled maps and then call Handler.Handle with it. It's not clear to me that the act of copying the map is going to be fast enough.

I don't think we need to input maps to achieve option (2) or (3), I think I should clarify. This technique doesn't involve any change that isn't internal to Handler implementation, everything else looks the same as before. There's no map copying; there is one blank, probably already-allocated map obtained from a pool when the output Buffer is obtained by Handler.Handle. (edit: some attention to pre-formatted attributes is needed here, but it can be pretty efficiently done)

This map can be populated with keys visited when the handler traverses the attributes of the record, writing to output. A couple of quirks are that the traversal is made more interesting and dynamic by things like group Attrs, LogValuers that expand into groups, or ReplaceAttrs, etc. ReplaceAttr needs to happen before a key goes into the map, but this is tractable. Also, precisely, it's not just key strings that need to be matched, it's also the group prefix - a little tricky but possible to do safely and without allocating.

If this map is populated with visited keys, we can detect whether some key has been committed previously in the log line, and trigger behavior option (2) or (3) as a result. Big picture, I'm not enthusiastic about this as a solution, but I was surprised that it seemed more performant than I would've guessed.

[08d2]

Like what? The other security issue I can imagine is a DOS attack where you're tricked into logging something massive, but that's a distinctly different class of security issue from being able to hide activity or perform remote-code-execution.

Consider a type which renders in one way in some version X, and a different (and richer) way in a subsequent version Y. Consider the case where that change results in sensitive information from that type which was previously not logged becoming logged. This happens easily if the type grows a new sensitive field in a new version. Such changes are in practice never considered breaking in terms of versioning. A downstream application which logs a value of that type will have one behavior if it includes version X of the relevant package, and a different behavior when it includes version Y, and that version upgrade is in practice never going to be noticed, or tested, by the downstream application.

Printf based logging does not generally have this problem, because it is not conventional to pass values directly as a Printf log arguments.

[veqryn]

fwiw we once used a database driver that logged our database password in plain text when in debug mode, using regular old Printf (it was patched right away)

I don't really see any security issues with structured logging that don't already exist with regular logging.

And its not like this is log4j and we are executing the statements in the logs or trying resolve remote xml or something. If the data is strings, the worst case I can think of is accidental logging of sensitive data.

[jba]

A few points:

@veqryn's specific problem, duplicate "source" keys, will be easy to resolve with a ReplaceAttr function once #59280 is approved. You can then distinguish a "source" attribute that comes from the handler by seeing if the value is a *slog.Source. It's also going to appear before a user-supplied attribute, a fact you can use right now. The design doesn't promise that that will always be true, but I can't see why we'd change it.

In general, if you don't like duplicate keys, you can write a wrapping Handler. That can return an error if you like, but it's worth remembering that logging is often the last resort for debugging in the wild. If you've ever been in a position where you can't debug because you have a picky logger that refuses to output everything it has, you know how frustrating that can be, and how happily you'd jettison your principles of careful error detection and failing fast for just a few more characters on the screen.

This isn't a proposal, but if it were I'd vote to decline. We can always add dedup functionality later (though not as the default, it's true). As for json v2 compatibility, @dsnet has provided a good blueprint for how to achieve that when the time comes. But you go into the standard library with the packages you have, not the packages you wish you had. slog should behave like the current encoding/json.

[carlmjohnson]

What if the fields are changed to dupe.1 dupe.2 etc but to keep it from going unnoticed there’s a new field added with a canonical name like “duplicated-fields”: [“dupe”] and if you see anything called duplicated-fields in your logs you know to go figure out what’s misconfigured.