Logging


#21

Moderator note: don't make posts like this

Ewww, logging tons of stuff is disgusting, don't pollute Swift with elaborate "solutions" to this "problem".


#22

This is highly inappropriate. Please follow the Swift code-of-conduct.


#23

Of course, always!


(John McCall) #24

It's completely reasonable for Swift to include a nice logging library, especially on a server. Once that library exists, there's nothing Swift can reasonably do to prevent it from being over-used (accepting that that's a problem for the sake of argument). So your comment isn't providing any value here; you're just throwing insults at other people's work.


#25

Swift already has a nice logging library:

func print(_ items: Any..., separator: String = default, terminator: String = default)

Everything that requires something more complex than that is overuse.


(Phil J. Łaszkowicz) #26

I’ve used a variety of logging solutions across platforms, including .NET, Java, Python, C++, Rust, etc.

I’m in favor of combining two approaches into one API.

The default logging straight to stdout (print, printf, System.out.println, Console.log, etc) is a great tool for most to start with. So basing the API on extending that makes sense (I do this with NodeJS).

I wrote a logging lib because none of the existing solutions did exactly what I wanted or made it worthwhile to extend the existing codebase (sorry all).

So as far as extending goes (and I know LoggerAPI tried this), I think a base protocol to extend and implement is good.

I prefer Apache’s approach because there are multiple outputs, formats, and levels to configure in complex systems. Depending on a service to only output to stdout ignores the real complexities with enterprise logging.

If I write:

trace(“Beginning NATS background execution”)

I may want that to be printed as JSON to a web hook for a logging sidecar, as CSV for a file to be aggregated and picked up, and to the console when in DEBUG mode during development, whilst being ignored for console during production.

I wholeheartedly believe it’s inportant to keep these types of use cases in sight, whilst leaving a lightweight API-based solution inbuilt. Logging can be externalized easily enough with API-compliant projects.


#27

Or you could just begin the NATS background execution without logging it, I mean who cares?


(Phil J. Łaszkowicz) #28

I’m not going to argue a hypothetical logging message.


#30

This person is just straight out trolling. Do not feed the troll.


(John McCall) #31

I'm dealing with Vogel about their posts here; please continue with the discussion.


(Gopal Sharma) #32

I'm a out of my depth here, but I think lazy rendering would still be helpful in the context of sending logs to Splunk, etc.

As I understand it (and I may be totally wrong), Log4J version 1 wrote out log messages synchronously. We were writing structured log messages to a file, and a daemon would read those files and move them over to Splunk for indexing. Most of the time this worked just fine. However, we ran into a situation once where the logging slowed down so much, that our server would run out of threads to allocate to requests. We switched over to Log4J version 2, which I believe buffers log messages instead, and writes them out to file on a separate thread, and the problem went away.


(Gopal Sharma) #33

IMO, this is the way to go. Java does something similar as well. You have a factory to create loggers, so you'd have something like this at the beginning of most classes:

private static final Logger LOGGER = Logger.getLogger(Some.class);

The LOGGER object inherits all of its configuration from the global factory. The logger is named with Some (with the package name prefixed).

And at the log site something like this:

LOGGER.warn("some message");

(Adrian Zubarev) #34

I think the general idea behind this thread is great. I wonder if it would make sense to polish the stdlib part in that section and then start building the logging API on top of it. It would be great if the steaming API could be extended or overhauled so that we can inject listeners to destination streams at link time. Quite some time ago I started a thread in that direction.


(Helge Heß) #35

More general back-pressure aware streaming would be very useful before starting to write libraries, which then have to do all that on their own in some way. That "logging thread" is nothing but a custom solution here (and kinda a step back wrt async ...).

The console implementation in Noze.io is a little outdated and not super nice (and I don't want to bring it up as the "solution"), but it actually uses async streams to do the logging. I.e. the logging doesn't block the event loop, it rather schedules the log writes on the same and makes use of the same async underpinnings.

In the context FileChannels and PipeChannels would be required / nice to have.


(Gopal Sharma) #36

None of the above implementations seem to account for structured logging. This has been incredibly useful to me in the past.

Python's structlog has a handy page that explains why: http://www.structlog.org/en/stable/why.html.

Java achieves something similar with MDC.

IMO, it'd be great to include this as part of the base logging abstraction.

A problem that neither Python nor Java solve well (AFAIK) is the ability to "extract" structured log data from exceptions. I rarely want to log when I throw an exception, but I would ideally like to have the exception include structured log data that gets written when the exception is actually logged (likely in some error handling middle ware). I don't really have a good solution though...

What I've done in the past is to write exception message strings in some predictable format (like say key1="value1", key2="value2") that gets parsed when logging, but that's ugly and slow.


(Johannes Weiss) #37

Do you have good documentation links on what form of lazy rendering Splunk supports when ingesting data? I'm still unconvinced that the benefit is large enough to add that extra complexity given the variety of log ingesting solutions people use. But I'm very happy to be convinced otherwise if it's a thing that most solutions support and is shown to have actual benefit.

Yes, all synchronous logging from the calling thread is probably not a good idea. I'd also recommend to use a 'large enough' buffer. The 'running out of threads' issue is a different issue though, right?


(Helge Heß) #38

I think the Apache example is essentially "structured logging", just not a general purpose one. (and I agree with you, it makes sense to think about this. logging doesn't have to be reduced to string lines).


(Gopal Sharma) #39

AFAIK, Splunk itself doesn't support it. I'm likely not being very clear here, or I may be misunderstanding what "lazy rendering" means in this context. Let me try again.

In the setup we used, a log like this:

logger.info("Response sent");

would get written to a file like this:

[start event]
timestamp=<some timestamp>
thread=<some thread>
level=INFO 
logger=<some logger>
clientIP=Unknown
method=POST
msDB=79
msRequest=134
numDB=1
path=<some path>
queryParams=
requestId=88472589-16e0-4240-bba3-1c7073d49cca
statusCode=200
userAgent=python-requests/2.18.1
message=Response sent
[end event]

A separate process parses this file, and pushes data into Splunk's indexers (I don't know how exactly this worked since I didn't write it).

IMO it could be useful for the conversion from "Response sent" to the large String that actually gets logged to not happen on thread that calls logger.info?

Makes sense. FWIW, in the event of the buffer running out, my preference would be to block new logs, rather than to drop older logs.

Yes - I only mentioned it because it was a consequence of synchronous logging.


(Johannes Weiss) #40

Agreed! Structured logging would be great to have but it's somewhat orthogonal to the question of how we reach the logger itself (local v. global). Not sure if we should discuss both axes at the same time or not.


(Helge Heß) #41

Maybe how you "reach it" should actually be framework specific? General purpose libraries would probably need to get it injected. (and this project would essentially just provide the protocol/API?)