the (2) is a good idea and may be enough for all plugin developer. I have thought about to add a special interface which have a method.If the method can run successfully, the plugin should be activated.This way may be more flexible for plugin developer.
在 2020/12/29 上午11:45，“Sheng Wu”<[email protected]> 写入: I am a little confused, are you asking for a bug? Or something internal. I thought you said this is an internal thing. For the case you mentioned, I could provide some options 1. This is a return value, so consider an object as type, and use instanceof to check. Because this is an HTTP level invocation, the reflection usually should be quick enough. 2. Enhance the witness class with a supporting about method signatures, Such as add witness method signature, you provide class name + method name + types of all parameters and return value to activate the plugin. We should only consider (2) if it is really necessary actually. Sheng Wu 吴晟 Twitter, wusheng1108 Li BingLong(智能平台) <[email protected]> 于2020年12月29日周二 上午11:39写道： > I know. > The method I intercept is Dispatcher.dispatch. > And In the > org.apache.skywalking.apm.plugin.mrd.DispatcherInterceptor#beforeMethod > There is some code " FullHttpResponse response = > simpleHttpResponse.getHttpResponse();" > The getHttpResponse method has different returnType in different version. > > 在 2020/12/29 上午11:34，“Sheng Wu”<[email protected]> 写入: > > Hi > > Do you know the instrumentation#InstanceMethodsInterceptPoint could add > `ReturnType` as a match condition? > > Sheng Wu 吴晟 > Twitter, wusheng1108 > > > Li BingLong(智能平台) <[email protected]> 于2020年12月29日周二 > 上午11:32写道： > > > Its not the method intercepted throw error. > > For example, in the following code, > simpleHttpResponse.getHttpResponse() > > throws error because different version has different returnType. > > > > @Override > > public Object afterMethod(EnhancedInstance objInst, Method > method, > > Object[] allArguments, Class<?>[] argumentsTypes, > > Object ret) throws Throwable { > > > > SimpleHttpResponse simpleHttpResponse = (SimpleHttpResponse) > > allArguments[1]; > > FullHttpResponse response = > simpleHttpResponse.getHttpResponse(); > > > > AbstractSpan span = ContextManager.activeSpan(); > > if (response.status().code() >= 400) { > > span.errorOccurred(); > > Tags.STATUS_CODE.set(span, > > Integer.toString(response.status().code())); > > } > > ContextManager.stopSpan(span); > > return ret; > > } > > > > 在 2020/12/28 下午11:16，“Sheng Wu”<[email protected]> 写入: > > > > You could create this kind of scenario, yes, definitely. > > My question is, SkyWalking's agent technology is for resolving > tracing > > and > > metrics collection, which related to RPC related methods mostly, > and > > only > > very few parameter types or methods. > > You are keeping changing codes, but not randomly, and if the > method is > > not > > matching for instrumentation declaration, SkyWalking agent core > will > > automatically ignore it already. > > > > Why still you are facing `invoke error`? If you define > instrumentation > > only > > by class name and method name, yes, for some internal codes, > maybe. > > But why > > don't you provide more accurate match rules? > > > > Sheng Wu 吴晟 > > Twitter, wusheng1108 > > > > > > Li BingLong(智能平台) <[email protected]> 于2020年12月28日周一 > > 下午10:36写道： > > > > > It is a real case in our inner project "http-server" which is > build > > using > > > netty. > > > Usually we add one or more class to add a new future.Then > update the > > > version, Gray released. > > > > > > Recently, I resolve the netty conflict using maven shade > plugin and > > update > > > the version.All version not equal to the target version will > using > > > different instrumentation. > > > I think this logic is very difficult to express by the witness > class. > > > > > > On the other hand, If people have to use the Uranus project to > find a > > > witness, It's a little more expensive to resolve the problem. > > > Think about this: If I have a runtime method invoke error,Why > should > > I use > > > a witness class to resolve it? > > > > > > > > > 在 2020/12/28 下午8:18，“Sheng Wu”<[email protected]> 写入: > > > > > > Hi > > > > > > I am not sure whether this is a real case. Is there really > > existing > > > case 2 > > > versions including 100% the same codes? > > > There is a project created by the community called, uranus, > > which you > > > could > > > find on the website. > > > https://skywalking.apache.org/docs/ > > > > > > This is a tool to help on locating witness class. > > > > > > Sheng Wu 吴晟 > > > Twitter, wusheng1108 > > > > > > > > > Li BingLong(智能平台) <[email protected]> > 于2020年12月28日周一 > > > 下午7:33写道： > > > > > > > Sometimes its too difficult to find a witness class. > > > > Maybe the version 2 has all class in version 1.How can I > > define a > > > > instrumentation if I want the version 1 to be instrument? > > > > In my opinion,we usually want to instrument a specified > > method. If we > > > > dont have other ways,some code about a method may be a > good > > “witness > > > > code”other than witness class. > > > > > > > > > > > > > > > >
