Investigating Kallithea performance issues

Thu Mar 23 05:26:27 UTC 2017

Hi,

> On Tue, Mar 21, 2017 at 11:11 AM, Adi Kriegisch <adi at cg.tuwien.ac.at> wrote:
>>
>> Hi!
>>
>> > Sometimes the performance of our Kallithea instance takes a hit, such that
>> > page loading can take many many seconds, and it is unclear what causes this.
>> >
>> > Could you suggest ways to investigate this?
>> Step 1: get data. :-)
>>
>> Actually one of the reasons I use uwsgi for deployment is that I get execution
>> times 'for free': 'generated 309 bytes in 268 msecs' or 'generated 436 bytes
>> in 56 msecs' including the uri and all kinds of useful information.
>>
> That kind of monitoring is indeed what we do now, with uwsgi:
> {address space usage: 1769336832 bytes/1687MB} {rss usage: 1298124800 bytes/1237MB} [pid: 52669|app: 0|req: 349/9564] 1
> 35.252.28.229 () {40 vars in 1686 bytes} [Wed Mar 22 07:39:38 2017] GET /review/ext/cvpsw-review/pull-request/15006/_/c
> vp80 => generated 76111 bytes in 924 msecs (HTTP/1.1 200) 5 headers in 149 bytes (1 switches on core 0)
> {address space usage: 1687105536 bytes/1608MB} {rss usage: 1218387968 bytes/1161MB} [pid: 49555|app: 0|req: 209/9565]
> 35.252.28.229 () {42 vars in 1591 bytes} [Wed Mar 22 07:39:40 2017] GET /js/graph.js?ver=0.2.1
> => generated 0 bytes in 156 msecs (HTTP/1.1 304) 3 headers in 98 bytes (1 switches on core 0)
> 2017-03-22 07:40:09.168 INFO  [kallithea.RequestWrapper] IP: 135.224.206.40 Request to /review/ms/sw-review/branches-tags time: 66.745s
>
> But it are these kind of peeks we want to understand on what Kallithea was exactly waiting upon, is it pure disk/io, is it a qeury to the database that took long (postgresql with monitoring with pg_activity)
> Is there anything Kallithea debug info or pythong profiler can provide us (without overloading kallithea itself with all the debug) that we today are not aware of?
>
> Our database is 'only' 270MB big, so besides inefficient queries, I cannot understand if it would be database access that is causing hickups of more then a minute. Either we configured our database wrong, or I would think postgresql can keep 270MB in memory (we might need to force it to keep it in memory, even if other processes consume alot of memory), the server has 125GB of memory, with a 'free' count (even with caches) of 20GB (let me be clear, the sever is used for other applications as well).
>
> What I do see as queries running long on the database are these, (But I haven't correlated these directly with the above peeks yet (but it could well be).)
> 47362  kallithea            kallithea        127.0.0.1    0.0  0.0    0.00B    0.00B  00:47.10  N    N  SELECT cache_invalidation.cache_id AS cache_invalidation_cache_id, cache_invalidation.cache_key AS cache_invalidation_cache_key, cache_invalidation.cache_args AS cache_invalidation_cache_args, cache_invalidation.cache_active AS cache_invalidation_cache_active FROM cache_invalidation WHERE cache_invalidation.cache_key = 'devws048-32241review/ms/sw-review'
>
> As you see, that query runs for 47 seconds at the moment I catch it.
>
>  Anybody who can explain what this 'triggers' in the database, as that table itself is quite small.
>  public | cache_invalidation                                              | table    | kallithea | 992 kB     |
>  public | cache_invalidation_cache_id_seq                                 | sequence | kallithea | 8192 bytes |
>
> I haven' looked googled yet what this could mean, so forgive me if I'm asking basic questions ;-)

As far as I understand it now, these queries are the result of updated
repositories from 'outside' kallithea and the accompanied repo-'cache
invalidation' trigger.
If this is really hitting us that would mean:
* requests to kallithea webgui have to wait until these queries are over?
* we need to find a way to have these repos updated without the need
of the cache invalidation, so instead of doing an 'outside pull',
would a 'push to kallithea' prevent the need of invalidating the
cache?

Br,

Jan

>
>
>> My setup is similar to this:
>> https://lists.sfconservancy.org/pipermail/kallithea-general/2015q1/000130.html
>>
>> You may even monitor execution times or graph them or do some statistics work
>> on them or -- simply put -- correlate higher latency and execution time with
>> system or I/O load (even simple tools like sysstat/sadc/sar are sufficient to
>> gain insights in this regard).
>
>
> I will already have a look at these simple ones, I must admit I didn't use these until now (I did use iowait/top/...)
> http://www.thegeekstuff.com/2011/03/sar-examples/?utm_source=feedburner
>
>
>>
>>
>> Hope, this helps...
>>
>> All the best,
>>         Adi