There exists a known race condition in the HTTP keepalive mechanism:
HTTP KeepAlive connection closed by server but client had sent a request in the mean time
https://github.com/mikem23/keepalive-race
As I understand, I need my HTTP client either to have a shorter timeout than my HTTP server, or retry when getting TCP-FIN or TCP-RST.
My question is, how do today's web-browsers, that use the HTTP keepalive feature, handle this race condition. Do they retry?
I'll be happy for references, a google search hasn't come up with anything.
According the the RFC, in these cases, a server should respond with a 408 error code, signalling to the client that the connection has already been closed on its side. As the RFC states:
If the client has an outstanding request in transit, the client MAY
repeat that request on a new connection.
This means that it's up to the client (aka each browser) to decide how a 408 response will be handled. There are 2 alternatives:
handle it gracefully: retrying the remaining requests in a new connection automatically, so that the user stays completely unaware of the underlying failure that happened
fail-fast: showing the user a failure with the appropriate 408 error message
For example, it seems that Chrome in the past was following the second approach until a point, where people started considering this as a "buggy" behaviour and switched to the first one. You can find the bug thread related to the Chromium bug here and the associated code change here.
Note: As you can read in the final emails in the linked thread, Chrome performs these retries, only when some requests have succeeded in this connection. As a result, if you try to reproduce that with a single request, returning a 408 response, you'll notice that Chrome won't probably retry in that case.
Related
I stumbled upon a case where a request to an endpoint might take more than 60 seconds (let's say that's the timeout value), in which case the server sends a response and continues processing the request in the background. There are also cases where the same request would be processed before it times out and a successful response would be sent from the server to the client.
What would be the best HTTP code to use in those first case? I read HTTP server timeout. When should it be sent, which suggests 503 or 504, and HTTP status code for 'Loading', which mentions that the request can be deemed successful and return 200. But I'm not convinced by any of those suggestions more than the others yet.
No
HTTP protocol doesn't work that way.
A server would receive a request, process it and sends a reply. The cycle ends there.
HTTP is never intended to send multui-stage replies with different states. You need to work on a custom protocol built on top of HTTP if you want to do that.
Sending timeout error as an indication of an unfinished response is an anti pattern. If your server takes more time than usual to process a request, you should send a success response with an ID which can be used to poll the state of the initial request and get the results.
So to summarize from your question and comments: you have an HTTP API that takes a command and executes it, and sends a callback-reply through a webhook. If the execution takes longer than a minute, you have to send some form of reply that indicates the request is still being processed.
There are various problems with executing long-running work in an HTTP request handler. For starters, you tie up HTTP server resources (threads, sockets) while processing non-HTTP work, you can't restart the HTTP server without losing work, and so on.
So I would opt for a queuing mechanism that takes in the work, replies 200 OK or 201 Created immediately, and then schedules the work for processing on a background thread or even a different service. When finished, you execute the webhook callback.
Any error response to the initial call will leave the caller confused: they won't know whether their requested work will finish, unless you use an "exotic" status code that actually differs from real error conditions, and document that they can expect that.
Charlie and CodeCaster suggested to return 200 or 201 and I took a look at the other 2xx codes and found 202 Accepted:
From https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/202
The HyperText Transfer Protocol (HTTP) 202 Accepted response status
code indicates that the request has been accepted for processing, but
the processing has not been completed; in fact, processing may not
have started yet. The request might or might not eventually be acted
upon, as it might be disallowed when processing actually takes place.
202 is non-committal, meaning that there is no way for the HTTP to
later send an asynchronous response indicating the outcome of
processing the request. It is intended for cases where another process
or server handles the request, or for batch processing.
I wonder if this would fit best.
This question already has answers here:
FIN vs RST in TCP connections
(3 answers)
Closed 2 years ago.
I am using iOS but I am asking for networking in general. What does it mean to cancel a network request? Is there a message sent to the server or does the server acknowledge the socket being disconnected?
As you mention using NSURLSessionTask as your way to request, cancel() means a urlSession(_:task:didCompleteWithError:) will be send to the tasks delegate. But passing in a global error code NSURLErrorCancelled (-999) to the defined NSURLErrorDomain.
It is possible that cancelation is later called on the task as a complete processing of the request message is done. So it's up to you to act accordingly once your ErrorDomain is getting the error code NSURLErrorCancelled marking your intention to cancel, and therefore would want to throw away any data that is received since last request.
The Server gets possibly a complete request but your client is not receiving answers anymore. Or the request sequence is not complete so the Server recognises not correct what was intended but would work thru the request until it fails cause of incomplete or wrong formatted request data.
When your receiver callback is down do to canceling you just don't parse any answer of the Server and if you could still parse the Server data that would mean your task is still running. Any result after cancel() should be treated as possibly incomplete or misleading/wrong/invalid. This is why you set a NSURLErrorCancelled error to a NSURLErrorDomain, you want to know what the status is before you assume any received data is of value for you.
By the way NSURLErrorCancelled is also thrown when NSURLSessionAuthChallengeCancelAuthenticationChallenge is marking a server with no trust. So it's actually the same procedure, you decide if any received data is something you want to trust to.
If a socket is disconnected, there is no connection at all, no data passing thru, nothing to receive. nothing to request from. Any Error on both sides can't be exchanged. Server and Client are disconnected then.
Canceling a request does not imply a socket is stopped from working.
It just means the data since the last request is to be handled as invalid.
Why is this?
Because you can construct your own sockets, ignoring ErrorDomain stuff with a complete different request pattern.
Also means in case of client error/crash/canceling nothing is send, you just do not accept any answer as valid even if it was delivered thru the sockets.
For this reasons there are Protocols that define how a message should look like and what should happen in case it was incomplete or would need any kind of validation in a given pattern that validates any data that was send. TCP, UDP, JS-Websocket with handshake and ongoing "dataflow", even OSC etc. and lots of other protocols.
I am talking about only one case here.
client sent a request to server -> server received it and returned a response -> unfortunately the response dropped.
I have only one question about this.
Is this case even possible? If it's possible then what should the response code be, or will client simply see it as read timeout?
As I want to sync status between client/server and want 100% accuracy no matter how poor the network is, the answer to this question can greatly affect the client's 'retry on failure' strategy.
Any comment is appreciated.
Yes, the situation you have described is possible and occurs regularly. It is called "packet loss". Since the packet is lost, the response never reaches the client, so no response code could possibly be received. Web browsers will display this as "Error connecting to server" or similar.
HTTP requests and responses are generally carried inside TCP packets. If a TCP packet carrying the HTTP response does not arrive in the expected time window, the request is retransmitted. The request will only be retransmitted a certain number of times before a timeout error will occur and the connection is considered broken or dead. (The number of attempts before TCP timeout can be configured on both the client and server sides.)
Is this case even possible?
Yes. It's easy to see why if you picture a physical cable between the client and the server. If I send a request down the cable to the server, and then, before the server has a chance to respond, unplug the cable, the server will receive the request, but the client will never "hear" the response.
If it's possible then what should the response code be, or will client simply see it as read timeout?
It will be a timeout. If we go back to our physical cable example, the client is sitting waiting for a response that will never come. Hopefully, it will eventually give up.
It depends on exactly what tool or library you're using how this is wrapped up, however - it might give you a specific error code for "timeout" or "network error"; it might wrap it up as some internal 5xx status code; it might raise an exception inside your code; etc.
When a TCP connection gets cancelled by the client while making a HTTP request, I'd like to stop doing any work on the server and return an empty response. What HTTP status code should such a response return?
To be consistent I would suggest 400 Bad Request now if your backend apps are capable of identifying when the client gets disconnected or if you reject or close the connection, probably you could return Nginx' non-standard code 499 or 444.
499 Client Closed Request
Used when the client has closed the request before the server could send a response.
444 No Response
Used to indicate that the server has returned no information to the client and closed the connection.
HTTP (1.0/1.1) doesn't have a means to cancel a request. All that a client can do if it no longer wants the response is to close the connection and hope that the server contains an optimization to stop working on a response that can no longer be delivered. Since the connection is now closed, no response nor status code can actually be delivered to the client and so any code you "return" is only for your own satisfaction. I'd personally pick something in the 4xx range1 since the "fault" - the reason you can no longer deliver a response - is due to the client.
HTTP 2.0 does allow an endpoint to issue END_STREAM or RST_STREAM to indicate that they are no longer interested in one stream without tearing down the whole connection. However, they're meant to just ignore any further HEADERS or DATA sent on that stream and so even though you may theoretically deliver a status code, the client is still going to completely ignore it.
1Probably 400 itself since I can't identify a more specific error that seems entirely appropriate.
There are just a few plausible choices (aside from 500, of course):
202 Accepted
You haven't finished processing, and you never will.
This is appropriate only if, in your application domain, the original requestor "expects" that not all requests will be satisfied.
409 Conflict
…between making and cancelling the request.
This is only weakly justified: your situation does not involve one client making a request based on out of date information, since the cancellation had not yet occurred.
503 Service Unavailable
The service is in fact unavailable for this one request (because it was cancelled!).
The general argument of "report an error as an error" favors 409 or 503. So 503 it is by default.
There really is little to do. Quoting from RFC 7230, section 6.5:
A client, server, or proxy MAY close the transport connection at any time.
That is happening at TCP-, not HTTP-level. Just stop processing the connection. A status code will confer little meaning here as the intent of an incomplete/broken request is mere speculation. Besides, there will be no means to transport it across to the client.
I understand an HTTP request will result in a response with a code and optional body.
If we call the originator of the request the 'client' and the recipient of the request the 'server'.
Then the sequence is
Client sends request
Server receives request
Server sends response
Client receive response
Is it possible for the Server to complete step 3 but step 4 does not happen (due to dropped connection, application error etc).
In other words: is it possible for the Server to 'believe' the client should have received the response, but the client for some reason has not?
Network is inherently unreliable. You can only know for sure a message arrived if the other party has acknowledged it, but you never know it did not.
Worse, with HTTP, the only acknowledge for the request is the answer and there is no acknowledge for the answer. That means:
The client knows the server has processed the request if it got the response. If it does not, it does not know whether the request was processed.
The server never knows whether the client got the answer.
The TCP stack does normally acknowledge the answer when closing the socket, but that information is not propagated to the application layer and it would not be useful there, because the stack can acknowledge receipt and then the application might not process the message anyway because it crashes (or power failed or something) and from perspective of the application it does not matter whether the reason was in the TCP stack or above it—either way the message was not processed.
The easiest way to handle this is to use idempotent operations. If the server gets the same request again, it has no side-effects and the response is the same. That way the client, if it times out waiting for the response, simply sends the request again and it will eventually (unless the connection was torn out never to be fixed again) get a response and the request will be completed.
If all else fails, you need to record the executed requests and eliminate the duplicates in the server. Because no network protocol can do that for you. It can eliminate many (as TCP does), but not all.
There is a specific section on that point on the HTTP RFC7230 6.6 Teardown (bold added):
(...)
If a server performs an immediate close of a TCP connection, there is
a significant risk that the client will not be able to read the last
HTTP response.
(...)
To avoid the TCP reset problem, servers typically close a connection
in stages. First, the server performs a half-close by closing only
the write side of the read/write connection. The server then
continues to read from the connection until it receives a
corresponding close by the client, or until the server is reasonably
certain that its own TCP stack has received the client's
acknowledgement of the packet(s) containing the server's last
response. Finally, the server fully closes the connection.
So yes, this response sent step is a quite complex stuff.
Check for example the Lingering close section on this Apache 2.4 document, or the complex FIN_WAIT/FIN_WAIT2 pages for Apache 2.0.
So, a good HTTP server should maintain the socket long enough to be reasonably certain that it's OK on the client side. But if you really need to acknowledge something in a web application, you should use a callback (image callback, ajax callback) asserting the response was fully loaded in the client browser (so another HTTP request). That means it's not atomic as you said, or at least not transactional like you could expect from a relational database. You need to add another request from the client, that maybe you'll never get (because the server had crash before receiving the acknowledgement), etc.