Node: Inspecting Node.js with Chrome DevTools

[yury-s]

Objective

We’ve been thinking about providing unified debugger support across various v8 embedders. One natural approach to that is to reuse Chrome DevTools and provide its JS debugger, performance and memory profiler functionality to Node.js users. What we aim for Node is what we have for Chrome for Android. You basically go to _chrome://inspect_ on a stable version of Chrome and it debugs your running Node instance. You might need to override / specify ports to both if you want them to be non-default. The rest is behind the scenes.

Internal design

DevTools is comprised of the two major components: Frontend providing UI and Backend that instruments inspected application. Frontend is essentially a web application that communicates to the Backend via remote debugging protocol and nothing prevents it from connecting to Node.js instead of Chrome. To support that we’ll need to extract parts of debugger and profiler functionality from Blink into a separate project that would implement DevTools remote debugging protocol. Let’s call the project v8-inspector. Node.js could then expose the protocol to the Frontend over a WebSocket connection.

Proof of concept

I’ve created a prototype demonstrating this approach:

https://github.com/yury-s/v8-inspector - implementation of DevTools protocol for debugger, performance and memory profilers. I started with forking Blink and then removed dependencies that wouldn’t make sense in Node.js. There is definitely more work to be done but I believe this is good enough as a proof of concept.
https://github.com/yury-s/io.js/tree/remote-debugger - io.js clone that uses the project decribed above and ws module to allow using DevTools Frontend for debugging/profiling.

Local Node debugging scenario with custom ports would look like this:

1. Start Node with --remote-debugging-port=9222 command-line flag
2. Start Chrome with --remote-debugging-targets=localhost:9222 command-line flag
3. Navigate Chrome to chrome:inspect to see list of inspectable Node.js targets

Longer term

Long-term solution would be to have both Blink and Node.js use the same native v8-inspector library linked against v8. DevTools Frontend development would continue in Blink. For each version of v8-inspector there is a compatible version of DevTools Frontend served from the cloud. This is how it works with Android debugging at the moment: when the user opens DevTools for Chrome running on Android, the desktop browser loads a version of DevTools UI corresponding to the version of Chrome running on Android and forwards all traffic from the device to that Frontend.

Open questions

Before moving forward with this approach and investing into Blink refactorings necessary for extracting common code into its own library we’d like to collect some feedback on whether Node.js community would be interested in that. In particular, is there an interest in integrating such debugger library into Node.js core?

• This will require compiling Node with the v8-inspector library and providing Node.js-specific implementation of interfaces required by the debugger.
• Remote debugging protocol will be exposed to third party Node.js modules as a simple API similar to what is available for Chrome extensions (https://developer.chrome.com/extensions/debugger). Client of such API should run on a separate thread not to interfere with JavaScript in the inspected Node application. In the prototype all debugger events are forwarded from the main thread to a node instance running on a separate thread.
• Implementing transport for the protocol. It can be a module that would use the debug API described above and send protocol commands over WebSocket. E.g. in the prototype it is a simple web server based on ‘ws’ module: https://github.com/yury-s/io.js/blob/remote-debugger/lib/remote_debugging_server.js. Bundling this module with the core would enable zero-conf rich devtools experience (on par with Chrome’s DevTools).
• As the project evolves it will require adding new instrumentation to use some new features such as asynchronous call stacks. On the DevTools side we’ll be providing generic API that would work for both Blink and Node.js
• There is some maintenance cost related to the debugger if Node.js decides to ship with it but I don’t expect it to be much higher than for the current debugger that includes debugger-agent.cc. Ideally, v8-inspector would be included as part of the v8 checkout and the obsolete debugging protocol implemented by v8 would be removed, but that is not a done deal yet.

[ofrobots]

/cc @trevnorris, @bmeck

[bmeck]

/cc @thlorenz @3y3 @bajtos

[Qard]

/cc @nodejs/tracing

[thlorenz]

I did some related work a while ago: https://github.com/thlorenz/debugium

Implementing the protocol is the easy part. However trying to reuse code from chromium that communicates with v8 turned out to be non-trivial.
Things are quite coupled in there as I realized when I tried to unfuddle things.

For anyone working on this https://github.com/thlorenz/chromium-remote-debugging-proxy may come in useful as it allows watching the messages sent between chrome DevTools and the debugged instance which helps in understanding the protocol used.

[paulirish]

To add a few items…

This project brings all of the DevTools's JS features to Node. That includes:

• the sampling profiler with flame chart, top-down and bottom-up tree views
• memory heap snapshot capture and viewing
• memory allocation profiler
• JS debugger with breakpoints, step-out/in/over. Asynchronous stacks
• Experimental efforts around Step into Async, Promises view, etc.

Moving forward with this would make the current node-inspector codebase somewhat redundant, but we'd like to work with @bajtos, @3y3 and friends on making sure the experience of launching a debugging/profiling session with these tools is smooth and enjoyable.

@thlorenz

At one point I was told that the chrome team was going to take that on, but afaik that never materialized.

This ticket is the materialization of that work. ;) @yury-s is an engineer on Chrome and owns the JS debugging tools and how they interface with V8.

[thlorenz]

@paulirish Awesome! Happy this is finally happening :)

[trevnorris]

Does --remote-debugging-port=9222 bind to localhost or 0.0.0.0? Could it be made configurable? Much usage will come from debugging remote processes.

Will we have to integrate the same type of code into our JS, along with our C++, like in v8/src/promise.js:

  if (DEBUG_IS_ACTIVE) {
    %DebugPromiseEvent({ promise: promise, status: status, value: value });
  }

Having this w/o LTS support reduces its utility in node. As such it will become a maintenance burden across 4 or 5 active branches. Is it even possible that this API is integrated in a V8 release that's 1-2 years old?

Also in regards to LTS support, I assume that there isn't indefinite support of Chrome to debugger API support. Or am I wrong?

[yury-s]

Does --remote-debugging-port=9222 bind to localhost or 0.0.0.0? Could it be made configurable? Much usage will come from debugging remote processes.

It can be bound to any interface that we want, I don't see any issues with that (except security). Chrome already accepts both host and port as values of --remote-debugging-targets and the host may well be remote one.

It makes sense to differentiate between transport and the debugger backend implementing the remote debugging protocol. Same debugger implementation can be used with different types of connections similar to how it works in Chrome at the moment where various transports are used to transfer DevTools traffic depending on the inspected target:

1. Chrome IPC in case of Chrome embedded DevTools.
2. Web Socket when connecting to remote Chrome instance.
3. Adb when connecting to Chrome and WebView running on Android devices.

Will we have to integrate the same type of code into our JS, along with our C++, like in v8/src/promise.js:

If Node uses alternative implementation of Promises and wants to leverage debugging capabilities provided by v8-inspector then yes, you will need to provide corresponding instrumentation that way or another. v8-inspector is going to have C++ API so you'll need at least one binding to notify the debugger. The call will go through the v8-inspector C++ API not through some internal v8 API.

Is it even possible that this API is integrated in a V8 release that's 1-2 years old?

I'm afraid not, as it would require making v8-inspector work with such old releases of v8. This is a non-goal for the project. As I mentioned, the idea is to have particular version of v8-inspector compatible with particular version of v8 API. If v8 is updated, v8-inspector will likely need to be updated as well.

Also in regards to LTS support, I assume that there isn't indefinite support of Chrome to debugger API support. Or am I wrong?

At the moment, there is no official v8 debugger API which would be well-supported. There is a set of C++ methods in v8-debug.h as well as debug context which provides access to the debugger internals. This is one of the issues we'd like to address with v8-inspector. It should become a recommended way to debug v8. Eventually we'd like to drop most of v8/src/debug/debug.js and prohibit using the debug context directly.

Such approach would leave us the freedom of changing remote debugging protocol without committing to its backward compatibility as there will always be a version of Chrome DevTools compatible with given version of the backend. There are already two types of protocol commands: public and hidden. Public ones are those for which we commit to backward compatibility. We're hesitant on adding more public commands as it limits our development pace.

In terms of LTS, similar to v8 we don't want to commit to keeping the API stable as it would be too restrictive. What we can do is to follow the same public API compatibility as v8 does (or at least declares): https://code.google.com/p/v8-wiki/wiki/Source#V8_public_API_compatibility

Since Blink is going to use v8-inspector too, we'll have to keep it up to date with latest v8 API changes. Should we make any breaking changes they will be easy to detect in nightly builds and can be reverted. On our side we'll have to make sure that changes to v8-inspector API would be made in a non-breaking manner, follow the same public API deprecation policy as v8 does and the new APIs would work for Node too, not only for Chrome.

Will this work for you?

[3y3]

@thlorenz ,

Implementing the protocol is the easy part. However trying to reuse code from chromium that communicates with v8 turned out to be non-trivial.

At current iteration node-inspector tries to inject some parts of Chrome debugger API to app.
There exists some outdated experiments with InjectedScriptSource.js and DebuggerScript.js from blink source.

v8-debug has experimental method enableWebkitProtocol
(but work in progress)

Main problem of this experiments - compatibility:

1. Compatibility of js codebase - chrome debugger uses all latest js features implemented in v8 (generators, arrow functions, for of) (I say here about InjectedScriptSource.js and DebuggerScript.js)
2. Compatibility of features (c++ codebase) - async call stack, workers, network debugging etc. This is also described in issue open questions

As the project evolves it will require adding new instrumentation to use some new features such as asynchronous call stacks. On the DevTools side we’ll be providing generic API that would work for both Blink and Node.js

For this reason I also worried about @trevnorris questions:

Is it even possible that this API is integrated in a V8 release that's 1-2 years old?

@paulirish ,

Moving forward with this would make the current node-inspector codebase somewhat redundant

I will be happy, if node-inspector will change his main target from "supporting compatibility with DevTools" to something like "extended debugging environment for nodejs"

@yury-s , are you want to backport async call stack from chrome to node? And some other features?

I'd like to think about v8-inspector.h like about current v8-debug.h in v8. His provide main api to implement debugger into nodejs, but also provides entry points for external developers.
In context of v8-debug.h main entry point is getDebugContext.
So I'd like to see more api for external developers =)

[3y3]

/cc @auchenberg

[trevnorris]

@yury-s Thank you for the comprehensive response.

If Node uses alternative implementation of Promises and wants to leverage debugging capabilities provided by v8-inspector then yes

Our process.nextTick() emulates the micro task queue. This and our handling of timers would most likely require insertion.

Is it even possible that this API is integrated in a V8 release that's 1-2 years old?

I'm afraid not, as it would require making v8-inspector work with such old releases of v8. This is a non-goal for the project.

While I understand, it's also unfortunate. Our main stable release is 6 month. Taking into account V8's rapid release cycle I will make the assumption that the same API could break before that is over. If we can't guarantee support for at least the 6 months of latest stable then I'm afraid that it won't do us any good.

Beyond that, saying it would be available for the 6 months of stable but not for the 18 months of LTS would need to have serious discussion as to whether bringing it is still viable. For the time I'm ignoring the additional 12 months a release will be in maintenance. Which brings full support to 3 years.

I do hope we can work out these issues, as having this available would be very helpful.

[yury-s]

For this reason I also worried about @trevnorris questions:

Is it even possible that this API is integrated in a V8 release that's 1-2 years old?

No. See my reply above.

@yury-s , are you want to backport async call stack from chrome to node? And some other features?

As Paul Irish wrote above the project will come with all debugging features available in Chromium. Async call stacks is not an exception. There is a generic API resembling the one described in this MSDN article which needs to be called from appropriate places where we'd like to capture async call stacks. I can help with adding such instrumentation to Node but we'll need some person knowledgeable better than me to identify places where the hooks should be added.

I'd like to think about v8-inspector.h like about current v8-debug.h in v8. His provide main api to implement debugger into nodejs, but also provides entry points for external developers.
In context of v8-debug.h main entry point is getDebugContext.

This is a good way to think about v8-inspector. Ideally, we should be able to drop v8-debug.h in favor of v8-inspector. This would make the API more explicit and allow to provide better support for it.

[yury-s]

While I understand, it's also unfortunate. Our main stable release is 6 month. Taking into account V8's rapid release cycle I will make the assumption that the same API could break before that is over. If we can't guarantee support for at least the 6 months of latest stable then I'm afraid that it won't do us any good.

How does this work with v8 at the moment? There may be 4 or 5 major v8 releases in 6 months which means there may be a lot of breaking changes. Assuming that v8-inspector is shipped as part of v8 and its API is defined in /include/v8-inspector.h (which would be ideal case for us long-term) how would that work with Node?

I do hope we can work out these issues, as having this available would be very helpful.

I hope so too.

[trevnorris]

How does this work with v8 at the moment? There may be 4 or 5 major v8 releases in 6 months which means there may be a lot of breaking changes.

Once a major release branch is cut from master the V8 minor is no longer updated on that branch.

Assuming that v8-inspector is shipped as part of v8 and its API is defined in /include/v8-inspector.h (which would be ideal case for us long-term) how would that work with Node?

From what I've read, it wouldn't. It doesn't make sense for us to integrate an API that will break so quickly. It's an added disadvantage since Google is insistent that no one have older versions of Chrome installed. If we were allowed to do that then it might be possible to simply use the corresponding Chrome version with a given node release.

[benjamingr]

I think it is very important that the debugger ships with node as a standalone (rather than connected via chrome://somename for compatibility issues, I still want to be able to easily debug io.js 3.0 when Chrome gets 5 versions ahead. Something like an atom/electron wrapper over a specific version of Chromium would be ideal and it would also solve the compatibility issues.

I think it would also be very important that all the instrumentation hooks we need for async are there - and that integration when running non js modules in node works (reasonably) well. Of course, userland hooks might also be interesting (so people can signal to the debugger things they want it to ignore, for instance - C# has this and it's very useful).

Huge +1 for the idea by the way.

[auchenberg]

This is great! Glad to see this is moving forward.

I think it is very important that the debugger ships with node as a standalone (rather than connected via chrome://somename for compatibility issues, I still want to be able to easily debug io.js 3.0 when Chrome gets 5 versions ahead. Something like an atom/electron wrapper over a specific version of Chromium would be ideal and it would also solve the compatibility issues.

As I understand it, the debugger back-end is separate from the core via v8-inspector, and since it would implement the Chrome Remote Debugging Protocol over WebSocket and HTTP, the debugger back-end would be compatible with external front-ends like Chrome DevTools App (based upon Electron)

By having a runtime/protocol version we will be able to match the front-end with the back-end. Eventually this would need to use some kind of feature detection, but that's further out in the future, as I see it.

[Qard]

Yeah, the only way I see this working is shipping the UI as an electron app with each release, or ensuring backwards compatibility in the protocol long enough to support the LTS releases.

I think backwards compatibility for the dev tools should, in theory, be easier than V8 itself. It's only the protocol that is important here. I don't expect anyone should be touching the API outside of core, so node core doesn't need to worry about API compatibility.

[trevnorris]

@yury-s Meant to ask, how much do you think the actual debugging API that connect client to process would change? If that itself doesn't change much, or can be fixed more easily over time, then we'd be much closer to having a workable solution.

[targos]

@benjamingr @Qard from the OP:

For each version of v8-inspector there is a compatible version of DevTools Frontend served from the cloud. This is how it works with Android debugging at the moment: when the user opens DevTools for Chrome running on Android, the desktop browser loads a version of DevTools UI corresponding to the version of Chrome running on Android and forwards all traffic from the device to that Frontend.

Doesn't it mean that one could use the latest Chrome and still be able to debug any version of Node.js using an old API ?

[benjamingr]

@targos I'd much prefer it if it was possible to run the debugger without accessing code served from the cloud live. Node would have no control over which versions are supported and how - nor would it be able to easily edit the code (unless node starts serving it). I think it would be preferable to have a standalone app.

[trevnorris]

Node isn't going to control what tools are used to communicate with the debugger API regardless. I'm not sure how some companies would feel about needing an internet connection to debug processes on their local network, but I'd say that sounds like a promising solution.

[Qard]

Yeah, not sure how I feel about a service out on the net somewhere having
direct access to the VM my code is running in. I'd definitely prefer the
standalone app.
On Aug 26, 2015 12:22 AM, "Benjamin Gruenbaum" [email protected]
wrote:

@targos https://github.com/targos I'd much prefer it if it was possible
to run the debugger without accessing code served from the cloud live. Node
would have no control over which versions are supported and how - nor would
it be able to easily edit the code (unless node starts serving it). I think
it would be preferable to have a standalone app.

—
Reply to this email directly or view it on GitHub
https://github.com/nodejs/node/issues/2546#issuecomment-134874574.

[jkrems]

Would the scope include things like PageAgent.reload to restart a node process from inside the devtools or the NetworkAgent/ConsoleAgent/...? If not, then there's definitely still a lot of value in node-inspector & friends.

[Spy-Seth]

This feature could also lead to use the Firefox devtools with node thought the valence project :+1: :+1: :+1:

[yury-s]

@trevnorris :

@yury-s Meant to ask, how much do you think the actual debugging API that connect client to process would change? If that itself doesn't change much, or can be fixed more easily over time, then we'd be much closer to having a workable solution.

This one is stable. For basic debugging support it is as simple as this:

class FrontendChannel {
    virtual void sendToFrontend(const String& message) = 0;
}
class Backend {
    void connectFrontend(FrontendChannel* channel);
    void disconnectFrontend();
    void dispatchMessageFromFrontend(const String& message);
}

See https://github.com/yury-s/v8-inspector/blob/embedded-in-io/Source/v8inspector/V8Inspector.h#L33 and https://github.com/yury-s/v8-inspector/blob/embedded-in-io/Source/core/inspector/InspectorFrontendChannel.h

The messages are strings that are opaque to Node, all parsing and dispatching of the protocol messages is performed by v8-inspector itself.

Once a major release branch is cut from master the V8 minor is no longer updated on that branch.

At that point v8-inspector is also branched along with v8. Also we know exact revision of the front-end that is compatible with that v8-inspector and can use it for debugging this particular release of v8 (embedded in Node, Chrome or something else). At the moment v8 branching is consistent with Chrome so we should always have good front-end that is capable of debugging that v8 version. More details below.

@trevnorris

Assuming that v8-inspector is shipped as part of v8 and its API is defined in /include/v8-inspector.h (which would be ideal case for us long-term) how would that work with Node?

From what I've read, it wouldn't. It doesn't make sense for us to integrate an API that will break so quickly. It's an added disadvantage since Google is insistent that no one have older versions of Chrome installed. If we were allowed to do that then it might be possible to simply use the corresponding Chrome version with a given node release.

@benjamingr

I think it is very important that the debugger ships with node as a standalone (rather than connected via chrome://somename for compatibility issues, I still want to be able to easily debug io.js 3.0 when Chrome gets 5 versions ahead. Something like an atom/electron wrapper over a specific version of Chromium would be ideal and it would also solve the compatibility issues.

@targos

Doesn't it mean that one could use the latest Chrome and still be able to debug any version of Node.js using an old API ?

@targos exactly. We already allow debugging Chrome 38 that can be found on many Android devices by connecting to it with Chrome 46. The way it works is for each Blink revision we have a copy of the front-end sources served from the cloud. After that we ask inspected target which front-end revision should be used for debugging and load it from the cloud into devtools window. E.g. to connect to chrome 44.0.2403.133 we use this link https://chrome-devtools-frontend.appspot.com/serve_rev/@199588/inspector.html

In case of Android remote debugging we load the front-end into a browser window that has special devtools bindings providing connection to the browser on the device over USB. This API is backwards compatible so that we could load old front-ends.

In case of Node we don't even have to do this since we can connect via plain WebSocket . The front-end can be loaded as a normal page and remote target url can be passed as 'ws' param: https://chrome-devtools-frontend.appspot.com/serve_rev/@199588/inspector.html?ws=localhost:9222/node This will work except for some features that require more privileges than regular web page can have, e.g. access to local file system. Loading as a privileged DevTools front-end provides better user experience.

[yury-s]

@benjamingr :

I'd much prefer it if it was possible to run the debugger without accessing code served from the cloud live. Node would have no control over which versions are supported and how - nor would it be able to easily edit the code (unless node starts serving it). I think it would be preferable to have a standalone app.

The front-end code can be deployed in a different manner, e.g. it could be provided as a .deb package. I just described how it is implemented in Chrome at the moment and it has been working pretty well for providing zero-conf Android debugging experience across all platforms. Being able to see Node instances as another type of inspectable target in chrome:inspect looks natural to me but this doesn't have to be the only way to debug Node.

[yury-s]

@jkrems :

Would the scope include things like PageAgent.reload to restart a node process from inside the devtools or the NetworkAgent/ConsoleAgent/...?

At the moment it includes Debugger, Runtime, Profiler and HeapProfiler agents which make sense for all v8 embedders. Console could be added to the list later. On the other hand, PageAgent doesn't make much sense for Node so we'd rather keep it in Blink. If we need "reload" functionality, the method could be moved to e.g. to Runtime domain or Node can have its own domain "Node" providing functionality specific to Node.js The front-end can either support this in the core or we can have it as a third-party module. This is definitely not something for v.1 and deserves separate discussion.

[yury-s]

@auchenberg :

As I understand it, the debugger back-end is separate from the core via v8-inspector, and since it would implement the Chrome Remote Debugging Protocol over WebSocket and HTTP, the debugger back-end would be compatible with external front-ends like Chrome DevTools App (based upon Electron)

Correct. Other front-ends should be able to connect using the same transport. Note though that we don't want to commit to backward compatibility of the whole protocol. It would be up to the front-end authors to make sure that their front-end is compatible with given back-end. We only promise to provide compatible version of DevTools front-end.

By having a runtime/protocol version we will be able to match the front-end with the back-end. Eventually this would need to use some kind of feature detection, but that's further out in the future, as I see it.

Note that this version only relates to the subset of public commands in the protocol. We have been reluctant to making more commands public so the protocol version hasn't changed for a while. Majority of the commands in the protocol are non-public and they have been changing quite often. To ensure full compatibility with the protocol discovery page returns exact version (198849) of the front-end that should be used.

[yunong]

This looks great. I'd love to be able to get the ability to walk the JS objects on the heap that is currently provided by mdb and v8.so https://www.joyent.com/blog/debugging-enhancements-in-node-0-12

Specifically, to be able to walk an object reference all the way back to the root object would be incredibly helpful in debugging memory leaks.

e.g. given a particular object's address, find references to all other objects that refer to this object, similar to what findjsobjects -r does.

[trevnorris]

Summary points:

• While the integration API is unstable, that doesn't matter much. What needs a certain amount of stability is the debugger protocol. It seems the debugger/profiler API have stabilized for the most part, but are still prone to breakage. How this will be handled over the course of an LTS needs to be solidified.
• The protocol requires web sockets. Does this mean node will have to natively support them as well?
• Discovery still needs to be done on all the injection points the API will need into core code.

[mscdex]

Is it possible to expose enough bits to make it transport-agnostic? This would allow core to avoid having to bundle a particular module for this to work.

[yury-s]

Is it possible to expose enough bits to make it transport-agnostic?

See second item in the "Open questions" section above and this comment above. There is a clear separation between v8-inspector and the transport provider. In the prototype remote_debugging_server.js requires three methods on node_debugger:

    node_debugger.connectToInspectorBackend(channel);
    node_debugger.dispatchOnInspectorBackend(message);
    node_debugger.disconnectFromInspectorBackend();

And provides single callback method:

   channel.sendMessageToFrontend = function(message) {...}

On top of that there can be a third-party WebSocket server or module providing another transport.

[yury-s]

@trevnorris :

While the integration API is unstable, that doesn't matter much. What needs a certain amount of stability is the debugger protocol. It seems the debugger/profiler API have stabilized for the most part, but are still prone to breakage. How this will be handled over the course of an LTS needs to be solidified.

While the debugger/profiler protocol may have been stable for a while there is still a chance that we'll need to make some breaking changes to it to move the project forward. We tried to keep stable subset of the protocol and it proved to be hard. We don't want to go that route if possible. Why do you think that we need stability of the remote debugging protocol with the proposed solution?

[JacksonTian]

Hope to see an improved debugger.

[trevnorris]

@yury-s

While the debugger/profiler protocol may have been stable for a while there is still a chance that we'll need to make some breaking changes to it to move the project forward.

If we still have access to the debugging console that can interpret the incoming JSON, that should be enough. This would actually be preferred. So the UI that developers have been using doesn't change over the lifetime of the release. To solidify what has been mentioned earlier, will we be able to use, and have access to, the older debugging consoles for older APIs?

[mattdesl]

One natural approach to that is to reuse Chrome DevTools and provide its JS debugger, performance and memory profiler functionality to Node.js users.

Another approach worth mentioning is leveraging Electron to launch a DevTools window with node integration. See hihat for an example. It has its own limitations, but doesn't involve remote debugging and provides a couple benefits over node-inspector (e.g. a more complete DevTools UX).

Having better debugging built-in to Node would be amazing, though. :smile: It's one of those things you take for granted when you start to do a lot of frontend development.

[yury-s]

@trevnorris

If we still have access to the debugging console that can interpret the incoming JSON, that should be enough. This would actually be preferred. So the UI that developers have been using doesn't change over the lifetime of the release. To solidify what has been mentioned earlier, will we be able to use, and have access to, the older debugging consoles for older APIs?

This is exactly the what should happen if I understand correctly what you mean. Let's assume that v8 stable branch used in Chrome 44 was created from v8 v.4.4.63 and that Blink used in Chrome 44 was branched from 195420. Let's also assume that DevTools backend implementation in Blink is based on revision 1024 of v8-inspector. As long as Node is compiled with v8-inspector r1024 the users should be able to debug it with r195420 of the front-end. Any Chrome version >= 44 can be used to host the front-end (as I mentioned above the source of the front-end will be served from https://chrome-devtools-frontend.appspot.com/serve_rev/@195420/inspector.html) and will work transparently to the user, the front-end could also be provided as Electron app or something else.

[yury-s]

Another approach worth mentioning is leveraging Electron to launch a DevTools window with node integration. See hihat for an example. It has its own limitations, but doesn't involve remote debugging and provides a couple benefits over node-inspector (e.g. a more complete DevTools UX).

There are already other debugging solutions for Node. The main question I'd like to answer after this discussion is to whether it would make sense to invest into extracting base debugger/profiler implementation from DevTools into v8-inspector. The benefit of the single source debugger is that it would make DevTools features available to other embedders and would allow them to use all power of the existing DevTools front-end for v8 debugging. Node.js is the second major client of v8 after Chrome so I'd like to first of all evaluate this solution with Node.js community.

It requires a lot of refactorings in Blink and adds a fair amount of complexity to Chrome itself since the debugger will have to be shipped as a third-party library. Before doing that I'd like to be sure that Node.js will be interested in integrating with the v8-inspector and that there is no another approach that would provide desirable debugging experience in a less intrusive manner/adding less maintenance cost to the projects.

[trevnorris]

We'll discuss this in the next TSC meeting. About whether we'll commit to bringing it into core or not.

[bajtos]

The main question I'd like to answer after this discussion is to whether it would make sense to invest into extracting base debugger/profiler implementation from DevTools into v8-inspector. The benefit of the single source debugger is that it would make DevTools features available to other embedders and would allow them to use all power of the existing DevTools front-end for v8 debugging.

From my point of view as a Node Inspector maintainer, this would be awesome. So far, we were playing a catch-me-up game, trying to keep in sync with the upstream changes made in DevTools UI and the protocol. We were loosing, simply because Node Inspector is a side project and our manpower cannot match speed of the Blink team of several dedicated developers.

I also think it's a waste of everybody's time when every debugger (Node Inspector, WebStorm, Visual Studio, Visual Code, etc.) has to reimplement (and keep maintaining) the same DevTools->V8 protocol bridge.

Even if Node core decides to not bring in the v8-inspector, I think it still makes sense to extract base debugger/profiler implementation from Blink into a module that can be reused by Node.js debuggers.

[trevnorris]

@yury-s Want to double check. Can we implement segments of the API and leave out others (e.g. async stack traces) and have it still work?

[yury-s]

@trevnorris Yes. Async stack traces is optional. You can opt to not support them, the core debugging/profiling functionality should still be available. In fact aforementioned prototype has debugger API for tracking async calls but it is not used.

[trevnorris]

@yury-s Awesome. That's great to hear.

[benjamingr]

Note that we might need to consider detecting if async stack traces are enabled somehow and disable certain optimizations in order to get meaningful results (like trampolining async calls).

[trevnorris]

@yury-s The node TSC has agreed to give a best effort at implementing "v8-inspector" into core.

[Qard]

Just a random comment: it'd be fantastic if this could be embedded in an atom plugin to have IDE-like functionality for node dev work.

[faceyspacey]

+1

[stevemao]

Have you guys seen https://github.com/Jam3/devtool? (I can't search for keyword "devtool" here so maybe someone already mentioned it).

[petkaantonov]

Related https://github.com/nodejs/promises/issues/8

[ofrobots]

I wanted to update the folks on this thread about the current status of this effort.

First, to recap the goals, we want to bring all the current (and future) DevTools's JS features to Node.js and other embedders (including sampling profilers, heap profilers, debugger, async stacks, etc.). To this end we have been working on extracting the DevTools backend from blink, extricate all its dependencies on Blink, and get it integrated into V8 instead. This way, the DevTools backend and the remote debugging protocol becomes available to all V8 embedders. We are calling this project v8-inspector.

Present status

https://github.com/repenaxa/node/tree/v8_inspector is a proof of concept integration with Node.js that @repenaxa (DevTools tech lead) has been working on. It is actually pretty functional, please do try it out and provide feedback.

• Complete stepping and breakpoint debugging, blackboxing
• Sourcemap support for transpiled code
• LiveEdit: JavaScript hot-swap evaluation w/ V8
• Console evaluation with support for ES6 and custom object formatting
• Sampling JavaScript profiler with flamechart
• Heap snapshot debugger, heap allocation profiler

All of the dependencies on blink have been removed, except for a dependency on the C++ container library from blink (wtf). We are actively working on swapping out the uses of wtf with C++ STL. We have a hacky integration with Node.js (see this patch: https://github.com/repenaxa/node/commit/86939cb60b22066f4ee6e818ef879c3dbd6282aa), but it does work.

Timeline

At this point, v8-inspector is not likely to land in V8 5.0. Based on this it seems that this will miss the boat for the next Node.js LTS (October 2016). It would be unfortunate if a better development experience wasn't available to Node.js LTS users until October 2017 (the subsequent LTS).

Even though the window for getting this in into V8 5.0 is closing, we do still have a runway for Node.js 6.0. Here's what we propose: v8-inspector does not necessarily need to land into V8 for Node.js to pick it up. We could integrate v8-inspector directly into Node.js (v6.0). This would unblock the work on the Node.js integration w/ v8-inspector.

Remaining Node.js integration efforts

• [ ] The integration of the debug server into the Node.js message loop. We have a hacky proof of concept (patch: https://github.com/repenaxa/node/commit/86939cb60b22066f4ee6e818ef879c3dbd6282aa) which should not be used in production.
• [ ] Ergonomics: What command line options do we want to expose for the debug server (host, port). Should we have a option that enables the debug server upon receipt of a signal (how would this work on Windows)? How does this relate to the existing Node.js debug server that talks to the JSON based V8 debug API?
• [ ] Bring in websockets (ws) as an internal module to avoid exposing it to user space.
• [ ] Wire asynchronous callstack support to Node
• [ ] Network inspection integration (v2)

We are continuing to focus on getting extricating the wtf dependency from v8-inspector, and getting it landed into V8 (probably V8 5.1 or 5.2 time-frame), in the meanwhile we'd be happy if folks from the community wanted to collaborate on the above Node.js integration tasks.

We would appreciate comments on the idea of getting v8-inspector landed into Node.js 6.0 as a direct dependency. ( +cc @nodejs/ctc )

It would be great to make better debug and profiling experience available to Node.js developers for the upcoming v6.0 release.

[cjihrig]

Bring in websockets (ws) as an internal module to avoid exposing it to user space.

This was discussed a while ago. I don't remember what the final verdict was. Would we be taking ws wholesale in the deps folder, using it as the baseline for a core implementation, starting a new implementation from scratch, something else?

[rvagg]

@ofrobots:

Even though the window for getting this in into V8 5.0 is closing, we do still have a runway for Node.js 6.0.

My recollection of the outcome from https://github.com/nodejs/LTS/issues/62 was that there was support for being flexible with upgrading V8 in Node.js v6 _if_ we can do it in a way that doesn't break ABI for addons as this is the primary concern that has lead to our V8-locking policy in the first place. For instance, the 4.8 changes have turned out to be pretty minimal from an API perspective and if we had have had this policy in place for earlier then Node.js v5 may have had an upgrade from 4.7 to 4.8 at some point. I'm going to have to revisit the CTC meeting where we discussed this and write up some proper minutes for that discussion because Chris was taking notes and had to leave so our minutes doc is incomplete. I believe that we agreed that the overriding concern was ABI breakage and then on top of that stability prior to cutting LTS, so there would have to be a buffer period before October where we rule out upgrading even if it's possible simply to create stability and predictability for users as they start testing v6 for LTS adoption.

_So_, if the worldview that I have in my head is an accurate reflection of the general agreement then that would mean that V8 5.0 isn't necessarily the final word for Node.js v6 LTS, we could have a path to 5.1 and even maybe 5.2 during the v6 stable period. So that would suggest that it may not be quite as difficult to get some form of this into v6 LTS.

@cjihrig relevant discussion on websocket support is here: https://github.com/nodejs/node/blob/master/doc/tsc-meetings/2015-09-02.md#inspecting-nodejs-with-chrome-devtools-2546

That, and my recollection goes something like this: We're open to adding websockets support purely for supporting this and we were much more happy with the idea of doing it privately _just_ for this to stop us jumping on an API we may not be happy supporting in core although there is a nice path to exposing such an API if we really wanted to. wrt ws specifically, I don't believe we ever went there in the TSC/CTC level and assumed we'd be implementing our own. There is some existing discussion here: https://github.com/nodejs/node/pull/4842 leading to an EP here: https://github.com/nodejs/node-eps/pull/6. My reading of this is that there is very little enthusiasm for going this route, but perhaps if it gives us a kick-start on getting this problem solved we should be open to it—i.e. afaik there are no decisions here so it's all up for discussion.

[benjamingr]

I just want to bring up the option of exposing the hooks required for a userland websockets library (or http library) to perform the necessary instrumentation rather than core endorsing ws or another specific module.

In general - I'm personally much more in favor of exposing more hooks for the debugger and not endorsing large modules that have to be maintained and de-facto enforce an opinion in Node.

I'm not sure how feasible that is but I think it should be discussed.

[bnoordhuis]

@benjamingr There's no endorsement, 'ws' won't be exposed to user code. It'd be a private implementation detail that can be replaced anytime.

Hooks or overrides on the other hand means committing to an API and supporting it over longer periods of times, possibly indefinitely. It gains us nothing that I can see, it would just make maintenance more of a hassle.

[cjihrig]

Is this something that the CTC needs to revisit?

[benjamingr]

@bnoordhuis hooks would mean userland libraries could use the instrumentation to provide very nice debugging tools. For example - if an HTTP2 library or a library for another protocol (IRC or FTP for example) could use the network panel in Chrome to track requests in a way that provides developers insights it might be interesting. Node is not only browser protocols like HTTP and WS.

That said, those two working is a lot better than a solution in which nothing works.

[jkrems]

@benjamingr If I understand it correctly, it's not about instrumenting ws so it shows up in the network panel. The devtools use web sockets to communicate with the node process and for that purpose ws would be used. Making web sockets used by the debugged app show up in the network tab is out of scope for all of this afaict.

[ofrobots]

@rvagg: The discussion in nodejs/LTS#62 led to a _possibility_ of being able to upgrade V8 in Node.js v6 _if_ the ABI differences could be bridged. Still, I think it is a good idea to bring v8-inspector into Node.js as a direct dependency for now. The work on integration with Node.js can begin, on the Node.js repository, with more feedback and collaboration from the community. Downstream vendors and module authors could potentially start building debug tooling built on top on the new debug protocol.

Down the road, _if_ we do end up picking up V8 5.1 or 5.2 for Node.js v6, we can simply drop the direct dependency on v8-inspector. I am afraid that waiting until then would reduce the runway to collaborate on the integration and get it polished in time for the v6 LTS.

[ChALkeR]

Bring in websockets (ws) as an internal module to avoid exposing it to user space.

Do we _absolutely_ need it? Are there any chances that we could move most of the code in question here to module outside of the core (i.e. that one has to explicitly npm install)?

Including another huge chunk of thirdparty code into core gives me an uneasy feeling. Even if it is not exposed, it will have to be supported by the collaborators here. For example, any security release in ws (there was one not so long ago, for example) would result in security releases of all Node.js branches that get this feature, and the same for all other modules that would be pulled in. That would also potentially break the schedule and could even force us to introduce some backwards incompatible changes.

For example, would an update of the DevTools that would shrink the range of supported Chrome/Chromium versions be treated as semver-major? What if it includes a security fix? There are more questions like those.

As for the hooks — even introducing a private-ish hook api that would somehow check that it's being used only by a specific module (and moving all the functionality into that module) seems like a more acceptable solution to me.

[jkrems]

@ChALkeR I guess any websocket implementation would do, including a simplified version of ws. I wouldn't expect a manual npm install being the final solution. I'd assume that even if ws would be picked, it would end up in deps/ including all of its dependencies.

[ChALkeR]

@jkrems I suppose that everything I noted above is still valid with the websocket implementation being in ./deps/.

[jkrems]

@ChALkeR Sorry, I completely mis-parsed your first paragraph. Thought you were talking about the fact that it currently _does_ require an explicit npm install when setting up (https://github.com/repenaxa/node/commit/86939cb60b22066f4ee6e818ef879c3dbd6282aa#diff-04c6e90faac2675aa89e2176d2eec7d8R21). But now re-reading your comment it sounds like you would prefer it if node would just ship with the hooks for the protocol and leaving it up to an external (possible npm-) module to actually hook it up to a websocket?

[ChALkeR]

@jkrems Yes, that's what I meant.

[ofrobots]

@ChALkeR: The problem with having part of the functionality being provided by user-space module is that it breaks the out-of-the-box debug experience. In many cases, at the time the user realizes that need to debug something, it may already be too late to get an npm module installed.

That would also potentially break the schedule and could even force us to introduce some backwards incompatible changes.

I am not sure I follow the point about backwards incompatible changes – since it is not going to be externally exposed. We (node-core) will need to address security vulnerabilities in upstream ws (or alternative websockets implementation) and dependent packages. In my opinion, the security impact would be _higher_ if core functionality (remote debug server) were to be dependent on externally provided modules that node-core cannot patch by providing a security update.

I think these are excellent discussions. I suspect that this thread will get unwieldy fast if we were to have these discussions _here_. Apropos https://github.com/nodejs/node/issues/2546#issuecomment-190301820, I propose that we create a branch to work on the v8-inspector integration and use individual PRs to discuss individual design points (websockets, event loop integration, ergonomics, etc.). If there are no objections, I'll proceed with that route.

[ChALkeR]

@ofrobots

I am not sure I follow the point about backwards incompatible changes – since it is not going to be externally exposed.

The debugger is going to be exposed, and it could lose compatibility with older browser versions at some point. Or am I misunderstanding something?

In my opinion, the security impact would be higher if core functionality (remote debug server) were to be dependent on externally provided modules that node-core cannot patch by providing a security update.

I meant — package it separately (perhaps installable with npm install), but keep in the nodejs org. Like nan, for example.

Btw, that means that we could force people to upgrade that tool if we _really_ need them to — Node.js could check the version of the debugger tool being used and refuse to work with it if it's too old.

[jkrems]

@ChALkeR Given that this interface is - afaik - meant as a replacement for the current JSON/TCP API, putting the interface into something external to node would mean that it would no longer be possible for something like an IDE to run a node program in debug mode without having to "patch in" that library. That doesn't sound like a great developer experience imo.

[trevnorris]

Only skimmed over the recent comments, so excuse any duplicates.

The debugger will require the websocket server to run off the main thread, IIRC. Along those lines, it would be more beneficial if that implementation was done w/o requiring a v8 stack. Which rules out using a third-party module.

[ofrobots]

@trevnorris

The debugger will require the websocket server to run off the main thread, IIRC. Along those lines, it would be more beneficial if that implementation was done w/o requiring a v8 stack. Which rules out using a third-party module.

Indeed. That's another option we are investigating. More specifically, we're looking at what would be involved in pulling out the websockets implementation devtools is using in blink. We haven't looked at it enough to know in detail whether we can extract it and plug it into Node. Note however, that we are more interested in getting devtools working with Node than we are in designing a websockets implementation for core. We're trying to find the path of least resistance here.

@ChALkeR

Btw, that means that we could force people to upgrade that tool if we really need them to — Node.js could check the version of the debugger tool being used and refuse to work with it if it's too old.

I think this stance is hostile to our users. If we think debug functionality is important to core, then we will need to own up to the security implications of supporting such functionality. There are cases where the user trying to debug may not be able to install a module from npm (e.g. shrink wrapped deps scenarios, etc.) Having said that, we are also looking at alternative (non-js) implementations of websockets (collaboration from the community on this is welcome). It is possible that we may be able to avoid this debate.

[jkrems]

I'd volunteer to port the needed parts of ws to C++ but I'm not sure if that would be a net positive long-term. Those parts would need some level of maintenance and I'm assuming that bigger chunks of C++ will always be scarier to contributors than JS. Is the current solution (additional V8 context for the debugger thread) causing known issues? There's some ugly parts (e.g. src/node.js having to do some branching if it's running as a debugger thread) but otherwise it seems - relatively - fine to me.

Maybe (random thought) the code for the devtools bindings could even live in the exact same place the current v8 protocol code lives. This might help make the branching between the two interfaces less confusing.

[ChALkeR]

@trevnorris

The debugger will require the websocket server to run off the main thread, IIRC. Along those lines, it would be more beneficial if that implementation was done w/o requiring a v8 stack. Which rules out using a third-party module.

Adding in a brand new c++ implementation of WebSockets gives me an even more uneasy feeling.

I might be missing something again, but why creating an IPC interface on the main thread and hooking to it and passing over WebSocket by a standalone process (that could reuse ws) is not an option?

[YurySolovyov]

There is node-lws which I think is fully C++

[joshgav]

@ofrobots (emphasis added)

To this end we have been working on extracting the DevTools backend from blink, extricate all its dependencies on Blink, and get it integrated into V8 instead.

Once extricated from Blink, would it make more sense to permanently keep v8-inspector as an independent component from v8? That would give us (Node) more flexibility as you already noted, and better facilitate external contributions of additional functionality to v8-inspector.

[ofrobots]

@joshgav The end goal is to get v8-inspector integrated directly into V8 as the fully featured debug API for V8. It is very closely tied to the internals of V8, and keeping them separate is going to be more work for embedders and the developers alike. We would love to see contributions from the community – I think contributing would be a lot easier once this is in V8, but if there is friction please do let us know. /cc @repenaxa.

[pavelfeldman]

Although v8-inspector will live outside of Blink, it will still be used there. As well as in other non-Chrome and non-Node V8 embedders. We were thinking about making it a third party dependency, but eventually decided against it:

• it creates a new step in the dependency roll: Blink -> v8-inspector -> v8 that we can not afford
• it limits our ability to iterate due to two-sided patches and the v8.h surface limitation
• v8-inspector exposes the debugging APIs (wired and strongly-typed) that needs to be versioned and introducing an dependent component with a separate API versioning sounded like too much overhead.

Having said that, we plan to move v8-inspector into v8 along with its contributors! So it would start with the Chrome DevTools and some V8 people and will grow from there. As @ofrobots mentioned, we can work on minimizing the associated friction.

[ofrobots]

I just opened a PR with initial v8_inpector support: https://github.com/nodejs/node/pull/6792.

[trevnorris]

Can this be closed now?

[cjihrig]

I think so. Please reopen if I'm wrong.

[xaxxon]

What is the current status on the effort to integrate in with vanilla v8? I'm currently building this functionality and have it minimally functional but would rather not, but I don't know how to find out about the progress being made.

Thank you.

[bmeck]

@xaxxon use node --inspect on v6+, should work fairly well

[xaxxon]

I'm not using node, though.

On Wed, Oct 5, 2016 at 8:23 AM, Bradley Meck [email protected]
wrote:

@xaxxon https://github.com/xaxxon use node --inspect on v6+, should
work fairly well

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/nodejs/node/issues/2546#issuecomment-251707069, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAIycuUi5Tj8B7sVus9suliVTUI_stEeks5qw8DWgaJpZM4FyAmT
.

[cjihrig]

You might want to check the V8 issue tracker then.

[eugeneo]

The code is in the V8 now, e.g. -
https://github.com/v8/v8/tree/ba41697cbd72ea7dcdd89e6dbe9090ecb156ce6c/src/inspector

On Wed, Oct 5, 2016 at 8:24 AM Colin Ihrig [email protected] wrote:

You might want to check the V8 issue tracker then.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/nodejs/node/issues/2546#issuecomment-251707826, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AARkrVMxf-2hCea0SHccmwmrYQR7_m7Wks5qw8E8gaJpZM4FyAmT
.

[xaxxon]

thank you.

On Wed, Oct 5, 2016 at 9:24 AM, Eugene Ostroukhov [email protected]
wrote:

The code is in the V8 now, e.g. -
https://github.com/v8/v8/tree/ba41697cbd72ea7dcdd89e6dbe9090
ecb156ce6c/src/inspector

On Wed, Oct 5, 2016 at 8:24 AM Colin Ihrig [email protected]
wrote:

You might want to check the V8 issue tracker then.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/nodejs/node/issues/2546#issuecomment-251707826, or
mute
the thread
2hCea0SHccmwmrYQR7_m7Wks5qw8E8gaJpZM4FyAmT>

.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/nodejs/node/issues/2546#issuecomment-251724922, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAIyckAW2RICpAYxAYi6zBOjV2PykixBks5qw88-gaJpZM4FyAmT
.

[msaspence]

I wonder if it is currently or would be possible to access this url programatically from the Node process?

[eugeneo]

No, not yet. We can expose it if it is really needed.

[msaspence]

I'm just think it would be nice to be able to expose it so in the process can open the tab for you automatically if you so wish. However It seems that you can actually determine the URL yourself if you know the port, which you can specify with --inspect=1234

[eugeneo]

UUID is not exposed.

On Wed, Nov 2, 2016 at 9:55 AM Matthew Spence [email protected]
wrote:

I'm just think it would be nice to be able to expose it so in the process
can open the tab for you automatically if you so wish. However It seems
that you can actually determine the URL yourself if you know the port,
which you can specify with --inspect=1234

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/nodejs/node/issues/2546#issuecomment-257927247, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AARkrVmlK-gwGRcNbq4PZSlaTl8O2cSEks5q6MCRgaJpZM4FyAmT
.

[msaspence]

With chrome-cli installed (open doesn't like the chrome-devtools protocol) I have the following proof of concept working on node 6.9.1

const { exec } = require('child_process');
  exec('chrome-cli list links', (_error, stdout) => {
    const url = `chrome-devtools://devtools/bundled/inspector.html?experiments=true&v8only=true&ws=localhost:${process.env.npm_package_config_debugPort || 9221}/node`;
    const match = stdout.match(new RegExp(`\\[(\\d+)\\] chrome-devtools://devtools/bundled/inspector.html\\?experiments=true&v8only=true&ws=localhost:${process.env.npm_package_config_debugPort || 9221}/node`));
    const tabId = match ? ` -t ${match[1]}` : '';
    exec(`chrome-cli open "${url}"${tabId}`);
  });

Where $npm_package_config_debugPort is provided to the node --inspect flag.

[june07]

I was having the same issue a few days ago and wrote a Chrome extension to solve it. Would love any feedback.

http://june07.com/nim

Direct Chrome Web Store link: https://chrome.google.com/webstore/detail/nim-node-inspector-monito/gnhhdgbaldcilmgcpfddgdbkhjohddkj