[Haskell-cafe] Timeout on pure code

[Kostiantyn Rybnikov]

Hi!

Our company's main commercial product is a Snap-based web app which we compile with GHC 7.8.4. It works on four app-servers currently load-balanced behind Haproxy.

I recently implemented a new piece of functionality, which led to weird behavior which I have no idea how to debug, so I'm asking here for help and ideas!

The new functionality is this: on specific url-handler, we need to query n external services concurrently with a timeout, gather and render results. Easy (in Haskell)!

The implementation looks, as you might imagine, something like this (sorry for almost-real-haskell, I'm sure I forgot tons of imports and other things, but I hope everything is clear as-is, if not -- I'll be glad to update gist to make things more specific):

https://gist.github.com/k-bx/0cf7035aaf1ad6306e76

Now, this works wonderful for some time, and in logs I can see both, successful fetches of external-content, and also lots of timeouts from our external providers. Life is good.

But! After several days of work (sometimes a day, sometimes couple days), apps on all 4 servers go crazy. It might take some interval (like 20 minutes) before they're all crazy, so it's not super-synchronous. Now: how crazy, exactly?

First of all, this endpoint timeouts. Haproxy requests for a response, and response times out, so they "hang".

Secondly, logs are interesting. If you look at the code from gist once again, you can see, that some of CandidateProvider's don't actually require any networking work, so all they do is actually just logging that they're working (I added this as part of debugging actually) and return pure data. So what's weird is that they timeout also! Here's how output of our logs starts to look like after the bug happens:

```

[2015-04-22 09:56:20] provider: CandidateProvider1

[2015-04-22 09:56:20] provider: CandidateProvider2

[2015-04-22 09:56:21] Got timeout while requesting CandidateProvider1

[2015-04-22 09:56:21] Got timeout while requesting CandidateProvider2

[2015-04-22 09:56:22] provider: CandidateProvider1

[2015-04-22 09:56:22] provider: CandidateProvider2

[2015-04-22 09:56:23] Got timeout while requesting CandidateProvider1

[2015-04-22 09:56:23] Got timeout while requesting CandidateProvider2

... and so on

```

What's also weird is that, even after timeout is logged, the string ""Got responses!" never gets logged also! So hanging happens somewhere in-between.

I have to say I'm sorry that I don't have strace output now, I'll have to wait until this situation happens once again, but I'll get later to you with this info.

So, how is this possible that almost-pure code gets timed-out? And why does it hang afterwards?

CPU and other resource usage is quite low, number of open file-descriptors also (it seems).

Thanks for all the suggestions in advance!

[Gregory Collins]

Given your gist, the timeout on your requests is set to a half-second so it's conceivable that a highly-loaded server might have GC pause times approaching that long. Smells to me like a classic Haskell memory leak (that's why the problem occurs after the server has been up for a while): run your program with the heap profiler, and audit any shared tables/IORefs/MVars to make sure you are not building up thunks there.

Greg

_______________________________________________
Haskell-Cafe mailing list
Haskel...@haskell.org
http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe

--

Gregory Collins <gr...@gregorycollins.net>

[Kostiantyn Rybnikov]

Gregory,

Servers are far from being highly-overloaded, since they're currently under a much less load they used to be. Memory consumption is stable and low, and there's a lot of free RAM also.

Would you say that given these factors this scenario is unlikely?

[Niklas Hambüchen]

You might already have considered it, but the GHC eventlog together with
Threadscope and ghc-events-analyze
(http://www.well-typed.com/blog/2014/02/ghc-events-analyze/) can be very
helpful to debug such issues.

[Kostiantyn Rybnikov]

Niklas,

This seems a very helpful tool indeed, but I'm not completely sure how could it be helpful in for specific problem. Was there anything specific you would suggest to measure with it, or just consider it as a general nice tool while approaching the problem?

Thanks!

[Niklas Hambüchen]

I meant in the general sense of approaching the problem, and the linked blog post takes as an example situation a web server that doesn't respond well after some amount of time, I associated that with your problem description.

_______________________________________________

[Gregory Collins]

Maybe but it would be helpful to rule the scenario out. Johan's ekg library is also useful, it exports a webserver on a different port that you can use to track metrics like gc times, etc.

Other options for further debugging include gathering strace logs from the binary. You'll have to do some data gathering to narrow down the cause unfortunately -- http client? your code? Snap server? GHC event manager (System.timeout is implemented here)? GC? etc

G

--

Gregory Collins <gr...@gregorycollins.net>

[Kostiantyn Rybnikov]

All right, good news!

After adding ekg, gathering its data via bosun and seeing nothing useful I actually figured out that I could try harder to reproduce issue by myself instead of waiting for users to do that. And I succeeded! :)

So, after launching 20 infinite curl loops to that handler's url I was quickly able to reproduce the issue, so the task seems clear now: keep reducing the code, reproduce locally, possibly without external services etc. I'll write up after I get to something.

Thanks.

[Kostiantyn Rybnikov]

An update for everyone interested (and not). Turned out it's neither GHC RTS, Snap or networking issues, it's hslogger being very slow. I thought it's slow when used concurrently, but just did a test when it writes 2000 5kb messages sequentially and that finishes in 111 seconds (while minimal program that writes same 2000 messages finishes in 0.12s).

I hope I'll have a chance to investigate why hslogger is so slow in future, but meanwhile will just remove logging.

[Mikhail Glushenkov]

Hi,

On 24 April 2015 at 21:25, Kostiantyn Rybnikov <k-...@k-bx.com> wrote:
> I hope I'll have a chance to investigate why hslogger is so slow in future,
> but meanwhile will just remove logging.

Have you considered using fast-logger?

[Kostiantyn Rybnikov]

Well I do now :) I'll also check out new but interesting haskell-logger [0]

I just think that it would be a good idea to investigate hslogger because it seems that a lot of people are using it already, so providing an update that would upgrade their performance with no additional work would make them happy.

[0]: https://github.com/wdanilo/haskell-logger

[David Turner]

Hi,

I've had a look at this as we use hslogger too, so I'm keen to avoid
this kind of performance issue. I threw a quick Criterion benchmark
together:

https://gist.github.com/DaveCTurner/f977123b4498c4c64569

The headline result on my test machine are that each log call takes
~540us, so 2000 should take about a second. Would be interested if you
could run the same benchmark on your setup as it's possible that
there's something else downstream that's causing you a problem.

A couple of things that might be worth bearing in mind: if you're
talking to syslog over /dev/log then that can block if the log daemon
falls behind: unix datagram sockets don't drop datagrams when they're
congested. If the /dev/log test is slow but the UDP test is fast then
it could be that your syslog can't handle the load.

I'm using rsyslogd and have enabled the feature that combines
identical messages, so this test doesn't generate much disk IO and it
keeps up easily, so the UDP and /dev/log tests run about equally fast
for me. Is your syslog writing out every message? It may be flushing
to disk after every message too, which would be terribly slow.

If you're not logging to syslog, what's your hslogger config?

Cheers,

David

[Kostiantyn Rybnikov]

Hi David.

I planned to create a detailed bug-report at hslogger's issues to start investigation there (as a better place) on Monday, but since I have code prepared already, it's easy to share it right now: https://gist.github.com/k-bx/ccf6fd1c73680c8a4345

I'm launching it as:

    time ./dist/build/seq/seq &> /dev/null

We don't use syslog driver, instead we have a separate file-to-syslog worker to decouple these components.

[David Turner]

I see.

The issue seems to be the default handler which writes the log to
stderr. Replacing 'addHandler h' with 'setHandlers [h]' makes it run
in a reasonable time as 'setHandlers' starts afresh; conversely it's
still slow even if you just use the default handler on its own, i.e.
removing the call to 'addHandler'.

I was a bit suspicious about the 'hFlush' in System.Log.Handler.Simple
but removing that didn't help. However, suspecting that the issue may
be too-much-flushing I eventually found
https://ghc.haskell.org/trac/ghc/ticket/7418 which says that writing
to stderr is slow because its default buffering mode is NoBuffering. I
added 'hSetBuffering stderr LineBuffering' and boom, it runs at a
sensible speed.

Recommend you get rid of writing to stderr and just log to the file
unless you've a good reason to send the output both ways, in which
case switch to LineBuffering as above.

If you're using rsyslog you also may be interested to read
http://www.rsyslog.com/doc/queues.html - this describes how it can
decouple the final log output from the input using a combination of
in-memory and on-disk buffers, and even discard lower-priority
messages from the queue if the going gets really tough. We judged it a
lot of effort to implement our own on-disk spool and were particularly
worried about it growing without bound if the downstream was too slow,
so this feature of rsyslog was just what we needed.

Hope that helps,

David

[Kostiantyn Rybnikov]

David, thank you very much. I confirm everything you wrote, both doing LineBuffering and removing output to stderr resolves the problem.

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).

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.

Anyway, thank you once again.