I'm still struggling with disconnects with my slow readers. Here is all that I have experimented with:
I downloaded the latest version of mod_wsgi source (5.0.0) and built it on my Centos7 system. This all seemed to work well and I am now running that version. I modified my app to return an iterator and stopped calling "write" directly. No real change in behavior. To eliminate network devices causing problems, I started experimenting with using localhost. Paradoxically, the localhost connections seem to be timing out *sooner* than remote ones and seem to time out 100% of the time! The timeout seems to occur at around 17-ish minutes. I have the apache config param "TimeOut" set to 43200 (12 hours). I know that is insane but that should make a send that is blocked by a slow reader sit there for 12 hours. Alas, it does not. My slow readers are still timing out. The error log has this: [Wed Feb 07 19:17:26.807222 2024] [wsgi:info] [pid 19013] [client 127.0.0.1:38570] mod_wsgi (pid=19013, process='', application='xcrutils.exegy-appliance.net|/xcr'): Reloading WSGI script '/var/web/sites/request_handler_wsgi.py'. [Wed Feb 07 19:34:03.493988 2024] [wsgi:debug] [pid 19013] src/server/mod_wsgi.c(2443): [client 127.0.0.1:38570] mod_wsgi (pid=19013): Failed to write response data: Connection timed out. My python wsgi application outputs this: Apache/mod_wsgi failed to write response data: Connetion timed out That error string is what is in the trapped exception itself. Looking at mod_wsgi, this call: rv = ap_pass_brigade(r->output_filters, self->bb); is resulting in rv being not equal to APR_SUCCESS and exception_when_aborted is false. Could there be some kind of timeout implemented in the bucket brigade code? On Friday, January 12, 2024 at 8:35:00 AM UTC-6 Greg Popp wrote: > Thank you again! The data IS in a file, but it requires an application to > extract the requested salient pieces. I will look at the file wrapper > extension. > > After more testing, I now think that I can fix my problem just by using > your second suggestion of increasing the Timeout configuration variable in > Apache. That is an easy fix and so far seems to be working well. > > On Thursday, January 11, 2024 at 4:22:38 PM UTC-6 Graham Dumpleton wrote: > >> If using the file_wrapper feature, make sure you also add: >> >> WSGIEnableSendfile On >> >> to mod_wsgi configuration as not on by default: >> >> >> https://modwsgi.readthedocs.io/en/master/release-notes/version-4.1.0.html#features-changed >> >> The file_wrapper mechanism would still have worked, but to use kernel >> sendfile feature have to also have the directive enabled. >> >> Can't remember if also need to add: >> >> EnableSendfile On >> >> to enable it in Apache itself. I don't think so. >> >> Graham >> >> On 12 Jan 2024, at 9:14 am, Graham Dumpleton <graham.d...@gmail.com> >> wrote: >> >> Also not sure whether it will help or not, but if the data you are >> sending is stored in a file and not generated on demand, then you might >> consider using the WSGI file_wrapper extension instead. >> >> >> https://modwsgi.readthedocs.io/en/master/user-guides/file-wrapper-extension.html >> >> I don't know how this will behave when buffer fills up since when working >> properly it is all handled in the OS kernel and not in Apache. >> >> Along similar lines, if is stored as a file, you might try mod_sendfile. >> It also would use kernel sendfile mechanism, but way it interacts may also >> see different behaviour in your situation. >> >> Graham >> >> On 12 Jan 2024, at 1:38 am, Greg Popp <pop...@gmail.com> wrote: >> >> Thank you very much! This is most helpful, though I don't think any of >> them will actually solve my issues, for many of the reasons you mentioned. >> >> I was thinking that perhaps the mod_wsgi interface had access to the file >> descriptor for the network socket used by Apache and could call "select" to >> see if it had enough buffer space for the requested write. If it didn't, it >> could (optionally) sleep some configurable duration and try again some >> configurable number of times. I understand though, that for most >> applications this would not be necessary. >> >> Yesterday, I tried implementing that same behavior in my wsgi app. I >> don't set the SendBufferSize and so use the system default. I grab the >> system TCP send buff value by running the 'sysctl' command. Then I keep >> track of the total bytes sent. If that value exceeds the system tcp send >> queue value, I run the 'ss' command from within my wsgi app to grab the >> Send-Q value for this connection (fortunately wsgi gives us the source-ip >> and source-port and I can filter the 'ss' output using that). If the Send-Q >> value is too high to accommodate another write, I sleep a second and try >> again until I get enough space. It's kind of a Rube Goldberg solution, but >> so far it seems to be working! >> >> Thank you for taking the time to answer my questions! I very much >> appreciate the assistance! >> >> On Wednesday, January 10, 2024 at 3:38:51 PM UTC-6 Graham Dumpleton wrote: >> >>> So what you are encountering is limitations in the socket buffer size >>> enforced by the operating system, in combination with Apache httpd applying >>> a socket timeout. >>> >>> In other words what happens is that the HTTP client isn't reading data >>> and so the operating system level socket buffer fills up. At that point the >>> Apache httpd write of the response data blocks with it eventually timing >>> out, causing the initial error you see. In that situation Apache httpd will >>> close down the connection, which results in you seeing the second error >>> when still trying to write out more data anyway. >>> >>> You may be able to adjust some Apache configuration settings to try and >>> solve this, but it would affect all requests in the context which you apply >>> the configuration (dependent on whether done in server, VirtualHost, >>> Directory or Location contexts). So not something you could selectively do >>> on a per client basis. >>> >>> The first Apache directive to look at is SendBufferSize. >>> >>> https://httpd.apache.org/docs/2.4/mod/mpm_common.html#sendbuffersize >>> >>> If this is not set it should default to 0, which means it uses the >>> operating system default. >>> >>> So you might be able to fiddle with this by setting it larger than the >>> operating system default (although there is still some upper bound set by >>> operating system you can go to). >>> >>> The next Apache directive to look at is Timeout. >>> >>> https://httpd.apache.org/docs/2.4/mod/core.html#timeout >>> >>> This would usually default to 60 seconds but some Linux distributions >>> may override this in the Apache configuration they ship. >>> >>> In very old Apache versions this actually defaulted to 300 seconds, but >>> it was made lower at some point. >>> >>> If playing with these, do be careful since they cause increased memory >>> usage or cause other undesirable effects depending on traffic profile your >>> server gets. >>> >>> One other thing you may be able to use is mod_ratelimit. >>> >>> https://httpd.apache.org/docs/2.4/mod/mod_ratelimit.html >>> >>> I have never actually used this and not exactly sure how it works, so is >>> a bit of a guess, but you may be able to use this to slow down how quickly >>> your application outputs the data. >>> >>> I am assuming here that this module will introduce waits into your >>> application, by blocking your writes for a bit, to keep the flow of data >>> being written by it under the rate limit. This would have the effect of not >>> stuffing so much data into the response pipeline such that things work >>> better with slower clients. Obviously using it would though penalise faster >>> clients but you might find an acceptable balance by setting a higher rate >>> limit for the initial burst of data and then using a lower rate after that. >>> >>> Graham >>> >>> >>> On 11 Jan 2024, at 7:09 am, Greg Popp <pop...@gmail.com> wrote: >>> >>> embedded >>> >>> On Wednesday, January 10, 2024 at 1:32:52 PM UTC-6 Graham Dumpleton >>> wrote: >>> >>>> Are you using mod_wsgi embedded mode or daemon mode? >>>> >>>> Graham >>>> >>>> On 11 Jan 2024, at 2:44 am, Greg Popp <pop...@gmail.com> wrote: >>>> >>>> Hello! >>>> >>>> My version of mod_wsgi is running on a Centos-7 system and is at >>>> version 3.4, (I know - very old) with python 2.7 >>>> >>>> I have been using mod_wsgi for a python application that runs a >>>> command-line program and marshals the output of the command line program >>>> back to an http client. The data being sent is binary and can be tens of >>>> gigs in size. >>>> >>>> This app is "unconventional", in that it calls 'write' directly, >>>> instead of returning an iterable. The problem I have had recently, is that >>>> some clients are slow to read the data and the TCP buffer gets filled up. >>>> When this happens, the next call to write on a full buffer causes a >>>> "failed >>>> to write data" exception (which I trap) but if I try again to send the >>>> data >>>> I get "client connection closed". >>>> >>>> Is there some config setting or methodology I can use to alleviate this >>>> issue? In other words, some way to back off and wait for the buffer to >>>> drain sufficiently to resume sending the data? OR - is there some way to >>>> get the current size (fullness) of the TCP write buffer on the connected >>>> socket? (Something like what you see from the 'ss' command line utility >>>> "Send-Q" column). If I could tell how full it is and what the max size is, >>>> I could implement a sleep/retry cycle of some kind. >>>> >>>> I have looked - even in the source code - but haven't been able to >>>> figure it out if there is a way to achieve this. Thanks in advance, for >>>> your attention. >>>> >>>> >>>> >>>> -- >>>> You received this message because you are subscribed to the Google >>>> Groups "modwsgi" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to modwsgi+u...@googlegroups.com. >>>> To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/modwsgi/3d97c06f-38ff-4345-af2f-eb86c2ef204cn%40googlegroups.com >>>> >>>> <https://groups.google.com/d/msgid/modwsgi/3d97c06f-38ff-4345-af2f-eb86c2ef204cn%40googlegroups.com?utm_medium=email&utm_source=footer> >>>> . >>>> >>>> >>>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "modwsgi" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to modwsgi+u...@googlegroups.com. >>> >>> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/modwsgi/9fc6ab3e-b791-4503-a3c0-20ba273b92bdn%40googlegroups.com >>> >>> <https://groups.google.com/d/msgid/modwsgi/9fc6ab3e-b791-4503-a3c0-20ba273b92bdn%40googlegroups.com?utm_medium=email&utm_source=footer> >>> . >>> >>> >>> >> -- >> You received this message because you are subscribed to the Google Groups >> "modwsgi" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to modwsgi+u...@googlegroups.com. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/modwsgi/2cdbd013-0f3f-4e38-bb05-dddbb5b0deaan%40googlegroups.com >> >> <https://groups.google.com/d/msgid/modwsgi/2cdbd013-0f3f-4e38-bb05-dddbb5b0deaan%40googlegroups.com?utm_medium=email&utm_source=footer> >> . >> >> >> >> -- You received this message because you are subscribed to the Google Groups "modwsgi" group. To unsubscribe from this group and stop receiving emails from it, send an email to modwsgi+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/modwsgi/36ccd27f-a765-45b5-882a-593f10f2a7d9n%40googlegroups.com.