Fix Bugs Faster using Activity Tracing

Session 714 WWDC 2014

Finding and diagnosing application bugs can be very difficult. Activity Tracing is a new technology that can dramatically improve your speed and efficiency finding bugs in your code. Log trace messages to an auto-generated ring buffer while associating them with the originating user action. Reduce the time sifting through unrelated log messages trying to understand what was going on when the failure happened.

Good morning, how's everybody doing?

Enjoying the conference, I hope.

My name's Eric Clements and I'm here to talk to you about a new technology we're introducing for iOS 8 and Mac OS X Yosemite called activity tracing.

The goal of this is to hopefully improve your ability to fix bugs in a much quicker fashion.

So, I'm going to cover a few topics today.

First of all, we'll go over a little bit of a background and why we developed this and what our goals were.

We'll talk about some new concepts, including things like activities, breadcrumbs, new concept of trace messages, as well as some tools and some considerations to take into account while adopting this new technology.

So, let's go into some background.

As many of you know, here at Apple we encourage asynchronous development very much so because it's very important to have a very responsive user interface and application.

And in doing so, we've introduced a lot of technology over the years, including XPC, GCD, NSOperationsQueue, to simplify those common tasks that were previously complex.

In doing so we realized that this became very difficult to debug.

How many of you have had problems trying to debug your XPC and-yeah.

So, we wanted to really find a way to improve that with this release and we've done a lot.

Not only in XCode with the new Queue debugging, but also with this new activity tracing technology.

And you'll hear more about this as we go along.

In addition to that, privilege separation has, in addition, has made that much worse because we have another process involved trying to separate privilege and now you have two processes involved that may log separately and you have no way to understand who's causing the problem.

So, this is just another piece of the puzzle that has complicated our lives.

The reality is debug tools have not kept pace.

We've done our best.

You'll see in this release, we've done a whole lot, as I mentioned, Queue debugging in a session that has happened as well.

The reality is, a lot of us log, and I'm sure and all of you in the room have used this log or NSLog or some logging mechanism to try to debug your situation.

What we really found that that's really not sufficient these days because there's really not enough context to understand where you were and how you got there and we really want to find a way to resolve that and get you that information you really need to diagnose the problem.

One of the diagnostic challenges we find is as you see going forward, you know, when you sample a Daemon or an XPCService, it may look very idle, looks like nothing's really happening, but the reality is, there may be two processes or multi-processes talking to this service that are blocked, but they also look idle, so how do you know that that's actually the situation?

We really wanted to find a way to improve this.

In addition to that, I'm sure, if you're like me, you probably have a desktop that has multiple windows open, including console, terminal, XCode, your application and any numerous number of applications, trying to understand what's happening at that moment in time.

We really wanted to find a better way to do this, so we think we developed something.

In addition to that, the complexity of the OS has really gone up over the years.

If you look at a simple query, in this case the touch that opendirectoryd, that query may actually talk to multiple daemons in a circular fashion.

And how do you really understand what's happening at that moment in time because you never know which Daemon is actually in flight, what they're doing and how you got there.

In addition to all this, we also have what I call a complex crash, where over time as an application is being used, it may invoke multiple Daemons or XPCServices and ultimately, at the end, you have the final event, as you see at the top, button click that talked to an XPCService, well, that XPCService happened to crash.

You may not have noticed it in the background and all that shows at the top of the screen is your query cannot be completed at this time.

Not very helpful.

So, how do we find a way to improve that?

So, we set out with several goals, we really wanted to make the additional data that we're providing in this release, available in a diagnostics report you see on a daily basis, whether that would be a Crash Report, a spin report or any other type of report that we provide.

In addition to that, we really wanted to reduce the time you spend guessing what, where and when.

How did I get here?

What was happening at that moment in time and why?

You know, that's really the biggest challenge in this complex environment and we really wanted to find a way to provide you that insight.

In addition to that, sometimes we look at problems as a single click or a single button that caused the problem.

It's not necessarily a single action that caused the problem, sometimes it's the interaction of actions over time.

Maybe there's a button click that created a new mail, deleted the trash and then tried to create a new mail.

Take the pick of your application, it's sometimes not just a single event, it's actually the combination of events that caused the problem.

So, how do we understand those types of events over time?

Whatever we did, we knew it had to be lightweight and really easy to use.

So, let's look at a Crash Report.

You can see a little hint here, we're looking to add some new additional information, we're obviously in a Crash Report, maybe at the top we'll actually have the events leading up to the crash.

Down below the crashing thread, some detailed messages about what was happening at that moment in time when the crash happened.

What's really exciting is to point out some performance gains here by adopting this new activity trace mechanism, you get some power back and CPU back to your application and focus on the actual work, versus trying to understand what was happening.

You see measured across various platforms, we've got really nice numbers.

Similarly on iOS, dramatic difference, gives you that CPU and power back to your application focused on that versus trying to understand what was happening.

So, to cover some terminology again and we put these boxes into place.

Those events I mentioned over time, those menu clicks and such, are what we call activities.

An activity flows from one process to another and this is kind of our scoping mechanism, so to speak.

On top of that we have a concept of breadcrumbs, and these are those high-level interesting or meaningful events that you as a developer understand because not every click or every menu option is interesting or meaningful, you get the opportunity to actually point out the ones that are meaningful to you as a developer.

In addition to that, you have trace messages that are actually part of this activity, so that you actually have a scoped view of that action and things over time versus a huge log to try and sift through to understand what was happening at that moment in time.

And we'll show you some great examples of what this does for you.

We do have a new Daemon, nothing you typically have to worry about, but it does come up in the background and it will-it's the one that handles all this behind the scenes and provides a diagnostic report mechanism.

So, let's talk a little bit about activities.

I alluded to an activity of what it was, but what really is an activity?

You think of an activity as a set of work that was triggered by an action.

As the diagram I showed before, an application could have done a simple query.

That query talked to another Daemon, which talked to other Daemons along the way.

That is a set of work that is triggered by an action.

But I totally understand that there are cases where you have explicit work, maybe you have an initialization phase, you're reindexing databases, you're building a new asset database, various reasons that you want to xScope that, where you can actually create that explicitly defined sort of work with an activity.

But what does that activity really mean to the system?

What I mean is, it's essentially a 64-bit identifier that's created by the operating system.

Something you don't have any control over, it's automatic as part of the API.

What's great about it is, it's also autopropagated, meaning it flows throughout the system across GCD queues, NSOperationQueues, as well as processes.

For example, take process A and process B, may have queues in flight and you can see here those-activity ID number 1 and activity ID number 2-you can see them flowing across the process and that happens to be XPC and GCD, all behind the scenes, nothing for you to do, it's all transparent.

So, you look at activities over time as I've showed previously in a previous chart, now you have a concept of a menu selection, has a single activity ID and then the next menu select has a new activity ID.

And this is happening every time a button is pressed on the screen or a click is done.

So, now you have a new scoping mechanism that tracks that activity across the system, very simple concept, automatic, nothing for you to track directly.

Well, we obviously wanted to make a way for you to create these activities and we wanted to make sure it was easy for you so you had nothing to do directly unless you wanted to.

So, we made it as simple as possible, we provided this automatically through the UIKit on iOS and the AppKit on OS X.

So, as part of every action on the screen, an activity is already created on your behalf, nothing for you to do, you just do your normal coding as you do every day.

And we obviously have an API to allow you to create activities explicitly for those cases where you may be startup phase, that database that I was talking about earlier, to really scope that information to the action that you are trying to track.

We have a concept of detached activities because now that we have this automatically created activity, how do you disassociate things?

Take, for example, an API callback comes in to your query and you realize the database is out of sync.

How do you disassociate that resync or rebuild the database from the action?

Because those are two separate activities that you want to track individually.

It wasn't the fact that the person did the query that caused the rebuild, so you want to track those separately and understand why the rebuild failed versus why the action failed.

So, here's a quick diagram of what detached activity looks like.

You can see activity ID number 1, went over across another queue and then you had an access you wanted to create, it will actually create a separate activity ID on that next thread on whatever work you're trying to initiate.

So, now you have two completely individual activities to scope this information into your diagnostic reports.

So, how do you initiate an activity?

So, I'm going to try to make this a very easy API to use.

OS activity, activity name, put some flags and a block.

One very important fact you'll see throughout this entire presentation, all these strings must be constant.

We do not allow you to pass arbitrary strings as part of these names.

This is for various reasons, including security, performance and privacy.

And obviously you saw the performance difference.

This is part of the reasons why we get this performance gain.

We provide a flag option so you can control the behavior.

This is when you would say I want to detach and activity or some other options provided in the future.

We do provide a non-block variant of course.

I didn't lay them all out today, but you can look at the headers and you can find the header, #include os/activity.h. So, let's talk about those automatic generated activities.

Take a process that has the search query about to initiate, type some characters, hit return, AppKit's going to actually do an OS activity start for you on your behalf behind the scenes.

It'll come up with a name, button pressed, control, send action.

And then your actual callback will be called just as it is today.

And then upon return, it's going to actually stop that scope because it realizes that you've returned from the function, so the scope of this activity is kind of returned, but that doesn't mean the activity itself has stopped because, as I mentioned earlier, these activities are flowing across the work that is occurring.

So, the work may still be in flight, but we've returned control back to the UI and this is just kind of ending that scope.

As you saw previously, to create an activity is pretty simple.

In the case here, I got a callback doing a checkCache OS activity initiate, flag default and the block, very simple.

Similarly, the detached activity-you have a searchField and you realize, as I mentioned earlier, maybe the database is out of sync, so therefore, you want to reinitiate a synchronize, or rebuild.

You create a new activity and say detach it from the current activity because it's really not the same operation.

I'm searching for data and I'm going to rebuild the database asynchronously, so you really want to separate them.

So, that's activities in a nutshell.

Most of the time you don't have to worry about it, they're there and if you need to create some explicitly, you have the option to do so.

So, let's talk about briefly about breadcrumbs.

Breadcrumbs are essentially a way to label those meaningful activities to you, as the developer.

As I mentioned, you understand your code, you understand those things that your application does because now that we have the automatic activities being created from the UI, they all look the same, you know, you have the Send action, you have a keypress.

Those aren't very interesting or meaningful to you.

You actually want to hone in on those things that are occurring like that composed mail, empty trash, whatever application you may have.

This gives us the opportunity to see those interactions at a microlevel.

So, as I mentioned before, you know, advanced essentially isn't a particular event that causes a problem, it's a combination of events that cause a problem.

So, this gives you the opportunity to see those over time.

You'll see this throughout, when you see the button clicks, now you can put an actual name to it.

So, instead of this just being a menu selection, it's actually, compose an e-mail.

Instead of being a button click, it's a send e-mail.

Go back to our original diagram with the activities with activities ID's in flight, same thing.

Now you have an activity ID with a name as well as a breadcrumb of those interesting meaningful activities over time.

One thing to note about breadcrumbs, they're a separate kind of ring buffer of their own.

We track the last 50 events over time.

So, you can look back in history and see the last 50 actions that occurred to see what was happening when an application crashed or had an issue.

So, how to do you add a breadcrumb?

Just like the previous API, very simple, OSactivitysetbreadcrumb and give it a string.

This is something you define and you control.

One key note is, this is only supported in the main process because it's an application concept at a high level, you really can't do this from plugins or libraries.

As with the previous API, these must be constant strings-for both security, privacy and performance.

Excuse me.

And as you saw previously, same include, #include os/activity.h. This here is in string example because many of its combined callbacks for the same routine, but in this case, I have a search which actually has an offline database.

I can actually have two different code paths that have two different profiles.

I can now highlight them separately and say, okay, while search was initiated, it was a button click or as a menu selection and this path actually, was searching the offline cache, which was has a different characteristic and different problems.

Whereas, if I'm online actually talking to a live database, I'll set a different breadcrumb, because then you see that microinteraction over time.

You understand what the person was doing at that moment in time.

And that's breadcrumbs in a nutshell, simple concept, microlevel interactions over time.

Now, this is the most important part and I think you're going to find this the most exciting.

Trace messages: We provide a new API to allow you to add trace messages to that activity that was created.

Uses a standard printf file, printf format that you're used to on a daily basis.

It's very lightweight, as you saw from the performance numbers.

What's very interesting about this is we understand that there's two modes of operation.

There's a development mode, where you're debugging and working on an application and there's the release mode, the part that's shipping on our shipping platform, whether that be an iOS device or a desktop.

So, we really wanted to take that into account when we developed this because it's a very different development cycle.

When there's a problem on a production system, you have less information available, but when you're actually trying to live debug a system, you may want to have more information; those debug messages, additional detail to understand what was happening.

All of this is stored in an in-memory ring buffer.

So, this is where we get our huge performance gains.

What's very unique about this is, unlike traditional logging, we actually craft through the PC of the caller.

We have the timestamp, we have the data that was provided as well as the sting offset and some additional data, including even the thread ID that was running at the moment that this was created.

As you can see offside here, we actually have, the entire entry for trace buffer is 64 bytes.

No matter how long your string is, how long your printf format string is, it's always 64 bytes.

So, what does a trace message look like?

As I said, looks just like printf.

In fact, if you use printf, you almost just do a search and replace.

One key difference and this is where privacy, security come into play.

This API will only take scalar formats.

So, you can only provide ints, longs, doubles and such.

And you'll see the benefits and we have some other things that are available, too, as well.

Part of this, of course, strings and characters are a privacy problem, they're a security problem.

You know, accidentally logging a password or a username.

Those are inappropriate to do and this is a safeguard, but most importantly, it gives us the performance that we cannot gain with arbitrary strings.

There's no way to copy an arbitrary string efficiently and not impact the performance of an application.

There's another header to include, #include os/trace.h. Now, this is the release trace message that we were talking about two loads earlier.

This is your release message.

We also have a debug trace message.

This is when you actually live debugging application.

Just like the release one, under debug is the only difference, very simple, very straightforward.

What's great about it is the debug messages are only saved in the ring buffer if you're in a debug mode.

So, a typical running application on a shipping device will not record these into the ring buffer, saving space for those more interesting messages that you care about for the application.

Additionally, in debug mode dash, it takes more resources.

Obviously, in this case, we actually increase the buffer size.

Well, we're also recording additional messages, so we want to, we may take more resources in this mode so we really don't want that on a production system.

But we obviously want you to have that while you're trying to debug your application.

One thing to note, there is a way to enable what we call the debug mode at launch.

All you do is set an environment variable called OSActivityModeDebug.

And then from there on, every tool application you run from that environment variable will run in debug mode.

We have an additional type of trace message called the payload.

Now, this one's really unique because as I mentioned earlier, we don't support strings.

So, we want you to be able to debug your application and such and have additional data that we can't provide you in the Crash Reports.

As you can see, there's one slight difference with this API, it actually takes a block as the last parameter.

It passes an XPC dictionary and you can populate the dictionary with any data you'd like, anything arbitrary can be output to this.

In this case, I'm setting a string of the interface string that I'm talking through.

What's great about this is the block is only called when it's requested.

So, in a typical running application, this block is skipped, but if I'm using one of the tools, and I say I want to see the block because I want to understand what's going on, you'll get these messages, just like you would have otherwise with all the other messages.

The great part is, now you get that additional data.

I'll give you a great example of this.

One thing to note, this does use XPC to transmit this, so this is extra overhead, when this mode is enabled because there's no simple way to process this data, that's why it's in XPC dictionary, but we do send any XPC message across your diagnosticd.

Just like I showed with the debug mode, there's also an environment variable, which you can set to enable stream mode at launch because sometimes you want to understand what was happening while the application launched, you're going to set the similar, OS ACTIVITY MODE=stream.

Now, this will stream all the events from that application to diagnosticd and to any listeners.

So, how does this fit into the picture?

You saw [inaudible] obviously the trace messages fit into that activity.

As I mentioned earlier, there's a ring buffer.

There's an individual ring buffer per process, per activity.

So, unlike the typical sys log and the other cases where everything goes into one big bucket, you have an individual ring buffer for that activity, in each application binary so that one cannot step on the other.

So, now you can get a very different view of your application, whether it's your own application or other Daemons involved.

So, let's show you what we did for you in Crash Reports.

First of all, up top we have our breadcrumb trail.

So, these are the events that were occurring up until the point that the application crashed.

And notice the last event listed up top or the first event listed at the top, has an activity ID associated with it.

This is the activity that was in flight while it crashed.

Then down below the thread, you actually have the activity information.

In this case you have an ID.

You have the name that it was created, in this case by UIkit or AppKit called sendAction.

You have the breadcrumb.

You have how long it's been running and any failure that we may have detected.

Our long-term goal is to be able to detect some situations and inform you of that because we can tell that from the information.

But, most importantly, down below, you have all the trace messages leading up to the crash.

And I'd like to point out here...

[ Applause ]

As you can see, it's not just the application that I wrote, Query Directory, it's also the Daemon behind the scenes.

There was nothing you had to do.

It was automatically provided and it's in the Crash Report for you to look at.

This could be any number of Daemon's involved, any number of processes, but you get that entire view of why your application crashed for the first time directly in a Crash Report.

So, how does this exactly work?

In a crashing process, as I showed earlier, you know, the XPCService may have failed while diagnosticd actually comes up when the process fails, the buffers from the individual processes that are running, are sent to diagnosticd along with the activities that were in flight and the breadcrumbs.

And this is how we get all this information into the diagnostic report, all automatic, nothing for you to do.

So, that's great for all the general cases, you got trace messages, you got lots of detail, but how do you report those unexpected errors?

You know, you're trying to debug a problem or you want to highlight the fact that an error occurred, you can also do that with another API called OS trace error.

It allows you to pass any amount of data, like you saw in the previous APIs...

[ Applause ]

All right, okay.

So, as mentioned, we have OS trace error.

We have OS trace error, it allows you to provide similar format string and whatever data you want to output.

The only difference about this is the fact that it actually signals that an error occurred.

And what does that do?

It sends those trace messages at that time to diagnosticd to evaluate and store, so you can retrieve this via tools or via any other mechanism we provide.

That's what you consider a soft error, but what about those fatal errors, you know, I'm about to crash, I got into a state I don't know how I got here.

We will provide an API for that as well, cut, OS trace fault.

This may be something you can call just before you're about to crash.

It could be something that you'd call to avoid a crash.

As I mentioned, it can be an impending catastrophic failure.

This is a really bad situation, I shouldn't be in this state, I want to gather this information.

The difference between this and the error is all the buffers from all the processes involved are sent, just like you saw in the crash, without it actually crashing.

To give you a view of this, just like you saw in the crash mode, the difference is the application stays up or the Daemon stays up.

The buffers are sent and accumulated by diagnosticd.

Simple API, you see in this example, we have an application main.

It's starting up some initialization phase, we create activity, the initialization fails.

And this is bad, so we initiate a fault and then exit.

In a typical situation, which is really bad, you crash because that's your only way to get information about what failed.

Now you have a new mechanism without actually crashing.

And I'll show you what this gets you.

With all this said, we should talk about what's supported as far as trace messages.

For example, this is a perfectly valid good trace message.

Similarly, I can pass other types whether they be longs or shorts and such.

You can see various forms.

You can even pass a pointer because sometimes a pointer to us is a correlation point and, you know, maybe it's a connection you're trying to follow and understand that's being shared across multiple activities.

You can even output floats.

In this case, I can output how many clients I had and the average time they were connected to that.

And that's additional data you didn't have in your application at the time it crashed.

So, let's talk about unsupported trace messages.

As I mentioned earlier, strings are not supported.

So, well, how do we protect that?

Well, you can try to trace one, but you're going to get unsupported string in the output, so it doesn't break anything, so if you actually do it, it's not going to hurt you, but you won't ever see the data out of it.

Now, it's partially because of the way the mechanism is built.

There's basically no way to provide a string because we don't copy data out of the binary.

But, providing a string doesn't break valid data-in this case, I can still pass the UID and you can see the password will not be admitted, you get unsupported string.

So, you can mix and match, it doesn't hurt you.

Just be aware that it's not supported.

Along the same lines, characters are not supported because we don't want to be accidentally outputting passwords with characters.

So, obviously we have limits.

Format string cannot exceed 100 characters.

That's a pretty long format string, so hopefully that's more than enough for you.

One thing to note, the trace message content will truncate, so if you put too much data or too long a string, it's going to truncate it in the output.

This is expected behavior.

Initially we're only providing you seven parameters, so you can output up to seven scalars.

One very important note, these trace messages and the buffers may vary by platform and mode.

As I mentioned earlier in your debug mode, we may provide a much larger ring buffer.

Whereas, when you're not in a debug mode, it's much smaller, or maybe on a particular iOS device, where we're low on memory, so we don't enable it at all.

Just be aware there are limits.

We can't enable it everywhere and it may vary by that by platform.

So, let's do a quick demo.

So, I have an application running here, Query Directory.

I'm just going to search for something that I know is in admin.

Okay, you'll see down here, there's the result.

Nothing bad happened.

Try it again and I'll do a NSOperatationQueue mode instead, still no failure.

Now I'm going to type something I know that will cause a crash.

Now, you see in the Crash Report, you have that breadcrumb trail that I was talking about.

You see the last operation queue that was in flight, the activity ID and if I scroll down to the crashing thread, you can see this was the NSOperationQueue that was in flight, but now you've got all this new info that you didn't have before, the activity and all the trace messages leading up to the event.

In this case, you can see query directory and opendirectoryd involved.

Exciting?

[ Applause ]

I know I'm excited because this has always been a problem for me.

That's just a quick demo, so you can see it for real.

So, it's great, but what about debugger support?

Well, it's great for having the Crash Reports and the string reports, but when you're really working on a problem, how do you see this live?

Well, we made it built-in to the debugger.

So you can actually see the activity that's on a thread, as well as any trace messages for that activity.

This is a very powerful mechanism.

In the case of lldb you can see here, I have a thread that was crashing, I can type thread info.

As part of that thread info, I have the activity name that was in flight, how many messages it accumulated and down below it you have all that detail in the debugger.

So, instead of switching between console and terminal, you can just stay in your debugger, output those messages and say, oh, this is how I got here and why did this happen?

Click it here so you can see that activity, the current breadcrumb and those trace messages.

So, let's demo that because that's the fun part.

And if the demo gods are good to me or I'm good to them, we'll have a demo that works.

So, let's do the same as I did before.

I'll search and you can see the result.

I'll do an OperationQueue, and let's crash.

Now, in the same window I can say thread info and, of course, it didn't work.

[ Background Sounds ]

So, there's that same info.

[ Applause ]

[ Background Sounds ]

So, you can try this with various combinations, obviously this works with all of our infrastructure, which is really exciting because it works with XPC, GCD and NSOperationQueue and obviously the tools.

So, we talked about what we have, let's talk about a new tool we provided, ostraceutil.

This will allow you to watch that live streaming I was talking about earlier, right?

All that XPC dictionary and additional data that you didn't have in the Crash Report.

You'll be able to watch for faults and errors across the system or by process.

So, while you're debugging the application, you can actually say ostraceutil watch fault application.

And you'll see the faults that your application are generating without having to enable any logs and having to do any extra effort.

In addition to that, you can actually look what kind of activities and trace messages the binary has, whether it be your binary to check it before it goes out into production, or to look at what other processes are providing in this mechanism.

So, you see a complete view of your development cycle.

This is a look at the live stream.

In this case, ostraceutilwatch Query Directory, very simple.

You'll notice down here in the middle, you actually have that payload I was talking about.

So, even though we don't support strings and characters, you can put whatever you like here.

It could be a binary block, it could be strings, data.

You still get it so you can still do your job and debug your application.

And in this case, you can see the case I've typed to crash the app every time is Z.

I got the Z as part of the query string.

Similarly, with faults.

This is kind of fun because I can actually watch faults across the entire system.

Inthis case, ostraceutilwatchfaults.

And you can see the Query Directory triggered a fault because something bad happened.

You get the Image uuid, the PID and the path.

You'll see the breadcrumbs and the timestamps with those breadcrumbs.

You'll see that same activity information as well as the trace messages.

So, for the first time without enabling any additional logging, you can see this data while you're working on your application.

So, what are some things to take into consideration?

Privacy-we're very concerned about privacy.

So, be very conscious about what you're tracing when you're adding trace messages.

Never trace information about a user device.

There's a session at 2:00.

If you don't get a chance to see it, please take a look at it, there's a privacy session on both OS X and iOS.

I think we have one every year and there's one again today at 2:00.

One thing I'd like to note is we've done our best to try to do what we call with propagation failures.

Because this mechanism is automatic, there are cases where a custom queuing mechanism is built into an application.

And that little connection is missing.

So, even though the work happened, we don't have the idea of the trace message or the activity.

And what ends up happening is you get two-thirds of the picture.

We've done our best to cover the bases and get these where we can.

Just be aware that, you know, we can't catch them all.

As you see from the chart, this is supported in our diagnostic tools, OS X Yosemite supports in Crash Reports, Spin Reports, Stackshots and lldb.

Please note that the lldb support that I showed you onstage is not active in your seat, but look forward to it in a future seat for Yosemite, but it is active for iOS 8.

So, you can actually look at your Crash Reports and lldb to get this information.

So, in summary, why use activity tracing?

Well, I think there is a lot of good reasons, as you may tell from the presentation.

It's already included in your diagnostic reports.

No additional work.

No more enable logging.

No more turn on this log or this messaging.

It's all part of your diagnostic report right out of the gate.

We have built-in debugger support, so you don't have to leave your debugger and figure out what was happening.

It works on both iOS and OS X, extremely fast and efficient, integrated with GCD, XPC, NSOperationQueue and Foundation.

And as I alluded to, it captures additional data that traditional logging doesn't-the caller of the API, the thread it was running on and other additional information.

But most exciting to me, is these trace messages are actually scoped to the action that caused the problem.

So, instead of sifting through tons and tons of logs, through different processes to understand what was happening, you finally have the information right in front of you at the time it happened.

For more information, please contact Paul Danbold our Evangelist.

We don't have official documentation, as you can see this is a very simple API, but there are headers and manned pages for them, please feel free to take a look at them and please contact the developer forums.

Some related sessions-some of these have already occurred, but I do encourage you to go to the privacy session if you can at 2:00, which of course conflicts with the diagnostics performance in GCD and XPC, but both of those are very important.

I would suggest taking a look at them.

Thank you.

[ Applause ]

Apple, Inc. AAPL
1 Infinite Loop Cupertino CA 95014 US