Hi Han Liu One more reminder, a trace id in one instance could have multiple threads sampling in theory, such as across threads scenarios. We also should set a threshold for this. Max 3 threads for one trace id maybe?
Sheng Wu 吴晟 Twitter, wusheng1108 Sheng Wu <[email protected]> 于2019年12月13日周五 下午1:03写道： > Hi Han Liu and everyone > > I have submitted a design draft to the doc. Please take a look, if you > have an issue, please let me known. We could set up a online meeting too. > > Sheng Wu <[email protected]>于2019年12月12日 周四下午8:49写道： > >> Hi Han Liu >> >> I have replied the design with the most important key points I expect. >> Let's discuss those. After we are on the same page, we could continue on >> more details. >> >> Sheng Wu 吴晟 >> Twitter, wusheng1108 >> >> >> han liu <[email protected]> 于2019年12月12日周四 下午2:26写道： >> >>> Due to formatting issues with previous mailboxes, they have been replaced >>> with new ones. >>> >>> I have completed some of the features in the google doc, and can provide >>> your comments and improvements. I will continue to improve the following >>> functions in the documentation. >>> The documentation is the same as you previously sent me. To prevent >>> trouble, I'll post the link again here. >>> >>> https://docs.google.com/document/d/1rxMf1WN3PaFaZp7r8JmtwfdkmjLTcFW_ETAZv5FIU-s/edit# >>> >>> Sheng Wu <[email protected]> 于2019年12月10日周二 上午10:46写道： >>> >>> > 741550557 <[email protected]> 于2019年12月9日周一 下午9:42写道： >>> > >>> > > Thank for your reply, the issues you mentioned are very critical and >>> > > meaningful. >>> > > There I will answer what you mentioned. Sorry, I'm not good at >>> comment >>> > > mode, so I use different colors and “ “ prefix to QA. >>> > > >>> > > >>> > > As we already have designed limit mechanism at backend and agent >>> > > side(according to your design), also the number would not be big(10 >>> most >>> > > likely), we just need a list to storage the trace-id(s) >>> > > >>> > > >>> > > If just need a list to storage trace-id(s), so how can I map to the >>> > > thread? I hope to use the map to quickly find thread info from >>> trace-id. >>> > > How can I get thread-stack information from your way? Could you >>> please >>> > > help elaborate? >>> > > >>> > >>> > Why do you need to do that? You just save a list of thread ids which >>> should >>> > do thread dump, or remove some thread id from them when the trace id is >>> > finished. >>> > This is easy to do this by doing a loop search in the list. Right? >>> > Thread-stack is in the list, they are stored in an element. Also, they >>> are >>> > in a list too. >>> > >>> > I think you were thinking the same all stack in a single map? That will >>> > cause a very dangerous memory and GC risk. >>> > >>> > >>> > >>> > > >>> > > >>> > > Could you explain the (2), what do you mean `stop`? I think if your >>> > > sampling mechanism should include the sampling duration. >>> > > >>> > > >>> > > As far as the communication between the sniffer and the OAP server, I >>> > hope >>> > > the sniffer only needs to obtain the thread-monitor task that needs >>> to be >>> > > monitored at this time. The termination condition can be stopped by >>> the >>> > > sniffer or the OAP server. >>> > > If It’s just an OAP server notification, it may be more complicated. >>> > Cause >>> > > OAP server need record sniffer has received the current command, and >>> > > sniffer is not stable, such as sniffer has shutdown when receiving >>> the >>> > > command, at this time, no thread information I have been collected. >>> > > >>> > > >>> > > I think that the active calculation termination by the OAP server can >>> > make >>> > > the monitoring more controllable, of course, the client can also >>> actively >>> > > report the end. >>> > > I think it’s necessary to provide a protection mechanism for the >>> sniffer >>> > > side, and it can be released quickly when the business peak period >>> or the >>> > > probe suddenly occupies a lot of CPU / memory resources. Therefore, >>> the >>> > > function of stopping monitoring can be provided in the UI interface, >>> so >>> > > that the sniffer can recover. >>> > > Sampling duration is required, but only as a default termination >>> > > thread-monitor condition. >>> > > >>> > >>> > But you should know, in the real case, the thread dump monitor is a >>> > sampling mechanism, you are even hard to know where they are happening. >>> > Then you have to send the stop notification to every instance. >>> > Even you could send the notification, but could you explain how you >>> know to >>> > stop? >>> > The scenario is, you are facing an issue, which trace and metrics can't >>> > explain, so you active thread dump, right? At the same time, you want >>> to >>> > stop? >>> > >>> > CPU and memory resources should be guaranteed by design level, such as >>> > 1. Limited thread dump task for one service. >>> > 2. Limited thread dump traces in the certain time window. >>> > For example, the OAP backend/UI would say, you only could >>> > 1. Set 3 thread dump commands in the same time window. >>> > 2. Every command will require the sampling thread dump number should be >>> > less than 5 traces. At the same time, in order to make this sampling >>> works, >>> > only active sampling thread dump after the trace executed more than >>> > 200ms(value is an example only). >>> > 3. Thread dump could be sent to the backend duration sampling to >>> reduce the >>> > memory cache. >>> > 4. Thread dump period should not less than 5ms, recommend 20ms >>> > 5. How depth the thread dump should do >>> > >>> > We need a very detailed design, above are just my thoughts, in order to >>> > share the idea, the safe of the agent should not be by UI button. >>> > Otherwise, your online system will be very dangerous, which is not the >>> > design goal of SkyWalking. >>> > >>> > >>> > >>> > > >>> > > >>> > > The sampling period depends on how you are going to visualize it. >>> > > >>> > > >>> > > Yes, I agree. I hope can provide a select/input let trace count and >>> time >>> > > windows can be configurable in UI. Of course, this is my current >>> idea, >>> > and >>> > > if there have other plains, I will adopt it. >>> > > >>> > > >>> > > Highly doubt about this, reduce the memory, maybe, only reduce if >>> the >>> > > codes >>> > > are running the loop or facing lock issue. But if it is neither of >>> these >>> > > two, they are different. >>> > > Also, please consider the CPU cost of the comparison of the stack. >>> You >>> > > need >>> > > a performance benchmark to verify if you want this. >>> > > >>> > > >>> > > I didn’t understand that first sentence. In my personal experience, >>> most >>> > > of the cases are blocking in the lock(socket/local) and running >>> loop. I >>> > > have not imagined any other cases? >>> > > For the second sentence, I think I can add a thread-stack-element >>> field >>> > to >>> > > storage the top-level element of last stack information. When get >>> stack >>> > > information next time, I can compare the current top-level element >>> that >>> > is >>> > > the same with that field. >>> > > I do this mainly to reduce duplicate thread-stack information form >>> taking >>> > > up too much memory space, this is a way to optimizing memory space. >>> It >>> > can >>> > > consider remove it, or do you have a better memory-saving solution? >>> After >>> > > all, memory and CPU resources are very valuable in the sniffer. >>> > > >>> > >>> > I know you mean about reducing the memory, but do you consider how >>> much CPU >>> > you will cost do a full thread dump comparison? The thread dump could >>> > easily be hundreds of lines in Java. >>> > I mean this is a tradeoff, CPU or memory. If you are just using limited >>> > memory, before you could send the snapshot to backend while collecting >>> new, >>> > even could save into the disk(if really necessary). >>> > In my experience, compress is always very high risk in the agent, if >>> you >>> > want to do that, you need a benchmark test to improve that, this CPU >>> cost >>> > is small enough. >>> > >>> > >>> > >>> > > >>> > > >>> > > The trace number and time window should be configurable, that is I >>> mean >>> > > more complex. Inthe current SamplingServcie, only n traces per 3 >>> > seconds. >>> > > But here, it is a dynamic rule. >>> > > >>> > > >>> > > I expect that it can be configured at the UI level for special trace >>> > count >>> > > and time windows as I said above. >>> > > For SamplingService, my previous tech design was not rigorous >>> enough, and >>> > > there were indeed problems. >>> > > Maybe we need to extend a new SamplingService, build a mapping base >>> on >>> > > endpoint-id and AtomicInteger. >>> > > For `first 5 traces of this endpoint in the next 5 mins`, just need >>> to >>> > > increment it. >>> > > For sampling, maybe use another schedule task to reset AtomicInteger >>> > value. >>> > > >>> > >>> > You could avoid map, by using ArrayList with >>> RangeAtomicInteger(SkyWalking >>> > provides that) to let the trace context to get the slot. >>> > Also, you are considering `active sampling after trace execution time >>> more >>> > than xxx ms`, you should add remove mechanism during runtime. >>> > Anyway, try your best to avoid using Map, especially this map could be >>> > changed in the runtime. >>> > >>> > >>> > >>> > > >>> > > >>> > > I think at least should be a level one new page called >>> configuration or >>> > > command page, which could set up the multiple sampling rule and >>> > visualize >>> > > the existing tasks and related sampling data. >>> > > >>> > > >>> > > I think it’s necessary to add a new page to the configuration >>> > > thread-monitor task, I think the specific UI display should be >>> designed >>> > in >>> > > detail. >>> > > For example, what I expected is similar to the trace page. The left >>> side >>> > > displays the configuration, and the right side quickly displays the >>> > related >>> > > trace list. When clicked, it quickly links to the trace page and >>> displays >>> > > the sidebox display. >>> > > I ’m not good at this. Do you have any good plans? >>> > > >>> > >>> > UI is the thing that is hard to discuss by text, so I am pretty sure, >>> we >>> > need some demo(could not be the codes, that is I mean drew by a tool) >>> > It is OK to show a trace with thread dumps on another page, even better >>> > linking to your task ID. >>> > But this kind of abstract description is hard to continue, no details I >>> > mean. >>> > >>> > >>> > >>> > > And I feel that the two of us have a different understanding of the >>> > > configuration object. I think it is more of a task than a command. I >>> > don't >>> > > know which way is better? >>> > > I suddenly thought of a problem. I think that some real problems are >>> > often >>> > > triggered at a specific period, such as a fixed business peak >>> period, and >>> > > we cannot guarantee that the user will operate on the UI. >>> > > So should the task mechanism be adopted to ensure that it can be run >>> at a >>> > > specific period? >>> > > >>> > >>> > This makes sense to me, and it is a just enhance feature. It is just a >>> > start time sampling rule. >>> > >>> > >>> > >>> > > >>> > > >>> > > We don't have separated thread monitor view table, how about we add >>> an >>> > > icon >>> > > at the segment list, and add icon at the first span of this segment >>> in >>> > > trace detail view? >>> > > I think the latter one should be an entrance of the thread view. >>> > > >>> > > >>> > > I think it's a good idea. The link I mentioned in one of the answers >>> > > above, I think it is also a convenient entry point. >>> > > The switch button I mentioned earlier is only a data filtering item >>> in >>> > the >>> > > query of the trace list and does not need a separate table UI. >>> > > >>> > >>> > As you intend to have a separated page for thread sampling, it is OK to >>> > >>> > >>> > > >>> > > >>> > > If you have some visualization idea, drawn by any tool you like >>> > supporting >>> > > comment, we could discuss it there. In my mind, we should support >>> > > visualize >>> > > the thread dump stack through the time windows, and support >>> aggregate >>> > them >>> > > by choosing the continued stack snapshots on the time window. >>> > > >>> > > >>> > > I think we should find a front-end who is better at discussing >>> together >>> > > because this depends on how the front-end UI can be displayed. >>> > > BTW: I can provide code for the OAP server and sniffer, and the >>> frontend >>> > > may need to look for help in the community alone. Hope that any >>> front-end >>> > > friends can participate in the topic discussion. >>> > > >>> > >>> > Once you have the demo, I could loop our UI committers in for UI side >>> > development. But UI committers may not be familiar with thread dump >>> context >>> > story. We need to resolve that first. >>> > Let's start up a demo, such as some slides on Google doc? >>> > >>> > >>> > > >>> > > >>> > > >>> > > >>> > > The above is my answer to all the questions, and I look forward to >>> your >>> > > reply at any time. As more and more discussions took place, the >>> details >>> > > became more and more complete. This is good. >>> > > Everyone is welcome to discuss together if you have any questions or >>> good >>> > > ideas, please let me know. >>> > > >>> > >>> > I think we could move the discussion to the design doc as the next >>> step. >>> > >>> > Please use this >>> > >>> > >>> https://docs.google.com/document/d/1rxMf1WN3PaFaZp7r8JmtwfdkmjLTcFW_ETAZv5FIU-s/edit# >>> > Trite the design including >>> > 1. Key features >>> > 2. Protocol >>> > 3. Work mechanism >>> > 4. UI design, prototype >>> > and anything you think important before writing codes. >>> > >>> > This is SkyWalking CLI design doc, you could use it as a reference. >>> > >>> > >>> https://docs.google.com/document/d/1WBnRNF0ABxaSdBZo6Gv2hMzCQzj04YAePUdOyLWHWew/edit# >>> > >>> > >>> > > >>> > > >>> > > 原始邮件 >>> > > 发件人:Sheng [email protected] >>> > > 收件人:[email protected] >>> > > 发送时间:2019年12月9日(周一) 10:50 >>> > > 主题:Re: A proposal for Skywalking(thread monitor) >>> > > >>> > > >>> > > Hi Thanks for writing this proposal with a detailed design. My >>> comments >>> > > are inline. 741550557 [email protected] 于2019年12月8日周日 下午11:22写道： >>> Thanks >>> > > for your reply, I have carefully read these issues you mentioned, >>> and >>> > > these issues mentioned are very meaningful and critical. I will >>> give you >>> > > technical details about the issues you mentioned below. I find these >>> > > issues are related, so I will explain them in different dimensions. >>> > use >>> > > a different protocol to transmission trace and thread-stack: 1. add >>> a >>> > > boolean field in segment data, to record has thread monitored. and >>> is >>> > good >>> > > for filter monitored trace in UI. 2. add new BootService, storage >>> Map to >>> > > record relate trace-id and trace-stack information. As we already >>> have >>> > > designed limit mechanism at backend and agent side(according to your >>> > > design), also the number would not be big(10 most likely), we just >>> need a >>> > > list to storage the trace-id(s) 3. listen >>> > > TracingContextListener#afterFinished if the current segment has >>> thread >>> > > monitored, mark current trace-id don’t need to monitor anymore. >>> (Cause >>> > if >>> > > for-each the step 2 map, the remove operation will fail and throw >>> > > exception). 4. when thread-monitor main thread running, It will >>> for-each >>> > > step 2 map and check is it don’t need monitor anymore, I will put >>> data >>> > > into new data carrier. 5. generate new thread-monitor gRPC >>> protocol to >>> > > send data from the data carrier. The agent side design seems pretty >>> > good. >>> > > the server receives thread-stack logic: 1. storage stack-stack >>> > > informations and trace-id/segment-id relations on a different >>> table. 2. >>> > > check thread-monitor is need to be stop on receiving data or >>> schedule. >>> > > Could you explain the (2), what do you mean `stop`? I think if your >>> > > sampling mechanism should include the sampling duration. reduce >>> CPU >>> > and >>> > > memory in sniffer: 1. through the configuration of thread >>> monitoring in >>> > > the UI, you can configure the performance loss. For example, set the >>> > > monitoring level: fast monitoring (100ms), medium speed monitoring >>> > > (500ms), slow speed monitoring (1000ms). The sampling period >>> depends on >>> > > how you are going to visualize it. 2. add new integer field on per >>> > > thread-stack, if current thread-stack last element same as last >>> time, >>> > > don’t need storage, just increment it. I think it will save a lot of >>> > > memory space. Highly doubt about this, reduce the memory, maybe, only >>> > > reduce if the codes are running the loop or facing lock issue. But >>> if it >>> > is >>> > > neither of these two, they are different. Also, please consider the >>> CPU >>> > > cost of the comparison of the stack. You need a performance >>> benchmark to >>> > > verify if you want this. 3. create new VM args to setting >>> thread-monitor >>> > > pool size, It dependence on user, maybe default 3? (this can be >>> > discussed >>> > > later) I think UI limit is enough. 3 seems good to me. 4. limit >>> > > thread-stack-element size to 100, I think it can resolve most of the >>> > > scenes already. It also can create a new VM args if need. multiple >>> > > sampling methods can choose :(just my current thoughts, can add >>> more) >>> > 1. >>> > > base on current client SamplingServcie, extra a new factor holder to >>> > > increment, and reset on schedule. Yours may be a little more complex >>> > than >>> > > the current SamplingServcie, right? Based on the next rule. 2. >>> `first 5 >>> > > traces of this endpoint in the next 5 mins`, it a good idea. My >>> > > understanding is that within a few minutes, each instance can send a >>> > > specified number of traces. The trace number and time window should >>> be >>> > > configurable, that is I mean more complex. Inthe current >>> SamplingServcie, >>> > > only n traces per 3 seconds. But here, it is a dynamic rule. UI >>> > settings >>> > > and sniffer perception: 1. create a new button on the dashboard >>> page, It >>> > > can create or stop a thread-monitor. It can be dynamic load >>> > thread-monitor >>> > > status when reselecting endpoint. I think at least should be a >>> level >>> > one >>> > > new page called configuration or command page, which could set up the >>> > > multiple sampling rule and visualize the existing tasks and related >>> > > sampling data. 2. sniffer creates a new scheduled task to check the >>> > > current service has need monitor endpoint each 5 seconds. (I see >>> current >>> > > sniffer has command functions, feel that principle is the same as >>> the >>> > > scheduler) Seems reasonable. thread-monitor on the UI:(That’s >>> just my >>> > > initial thoughts, I think there will have a better way to show) 1. >>> When >>> > > switch to the trace page, I think we need to add a new switch >>> button to >>> > > filter thread-monitor trace. 2. make a new thread-monitor icon on >>> the >>> > same >>> > > segment. It means it has thread-stack information. We don't have >>> > > separated thread monitor view table, how about we add an icon at the >>> > > segment list, and add icon at the first span of this segment in trace >>> > > detail view? I think the latter one should be an entrance of the >>> thread >>> > > view. 3. show on the information sidebox when the user clicks the >>> > > thread-monitor segment(any span). create a new tab, like the log >>> tab. >>> > If >>> > > you have some visualization idea, drawn by any tool you like >>> supporting >>> > > comment, we could discuss it there. In my mind, we should support >>> > visualize >>> > > the thread dump stack through the time windows, and support aggregate >>> > them >>> > > by choosing the continued stack snapshots on the time window. >>> They're >>> > > just a description of my current implementation details for >>> > thread-monitor >>> > > if these seem to work. I can do some time planning for these tasks. >>> > Sorry, >>> > > my English is not very well, hope you can understand. Maybe these >>> seem >>> > to >>> > > have some problem, any good idea or suggestion are welcome. Very >>> > > appreciated you to lead this new direction. It is a long term task >>> but >>> > > should be interesting. :) Good work, carry on. 原始邮件 发件人:Sheng >>> > > [email protected] 收件人:[email protected] >>> > > 发送时间:2019年12月8日(周日) 08:31 主题:Re: A proposal for Skywalking(thread >>> > > monitor) First of all, thanks for your proposal. Thread >>> monitoring is >>> > > super important for application performance. So basically, I agree >>> with >>> > > this proposal. But for tech details, I think we need more >>> discussion in >>> > > the following ways 1. Do you want to add thread status to the >>> trace? If >>> > > so, why don't consider this as a UI level join? Because we could >>> know >>> > > thread id in the trace when we create a span, right? Then we have >>> all >>> > the >>> > > thread dump(if), we could ask UI to query specific thread context >>> based >>> > > on timestamp and thread number(s). 2. For thread dump, I don't know >>> > > whether you do the performance evaluation for this OP. From my >>> > > experiences, `get all need thread monitor segment every 100 >>> > milliseconds` >>> > > is a very high cost in your application and agent. So, you may need >>> to >>> > be >>> > > careful about doing this. 3. Endpoint related thread dump with some >>> > > sampling mechanisms makes more sense to me. And this should be >>> activated >>> > > by UI. We should only provide a conditional thread dump sampling >>> > > mechanism, such as `first 5 traces of this endpoint in the next 5 >>> mins`. >>> > > Jian Tan I think DaoCloud also has customized this feature in your >>> > > internal SkyWalking. Could you share what you do? Sheng Wu 吴晟 >>> Twitter, >>> > > wusheng1108 741550557 [email protected] 于2019年12月8日周日 上午12:14写道： >>> Hello >>> > > everyone, I would like to share a new feature with skywalking, >>> called >>> > > “thread monitor”. Background When our company used skywalking to APM >>> > > earlier, we found that many traces did not have enough span to fill >>> up, >>> > > doubting whether there were some third-party frameworks that we >>> didn't >>> > > enhance or programmers API usage errors such as java CountDown >>> number >>> > is 3 >>> > > but there are only 2 countdowns. So we decide to write a new >>> feature to >>> > > monitor executing trace thread stack, then we can get more >>> information >>> > on >>> > > the trace, quick known what’s happening on that trace. Structure >>> > > Agent(thread monitor) — gRPC protocol — OAP Server(Storage) — >>> > > Skywalking-Rocketbot-UI More detail OAP Server: 1. Storage witch >>> traces >>> > > need to monitor(i suggest storage on the endpoint, add new boolean >>> field >>> > > named needThreadMonitor) 2. Provide GraphQL API to change endpoint >>> > monitor >>> > > status. 3. Monitor Trace parse, storage thread stack if the segment >>> has >>> > > any thread info. Skywalking-Rocketbot-UI: 1. Add a new switch >>> button on >>> > > the dashboard, It can read or modify endpoint status. 2. It will >>> show >>> > > every thread stack on click trace detail. Agent: 1. setup two new >>> > > BootService: 1) find any need thread monitor endpoint in current >>> > service, >>> > > start on a new schedule take and works on each minute. 2) start new >>> > > schedule task to get all need thread monitor segment each 100 >>> > > milliseconds, and put a new thread dump task to a global thread >>> > > pool(fixed, count number default 3). 2. check endpoint need thread >>> > monitor >>> > > on create entry/local span(TracingConext#createEntry/LocalSpan). If >>> > need, >>> > > It will be marked and put into thread monitor map. 3. when >>> > TraceingContext >>> > > finishes, It will get thread has monitored, and send all thread >>> stack to >>> > > server. Finally, I don’t know it is a good idea to get more >>> information >>> > on >>> > > trace? If you have any good ideas or suggestions on this, please >>> let me >>> > > know. Mrpro >>> > >>> >> -- > Sheng Wu 吴晟 > > Apache SkyWalking > Apache Incubator > Apache ShardingSphere, ECharts, DolphinScheduler podlings > Zipkin > Twitter, wusheng1108 >
