I have been exploring Netty for the past days, as I am writing a quick and tight HTTP server that should receive lots of requests, and Netty's HTTP server implementation is quite simple and does the job.
My next step is as part of the request handling, I need to launch an HTTP request to an external web server. My intuition is to implement an asynchronous client that can send a lot of requests simultaneously, but I am a little confused as what is the right approach. My understanding is that Netty server uses a worker thread for each incoming message, therefore that worker thread would not be freed to accept new messages until my handler finishes its work.
Here is the punch: even if I have an asynchronous HTTP client in hand, it won't matter if I need to wait for each response and process it back with my server handler - the same worker thread would remain blocking all this time. The alternative is to use the async nature of the client, returning a future object quickly to release the thread and place a listener (meaning I have to return 200 or 202 status to the client), and check my future object to indicate when the response is received and I can push it to the client.
Does this make sense? Am I way off with my assumptions? What is a good practice to implement such kind of Netty acceptor server + external client with high concurrency?
Thanks,
Assuming you're asking about Netty 4.
Netty configured with a ServerBootstrap will have a fixed number of worker threads that it uses to accept requests and execute the channel, like so:
Two threads accepting / processing requests
bootstrap.group(NioEventLoopGroup(2))
One thread accepting requests, two threads processing.
bootstrap.group(NioEventLoopGroup(1), NioEventLoopGroup(1))
In your case, you have a channel includes a bunch of Http Codec decoding/encoding stuff and your own handler which itself makes an outgoing Http request. You're right that you don't want to block the server from accepting incoming requests, or decoding the incoming Http message, and there are two things you can do to mitigate that, you've struck on the first already.
Firstly, you want to use an Async netty client to make the outgoing requests, have a listener write the response to the original requests channel when the outgoing request returns. This means you don't block and wait, meaning you can handle many more concurrent outgoing requests than the number of threads available to process those requests.
Secondly, you can have your custom handler run in its own EventExecutorGroup, which means it runs in a separate threadpool from the acceptor / http codec channel handlers, like so:
// Two separate threads to execute your outgoing requests..
EventExecutorGroup separateExecutorGroup new DefaultEventExecutorGroup(2);
bootstrap.childHandler(new ChannelInitializer<SocketChannel>() {
#Override
public void initChannel(SocketChannel ch) {
ChannelPipeline pipeline = ch.pipeline();
.... http codec stuff ....
pipeline.addLast(separateExecutorGroup, customHandler);
}
};
Meaning your outgoing requests don't hog the threads that would be used for accepting / processing incoming ones.
Related
I have fundumental question about how async requests work at top level.
Imagin if we have a top level route called HomePage(). This route is an async route and within this route we call to 10 different APIs before sending the response(image it takes like 5 seconds, remember this is an example to understand the concept and these numbers are for learning purposes). All of these api requests are awaited. So the request handler just releases the thread hanlding this request and goes to handle other requests until the response for these apis come back. So lets add this constraint. Our network card can handle only 1 connection and that one is held open till the response for the request to HomePage is ready. Therefor we cannot make any other requests to the server so whats the difference if this whole thing was sync from the beggining. We cannot drop the connection to the first request to HomePage because if that's the case then how are we ever going to send back the response for that request and we cannot handle new requests because the connection is kept open.
I suspect that my problem is how the reponse is sent back on top level async routes.
Can anybody give a deep dive explaination on how these requests are handled that can take more requests and still send back the response(because if it can send back a response the connection HAS TO HAVE KEPT ALIVE). Examples would be much appreciated.
So lets add this constraint. Our network card can handle only 1 connection
That constraint cannot exist. Network cards handle packets, not connections. Connections are a virtual construct that exist in the host computer.
Can anybody give a deep dive explaination on how these requests are handled that can take more requests and still send back the response(because if it can send back a response the connection HAS TO HAVE KEPT ALIVE).
Of course the connection is kept alive. The top-level async method will return the thread to the thread pool, where it is available to handle any other requests.
If you have some artificial constraint on your web app that prevents it from having more than one connection, then there won't be any other requests to handle, and the thread pool threads will do nothing.
Say I have a webserivce used internally by other webservices with an average response time of 1 minute.
What are the pros and cons of such a service with "synchronous" responses versus making the service return id of the request, process it in the background and make the clients poll for results?
Is there any cons with HTTP connections which stay active for more than one minute? Does the default keep alive of TCP matters here?
Depending on your application it may matter. Couple of things worth mentioning are !
HTTP protocol is sync
There is very wide misconception that HTTP is async. Http is synchronous protocol but your client could deal it async. E.g. when you call any service using http, your http client may schedule is on the background thread (async). However The http call will be waiting until either it's timeout or response is back , during all this time the http call chain is awaiting synchronously.
Sockets
Since HTTP uses socket and there is hard limit on sockets. Every HTTP connection (if created new every time) opens up new socket . if you have hundreds of requests at a time you can image how many http calls are scheduled synchronously and you may run of sockets. Not sure for other operation system but on windows even if you are done with request sockets they are not disposed straight away and stay for couple of mins.
Network Connectivity
Keeping http connection alive for long is not recommended. What if you loose network partially or completely ? your http request would timeout and you won't know the status at all.
Keeping all these things in mind it's better to schedule long running tasks on background process.
If you keep the user waiting while your long job is running on server, you are tying up a valuable HTTP connection while waiting.
Best practice from RestFul point of view is to reply an HTTP 202 (Accepted) and return a response with the link to poll.
If you want to hang the client while waiting, you should set a request timeout at the client end.
If you've some Firewalls in between, that might drop connections if they are inactive for some time.
Higher Response Throughput
Typically, you would want your OLTP (Web Server) to respond quickly as possible, Since your queuing the task on the background, your web server can handle more requests which results to higher response throughput and processing capabilities.
More Memory Friendly
Queuing long running task on background jobs via messaging queues, prevents abusive usage of web server memory. This is good because it will increase the Out of memory threshold of your application.
More Resilient to Server Crash
If you queue task on the background and something goes wrong, the job can be queued to a dead-letter queue which helps you to ultimately fix problems and re-process the request that caused your unhandled exceptions.
Since golang handles incoming requests in separate goroutines, it's unclear to me which types of tasks should be deferred for processing by a message queue e.g. NSQ consumers and which should be handled within the http request goroutine.
Since the net/http package runs each request you do not need to worry about blocking the request goroutine. The real question you should ask myself is "Do I need to do this before I return a response to the client, or can it be deferred until later". Generally if I need to fetch from a database to serve a response that will block the request goroutine, and that is ok. If I can return a response now and put a message on a queue to do stuff later, that can be ok too.
Since the request goroutine has little cost to exist, and it is isolated from other requests, you really don't need to worry about it that much. Do what makes sense for the client.
Is there a way to find out if a HttpServletRequest is aborted?
I'm writing an instant browser application (some kind of chat): The clients asks for new events in a loop using AJAX-HTTP-Requests. The server (Tomcat) handles the requests in a HttpServlet. If there are no new events for this client, the server delays the reply until a new event arrives or a timeout occurs (30sec).
Now I want to identify clients that are no longer polling. Therefore, I start a kick-Timer at the end of a request which is stopped when a new request arrives. If the client closes the browser window the TCP-Connection is closed and the HTTP-Request is aborted.
Problem: The client does not run into the kick-Timeout because the Servlet still handles the event request - sleeping and waiting for an event or timeout.
It would be great if I could somehow listen for connection abort events and then notify the waiting request in order to stop it. But I couldn't find anything like that in the HttpServletRequest or HttpServletResponse...
This probably won't help the OP any more, but it might help others trying to detect aborted HTTP connections in HttpServlet in general, as I was having a similar problem and finally found an answer.
The key is that when the client cancels the request, normally the only way for the server to find out is to send some data back to the client, which will fail in that case. I wanted to detect when a client stops waiting for a long computation on server, so I ended up periodically writing a single character to response body through HttpServletResponse's writer. To force sending the data to the client, you must call HttpServletResponse.flushBuffer(), which throws ClientAbortException if the connection is aborted.
You are probably using some sort of thread-notification (Semaphores or Object.wait) to hold and release the Servlet threads. How about adding a timeout (~10s) to the wait, then somehow checking whether the connection is still alive and then continuing the wait for another 10s, if the connection is still there.
I don't know whether there are reliable ways to poll the "liveness" of the connection (e.g. resp.getOutputStream not throwing an Exception) and if so, which way is the best (most reliable, least CPU intense).
It seems like having waiting requests could degrade the performance of your system pretty quickly. The threads that respond to requests would get used up fast if requests are held open. You could try completing all requests (and returning "null" to your clients if there is no message), and having a thread on the back-end that keeps track of how long it's been since clients have polled. The thread could mark a client as being inactive.
My web app must process and serve a lot of data to display certain pages. Sometimes, the user closes or refreshes a page while the server is still busy processing it. This means the server will continue to process data for several minutes only to send it to a client who is no longer listening.
Is it possible to detect that the connection has been broken, and react to it?
In this particular project, we're using Django and NginX, or Apache. I assumed this is possible because the Django development server appears to react to cancelled requests by printing Broken Pipe exceptions. I'd love to have it raise an exception that my application code could catch. It appears JSP can do this. So can node.js here.
Alternatively, I could register an unload event handler on the page in question, have it do a synchronous XHR requesting that the previous request from this user be cancelled, and do some kind of inter-process communication to make it so. Perhaps if the slower data processing were handed to another process that I could more easily identify and kill, without killing the responding process...
While #Oded is correct that HTTP is stateless between requests, app servers can indeed detect when the underlying TCP/IP connection has broken for the request being processed. Why is this? Because TCP is a stateful protocol for reliable connections.
A common technique for .Net web apps processing a resource intensive request is to check Response.IsClientConnected (docs) before starting the resource intensive work. There is no point in wasting CPU cycles to send an expensive response to a client that isn't there anymore.
private void Page_Load(object sender, EventArgs e)
{
// Check whether the browser remains
// connected to the server.
if (Response.IsClientConnected)
{
// If still connected, do work
DoWork();
}
else
{
// If the browser is not connected
// stop all response processing.
Response.End();
}
}
Please reply with your target app server stack so I can provide a more relevant example.
Regarding your 2nd alternative to use XHR to post client page unload events to the server, #Oded's comment about HTTP being stateless between requests is spot on. This is unlikely to work, especially in a farm with multiple servers.
HTTP is stateless, hence the only way to detect a disconnected client is via timeouts.
See the answers to this SO question (Java Servlet : How to detect browser closing ?).