I'm working on a video streaming service: server streams data back to client-side via grpc.
I'm using bidirectional streaming grpc, since
(1) client would send application-specific instructions to server-side, which happens really rarely;
(2) server sends back response to client-side in chunks;
(3) Though is a video streaming, I cannot loss any data, so doesn't consider UDP.
The overall throughput is about ~80MBps under ~110MBps upper bound throughput.
One problem confuses me a lot is:
On server-side grpc::ServerReaderWriter::Write() often returns false, which mean the stream has been closed: https://grpc.github.io/grpc/cpp/classgrpc_1_1internal_1_1_writer_interface.html#a89d051f5c00ede4469cfb216bd05b594
Meanwhile, one client-side grpc::ClientReaderWriter often gets deadline exceeded error.
I get that error status by:
// when streaming grpc read fails
auto fail_status = stream->Finish(); // deadline exceeded status returned here
This error happens every several seconds, which makes my video streaming application super choppy and totally un-usable.
I tried to investigate this issue by capture packets on both client-side and server-side. Some basic inspection:
I don't find any RST TCP packet
I don't see much TCP (fast) retransmission
My question is:
(1) Is there a way to get detailed error status on server-side?
(2) Can someone give me some suggestions on how to further investigate this issue?
Question on why pubsub requests seem to trigger such a high number of 503 errors? Is this something common? It seems other people see something similar but a majority of my requests end up that way
Similar to
Google Pubsub: UNAVAILABLE: The service was unable to fulfill your request
Catch error code from GCP pub/sub
This is expected behavior. Streaming pull, which is used by the client libraries, creates a bidirectional stream for receiving messages and sending back acknowledgements. These streams stay open for long periods of time and don't close with a successful response code when messages are received, they terminate with an error condition when the stream disconnects, perhaps due to a restart on the part of the server receiving the request or because of brief network blip. Therefore, even if you are receiving messages successfully, you'll still see error response codes for all of the streams themselves. The new streaming pull docs address this question directly.
we run a website that obtains location data through the Google Place API. We have 150k daily searches available, which we haven´t met yet as the website has been live for few weeks only. We have suddenly received a 502 error. A notification in the Console says: “The server encountered a temporary error and could not complete your request.”. Is this a temporary error? Is there any suggestions on what we can do? The website hasn’t been available for 40 minutes.
When you receive 5xx status or UNKNOWN_ERROR in the response, you should implement a retrying logic. Google has a following recommendation in their web services documentation:
In rare cases something may go wrong serving your request; you may receive a 4XX or 5XX HTTP response code, or the TCP connection may simply fail somewhere between your client and Google's server. Often it is worthwhile re-trying the request as the followup request may succeed when the original failed. However, it is important not to simply loop repeatedly making requests to Google's servers. This looping behavior can overload the network between your client and Google causing problems for many parties.
A better approach is to retry with increasing delays between attempts. Usually the delay is increased by a multiplicative factor with each attempt, an approach known as Exponential Backoff.
https://developers.google.com/maps/documentation/directions/web-service-best-practices#exponential-backoff
However, if retrying logic with Exponential Backoff doesn't help and the error persists for a long time you should file a bug in Google issue tracker
I hope this addresses your doubt!
UPDATE
There was an issue on Google side yesterday (November 6, 2017), you can refer to the following bug that explains the issue:
https://issuetracker.google.com/issues/68938173
I was using pyzabbix and trying to use trigger.get method to get all trigger info, but returned 504 Gateway Timeout exception. It never happened before, when I tried to get all the single host trigger info by specifying the host name using filter keyword it worked well. I thought it resulted from the increasing amount of hosts which means large number of trigger it returned. I have about 1800 hosts so far. Any solutions to this problem?
While Zabbix API in general has performance issues, and there are various PHP parameters to control timeouts like max_execution_time, HTTP response code 504 sounds suspicious. If you are using a proxy (maybe transparent, reverse etc), check the timeouts there and consider hitting the Zabbix API directly.
Closed. This question is opinion-based. It is not currently accepting answers.
Want to improve this question? Update the question so it can be answered with facts and citations by editing this post.
Closed 6 years ago.
Improve this question
I'm looking for guidance on good practices when it comes to return errors from a REST API. I'm working on a new API so I can take it any direction right now. My content type is XML at the moment, but I plan to support JSON in future.
I am now adding some error cases, like for instance a client attempts to add a new resource but has exceeded his storage quota. I am already handling certain error cases with HTTP status codes (401 for authentication, 403 for authorization and 404 for plain bad request URIs). I looked over the blessed HTTP error codes but none of the 400-417 range seems right to report application specific errors. So at first I was tempted to return my application error with 200 OK and a specific XML payload (ie. Pay us more and you'll get the storage you need!) but I stopped to think about it and it seems to soapy (/shrug in horror). Besides it feels like I'm splitting the error responses into distinct cases, as some are http status code driven and other are content driven.
So what is the industry recommendations? Good practices (please explain why!) and also, from a client pov, what kind of error handling in the REST API makes life easier for the client code?
A great resource to pick the correct HTTP error code for your API:
http://www.codetinkerer.com/2015/12/04/choosing-an-http-status-code.html
An excerpt from the article:
Where to start:
2XX/3XX:
4XX:
5XX:
So at first I was tempted to return my application error with 200 OK and a specific XML payload (ie. Pay us more and you'll get the storage you need!) but I stopped to think about it and it seems to soapy (/shrug in horror).
I wouldn't return a 200 unless there really was nothing wrong with the request. From RFC2616, 200 means "the request has succeeded."
If the client's storage quota has been exceeded (for whatever reason), I'd return a 403 (Forbidden):
The server understood the request, but is refusing to fulfill it. Authorization will not help and the request SHOULD NOT be repeated. If the request method was not HEAD and the server wishes to make public why the request has not been fulfilled, it SHOULD describe the reason for the refusal in the entity. If the server does not wish to make this information available to the client, the status code 404 (Not Found) can be used instead.
This tells the client that the request was OK, but that it failed (something a 200 doesn't do). This also gives you the opportunity to explain the problem (and its solution) in the response body.
What other specific error conditions did you have in mind?
The main choice is do you want to treat the HTTP status code as part of your REST API or not.
Both ways work fine. I agree that, strictly speaking, one of the ideas of REST is that you should use the HTTP Status code as a part of your API (return 200 or 201 for a successful operation and a 4xx or 5xx depending on various error cases.) However, there are no REST police. You can do what you want. I have seen far more egregious non-REST APIs being called "RESTful."
At this point (August, 2015) I do recommend that you use the HTTP Status code as part of your API. It is now much easier to see the return code when using frameworks than it was in the past. In particular, it is now easier to see the non-200 return case and the body of non-200 responses than it was in the past.
The HTTP Status code is part of your api
You will need to carefully pick 4xx codes that fit your error conditions. You can include a rest, xml, or plaintext message as the payload that includes a sub-code and a descriptive comment.
The clients will need to use a software framework that enables them to get at the HTTP-level status code. Usually do-able, not always straight-forward.
The clients will have to distinguish between HTTP status codes that indicate a communications error and your own status codes that indicate an application-level issue.
The HTTP Status code is NOT part of your api
The HTTP status code will always be 200 if your app received the request and then responded (both success and error cases)
ALL of your responses should include "envelope" or "header" information. Typically something like:
envelope_ver: 1.0
status: # use any codes you like. Reserve a code for success.
msg: "ok" # A human string that reflects the code. Useful for debugging.
data: ... # The data of the response, if any.
This method can be easier for clients since the status for the response is always in the same place (no sub-codes needed), no limits on the codes, no need to fetch the HTTP-level status-code.
Here's a post with a similar idea: http://yuiblog.com/blog/2008/10/15/datatable-260-part-one/
Main issues:
Be sure to include version numbers so you can later change the semantics of the api if needed.
Document...
Remember there are more status codes than those defined in the HTTP/1.1 RFCs, the IANA registry is at http://www.iana.org/assignments/http-status-codes. For the case you mentioned status code 507 sounds right.
As others have pointed, having a response entity in an error code is perfectly allowable.
Do remember that 5xx errors are server-side, aka the client cannot change anything to its request to make the request pass. If the client's quota is exceeded, that's definitly not a server error, so 5xx should be avoided.
I know this is extremely late to the party, but now, in year 2013, we have a few media types to cover error handling in a common distributed (RESTful) fashion. See "vnd.error", application/vnd.error+json (https://github.com/blongden/vnd.error) and "Problem Details for HTTP APIs", application/problem+json (https://datatracker.ietf.org/doc/html/draft-nottingham-http-problem-05).
There are two sorts of errors. Application errors and HTTP errors. The HTTP errors are just to let your AJAX handler know that things went fine and should not be used for anything else.
5xx Server Error
500 Internal Server Error
501 Not Implemented
502 Bad Gateway
503 Service Unavailable
504 Gateway Timeout
505 HTTP Version Not Supported
506 Variant Also Negotiates (RFC 2295 )
507 Insufficient Storage (WebDAV) (RFC 4918 )
509 Bandwidth Limit Exceeded (Apache bw/limited extension)
510 Not Extended (RFC 2774 )
2xx Success
200 OK
201 Created
202 Accepted
203 Non-Authoritative Information (since HTTP/1.1)
204 No Content
205 Reset Content
206 Partial Content
207 Multi-Status (WebDAV)
However, how you design your application errors is really up to you. Stack Overflow for example sends out an object with response, data and message properties. The response I believe contains true or false to indicate if the operation was successful (usually for write operations). The data contains the payload (usually for read operations) and the message contains any additional metadata or useful messages (such as error messages when the response is false).
Agreed. The basic philosophy of REST is to use the web infrastructure. The HTTP Status codes are the messaging framework that allows parties to communicate with each other without increasing the HTTP payload. They are already established universal codes conveying the status of response, and therefore, to be truly RESTful, the applications must use this framework to communicate the response status.
Sending an error response in a HTTP 200 envelope is misleading, and forces the client (api consumer) to parse the message, most likely in a non-standard, or proprietary way. This is also not efficient - you will force your clients to parse the HTTP payload every single time to understand the "real" response status. This increases processing, adds latency, and creates an environment for the client to make mistakes.
Modeling your api on existing 'best practices' might be the way to go.
For example, here is how Twitter handles error codes
https://developer.twitter.com/en/docs/basics/response-codes
Please stick to the semantics of protocol. Use 2xx for successful responses and 4xx , 5xx for error responses - be it your business exceptions or other. Had using 2xx for any response been the intended use case in the protocol, they would not have other status codes in the first place.
Don't forget the 5xx errors as well for application errors.
In this case what about 409 (Conflict)? This assumes that the user can fix the problem by deleting stored resources.
Otherwise 507 (not entirely standard) may also work. I wouldn't use 200 unless you use 200 for errors in general.
If the client quota is exceeded it is a server error, avoid 5xx in this instance.