<div dir="ltr">David, thank you very much. I confirm everything you wrote, both doing LineBuffering and removing output to stderr resolves the problem.<div><br></div><div>I will create a ticket for hslogger in order to propose somehow putting LineBuffering to stderr in case logging there is enabled. That should help people like myself who don't know they have the problem, but may suffer from it from time to time (at least every write to stderr slows down response-time quite significantly).</div><div><br></div><div>Thanks for the "queues" rsyslog article, I'll take a look. At the time of having task to "centralize logs" I didn't want to add rsyslog as one more required service, I also had an impression that all current "syslog" implementations are in a somewhat outdated stage with multiple badly supported forks, so I did minimal work of adding decoupled forwarding logs from files to central rsyslog, but maybe it's time to migrate everything to rsyslogs and get rid of forwarder services.</div><div><br></div><div>Anyway, thank you once again.</div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Sun, Apr 26, 2015 at 5:52 AM, David Turner <span dir="ltr"><<a href="mailto:dct25-561bs@mythic-beasts.com" target="_blank">dct25-561bs@mythic-beasts.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">I see.<br>
<br>
The issue seems to be the default handler which writes the log to<br>
stderr. Replacing 'addHandler h' with 'setHandlers [h]' makes it run<br>
in a reasonable time as 'setHandlers' starts afresh; conversely it's<br>
still slow even if you just use the default handler on its own, i.e.<br>
removing the call to 'addHandler'.<br>
<br>
I was a bit suspicious about the 'hFlush' in System.Log.Handler.Simple<br>
but removing that didn't help. However, suspecting that the issue may<br>
be too-much-flushing I eventually found<br>
<a href="https://ghc.haskell.org/trac/ghc/ticket/7418" target="_blank">https://ghc.haskell.org/trac/ghc/ticket/7418</a> which says that writing<br>
to stderr is slow because its default buffering mode is NoBuffering. I<br>
added 'hSetBuffering stderr LineBuffering' and boom, it runs at a<br>
sensible speed.<br>
<br>
Recommend you get rid of writing to stderr and just log to the file<br>
unless you've a good reason to send the output both ways, in which<br>
case switch to LineBuffering as above.<br>
<br>
If you're using rsyslog you also may be interested to read<br>
<a href="http://www.rsyslog.com/doc/queues.html" target="_blank">http://www.rsyslog.com/doc/queues.html</a> - this describes how it can<br>
decouple the final log output from the input using a combination of<br>
in-memory and on-disk buffers, and even discard lower-priority<br>
messages from the queue if the going gets really tough. We judged it a<br>
lot of effort to implement our own on-disk spool and were particularly<br>
worried about it growing without bound if the downstream was too slow,<br>
so this feature of rsyslog was just what we needed.<br>
<br>
Hope that helps,<br>
<br>
David<br>
<div class="HOEnZb"><div class="h5"><br>
<br>
On 25 April 2015 at 23:44, Kostiantyn Rybnikov <<a href="mailto:k-bx@k-bx.com">k-bx@k-bx.com</a>> wrote:<br>
> Hi David.<br>
><br>
> I planned to create a detailed bug-report at hslogger's issues to start<br>
> investigation there (as a better place) on Monday, but since I have code<br>
> prepared already, it's easy to share it right now:<br>
> <a href="https://gist.github.com/k-bx/ccf6fd1c73680c8a4345" target="_blank">https://gist.github.com/k-bx/ccf6fd1c73680c8a4345</a><br>
><br>
> I'm launching it as:<br>
><br>
> time ./dist/build/seq/seq &> /dev/null<br>
><br>
> We don't use syslog driver, instead we have a separate file-to-syslog worker<br>
> to decouple these components.<br>
><br>
> On Sun, Apr 26, 2015 at 12:08 AM, David Turner<br>
> <<a href="mailto:dct25-561bs@mythic-beasts.com">dct25-561bs@mythic-beasts.com</a>> wrote:<br>
>><br>
>> Hi,<br>
>><br>
>> I've had a look at this as we use hslogger too, so I'm keen to avoid<br>
>> this kind of performance issue. I threw a quick Criterion benchmark<br>
>> together:<br>
>><br>
>> <a href="https://gist.github.com/DaveCTurner/f977123b4498c4c64569" target="_blank">https://gist.github.com/DaveCTurner/f977123b4498c4c64569</a><br>
>><br>
>> The headline result on my test machine are that each log call takes<br>
>> ~540us, so 2000 should take about a second. Would be interested if you<br>
>> could run the same benchmark on your setup as it's possible that<br>
>> there's something else downstream that's causing you a problem.<br>
>><br>
>> A couple of things that might be worth bearing in mind: if you're<br>
>> talking to syslog over /dev/log then that can block if the log daemon<br>
>> falls behind: unix datagram sockets don't drop datagrams when they're<br>
>> congested. If the /dev/log test is slow but the UDP test is fast then<br>
>> it could be that your syslog can't handle the load.<br>
>><br>
>> I'm using rsyslogd and have enabled the feature that combines<br>
>> identical messages, so this test doesn't generate much disk IO and it<br>
>> keeps up easily, so the UDP and /dev/log tests run about equally fast<br>
>> for me. Is your syslog writing out every message? It may be flushing<br>
>> to disk after every message too, which would be terribly slow.<br>
>><br>
>> If you're not logging to syslog, what's your hslogger config?<br>
>><br>
>> Cheers,<br>
>><br>
>> David<br>
>><br>
>><br>
>> On 24 April 2015 at 20:25, Kostiantyn Rybnikov <<a href="mailto:k-bx@k-bx.com">k-bx@k-bx.com</a>> wrote:<br>
>> > An update for everyone interested (and not). Turned out it's neither GHC<br>
>> > RTS, Snap or networking issues, it's hslogger being very slow. I thought<br>
>> > it's slow when used concurrently, but just did a test when it writes<br>
>> > 2000<br>
>> > 5kb messages sequentially and that finishes in 111 seconds (while<br>
>> > minimal<br>
>> > program that writes same 2000 messages finishes in 0.12s).<br>
>> ><br>
>> > I hope I'll have a chance to investigate why hslogger is so slow in<br>
>> > future,<br>
>> > but meanwhile will just remove logging.<br>
>> ><br>
>> > On Thu, Apr 23, 2015 at 4:08 PM, Kostiantyn Rybnikov <<a href="mailto:k-bx@k-bx.com">k-bx@k-bx.com</a>><br>
>> > wrote:<br>
>> >><br>
>> >> All right, good news!<br>
>> >><br>
>> >> After adding ekg, gathering its data via bosun and seeing nothing<br>
>> >> useful I<br>
>> >> actually figured out that I could try harder to reproduce issue by<br>
>> >> myself<br>
>> >> instead of waiting for users to do that. And I succeeded! :)<br>
>> >><br>
>> >> So, after launching 20 infinite curl loops to that handler's url I was<br>
>> >> quickly able to reproduce the issue, so the task seems clear now: keep<br>
>> >> reducing the code, reproduce locally, possibly without external<br>
>> >> services<br>
>> >> etc. I'll write up after I get to something.<br>
>> >><br>
>> >> Thanks.<br>
>> >><br>
>> >> On Wed, Apr 22, 2015 at 11:09 PM, Gregory Collins<br>
>> >> <<a href="mailto:greg@gregorycollins.net">greg@gregorycollins.net</a>> wrote:<br>
>> >>><br>
>> >>> Maybe but it would be helpful to rule the scenario out. Johan's ekg<br>
>> >>> library is also useful, it exports a webserver on a different port<br>
>> >>> that you<br>
>> >>> can use to track metrics like gc times, etc.<br>
>> >>><br>
>> >>> Other options for further debugging include gathering strace logs from<br>
>> >>> the binary. You'll have to do some data gathering to narrow down the<br>
>> >>> cause<br>
>> >>> unfortunately -- http client? your code? Snap server? GHC event<br>
>> >>> manager<br>
>> >>> (System.timeout is implemented here)? GC? etc<br>
>> >>><br>
>> >>> G<br>
>> >>><br>
>> >>> On Wed, Apr 22, 2015 at 10:14 AM, Kostiantyn Rybnikov <<a href="mailto:k-bx@k-bx.com">k-bx@k-bx.com</a>><br>
>> >>> wrote:<br>
>> >>>><br>
>> >>>> Gregory,<br>
>> >>>><br>
>> >>>> Servers are far from being highly-overloaded, since they're currently<br>
>> >>>> under a much less load they used to be. Memory consumption is stable<br>
>> >>>> and<br>
>> >>>> low, and there's a lot of free RAM also.<br>
>> >>>><br>
>> >>>> Would you say that given these factors this scenario is unlikely?<br>
>> >>>><br>
>> >>>> On Wed, Apr 22, 2015 at 7:56 PM, Gregory Collins<br>
>> >>>> <<a href="mailto:greg@gregorycollins.net">greg@gregorycollins.net</a>> wrote:<br>
>> >>>>><br>
>> >>>>> Given your gist, the timeout on your requests is set to a<br>
>> >>>>> half-second<br>
>> >>>>> so it's conceivable that a highly-loaded server might have GC pause<br>
>> >>>>> times<br>
>> >>>>> approaching that long. Smells to me like a classic Haskell memory<br>
>> >>>>> leak<br>
>> >>>>> (that's why the problem occurs after the server has been up for a<br>
>> >>>>> while):<br>
>> >>>>> run your program with the heap profiler, and audit any shared<br>
>> >>>>> tables/IORefs/MVars to make sure you are not building up thunks<br>
>> >>>>> there.<br>
>> >>>>><br>
>> >>>>> Greg<br>
>> >>>>><br>
>> >>>>> On Wed, Apr 22, 2015 at 9:14 AM, Kostiantyn Rybnikov <<a href="mailto:k-bx@k-bx.com">k-bx@k-bx.com</a>><br>
>> >>>>> wrote:<br>
>> >>>>>><br>
>> >>>>>> Hi!<br>
>> >>>>>><br>
>> >>>>>> Our company's main commercial product is a Snap-based web app which<br>
>> >>>>>> we<br>
>> >>>>>> compile with GHC 7.8.4. It works on four app-servers currently<br>
>> >>>>>> load-balanced<br>
>> >>>>>> behind Haproxy.<br>
>> >>>>>><br>
>> >>>>>> I recently implemented a new piece of functionality, which led to<br>
>> >>>>>> weird behavior which I have no idea how to debug, so I'm asking<br>
>> >>>>>> here for<br>
>> >>>>>> help and ideas!<br>
>> >>>>>><br>
>> >>>>>> The new functionality is this: on specific url-handler, we need to<br>
>> >>>>>> query n external services concurrently with a timeout, gather and<br>
>> >>>>>> render<br>
>> >>>>>> results. Easy (in Haskell)!<br>
>> >>>>>><br>
>> >>>>>> The implementation looks, as you might imagine, something like this<br>
>> >>>>>> (sorry for almost-real-haskell, I'm sure I forgot tons of imports<br>
>> >>>>>> and other<br>
>> >>>>>> things, but I hope everything is clear as-is, if not -- I'll be<br>
>> >>>>>> glad to<br>
>> >>>>>> update gist to make things more specific):<br>
>> >>>>>><br>
>> >>>>>> <a href="https://gist.github.com/k-bx/0cf7035aaf1ad6306e76" target="_blank">https://gist.github.com/k-bx/0cf7035aaf1ad6306e76</a><br>
>> >>>>>><br>
>> >>>>>> Now, this works wonderful for some time, and in logs I can see<br>
>> >>>>>> both,<br>
>> >>>>>> successful fetches of external-content, and also lots of timeouts<br>
>> >>>>>> from our<br>
>> >>>>>> external providers. Life is good.<br>
>> >>>>>><br>
>> >>>>>> But! After several days of work (sometimes a day, sometimes couple<br>
>> >>>>>> days), apps on all 4 servers go crazy. It might take some interval<br>
>> >>>>>> (like 20<br>
>> >>>>>> minutes) before they're all crazy, so it's not super-synchronous.<br>
>> >>>>>> Now: how<br>
>> >>>>>> crazy, exactly?<br>
>> >>>>>><br>
>> >>>>>> First of all, this endpoint timeouts. Haproxy requests for a<br>
>> >>>>>> response,<br>
>> >>>>>> and response times out, so they "hang".<br>
>> >>>>>><br>
>> >>>>>> Secondly, logs are interesting. If you look at the code from gist<br>
>> >>>>>> once<br>
>> >>>>>> again, you can see, that some of CandidateProvider's don't actually<br>
>> >>>>>> require<br>
>> >>>>>> any networking work, so all they do is actually just logging that<br>
>> >>>>>> they're<br>
>> >>>>>> working (I added this as part of debugging actually) and return<br>
>> >>>>>> pure data.<br>
>> >>>>>> So what's weird is that they timeout also! Here's how output of our<br>
>> >>>>>> logs<br>
>> >>>>>> starts to look like after the bug happens:<br>
>> >>>>>><br>
>> >>>>>> ```<br>
>> >>>>>> [2015-04-22 09:56:20] provider: CandidateProvider1<br>
>> >>>>>> [2015-04-22 09:56:20] provider: CandidateProvider2<br>
>> >>>>>> [2015-04-22 09:56:21] Got timeout while requesting<br>
>> >>>>>> CandidateProvider1<br>
>> >>>>>> [2015-04-22 09:56:21] Got timeout while requesting<br>
>> >>>>>> CandidateProvider2<br>
>> >>>>>> [2015-04-22 09:56:22] provider: CandidateProvider1<br>
>> >>>>>> [2015-04-22 09:56:22] provider: CandidateProvider2<br>
>> >>>>>> [2015-04-22 09:56:23] Got timeout while requesting<br>
>> >>>>>> CandidateProvider1<br>
>> >>>>>> [2015-04-22 09:56:23] Got timeout while requesting<br>
>> >>>>>> CandidateProvider2<br>
>> >>>>>> ... and so on<br>
>> >>>>>> ```<br>
>> >>>>>><br>
>> >>>>>> What's also weird is that, even after timeout is logged, the string<br>
>> >>>>>> ""Got responses!" never gets logged also! So hanging happens<br>
>> >>>>>> somewhere<br>
>> >>>>>> in-between.<br>
>> >>>>>><br>
>> >>>>>> I have to say I'm sorry that I don't have strace output now, I'll<br>
>> >>>>>> have<br>
>> >>>>>> to wait until this situation happens once again, but I'll get later<br>
>> >>>>>> to you<br>
>> >>>>>> with this info.<br>
>> >>>>>><br>
>> >>>>>> So, how is this possible that almost-pure code gets timed-out? And<br>
>> >>>>>> why<br>
>> >>>>>> does it hang afterwards?<br>
>> >>>>>><br>
>> >>>>>> CPU and other resource usage is quite low, number of open<br>
>> >>>>>> file-descriptors also (it seems).<br>
>> >>>>>><br>
>> >>>>>> Thanks for all the suggestions in advance!<br>
>> >>>>>><br>
>> >>>>>> _______________________________________________<br>
>> >>>>>> Haskell-Cafe mailing list<br>
>> >>>>>> <a href="mailto:Haskell-Cafe@haskell.org">Haskell-Cafe@haskell.org</a><br>
>> >>>>>> <a href="http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe" target="_blank">http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe</a><br>
>> >>>>>><br>
>> >>>>><br>
>> >>>>><br>
>> >>>>><br>
>> >>>>> --<br>
>> >>>>> Gregory Collins <<a href="mailto:greg@gregorycollins.net">greg@gregorycollins.net</a>><br>
>> >>>><br>
>> >>>><br>
>> >>><br>
>> >>><br>
>> >>><br>
>> >>> --<br>
>> >>> Gregory Collins <<a href="mailto:greg@gregorycollins.net">greg@gregorycollins.net</a>><br>
>> >><br>
>> >><br>
>> ><br>
>> ><br>
>> > _______________________________________________<br>
>> > Haskell-Cafe mailing list<br>
>> > <a href="mailto:Haskell-Cafe@haskell.org">Haskell-Cafe@haskell.org</a><br>
>> > <a href="http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe" target="_blank">http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe</a><br>
>> ><br>
><br>
><br>
</div></div></blockquote></div><br></div>