[ovs-dev] [PATCH ovn 1/2] Revert "northd: Don't poll ovsdb before the connection is fully established"

Han Zhou hzhou at ovn.org
Fri Sep 17 20:55:29 UTC 2021


On Fri, Sep 17, 2021 at 1:48 PM Renat Nurgaliyev <impleman at gmail.com> wrote:
>
> Hello Han,
>
> when I wrote this patch we had an issue with a very big SB database,
around 1,5 gigabytes. There were no controllers or northds running, so the
database server was without any load at all. Although OVSDB was idling,
even a single northd process could not fully connect to the database due to
its size, since it could not fetch and process the data in 5 seconds.

Hi Renat, thanks for the explanation. However, suppose SB is still huge, if
NB is not that big, the probe config in NB_Global will soon be applied to
ovn-northd, which would probe in proper interval (desired setting with the
SB size considered) instead of the default 5 sec, and it should succeed,
right?

Thanks,
Han

>
> Since then many optimizations were made, and the database size with the
same topology reduced to approximately twenty megabytes, so today I
wouldn't be able to reproduce the problem.
>
> However, I am quite sure that it would still cause troubles with a huge
scale, when SB grows to hundreds of megabytes. With the default timeout of
5 seconds, which is implemented in the same thread that also fetches and
processes data, we make an artificial database size limit, which is not so
obvoius to troubleshoot.
>
> Regards,
> Renat.
>
> Han Zhou <hzhou at ovn.org> schrieb am Fr., 17. Sept. 2021, 23:34:
>>
>>
>>
>> On Thu, Sep 16, 2021 at 8:05 PM Zhen Wang <zhewang at nvidia.com> wrote:
>> >
>> > From: zhen wang <zhewang at nvidia.com>
>> >
>> > This reverts commit 1e59feea933610b28fd4442243162ce35595cfee.
>> > Above commit introduced a bug when muptiple ovn-northd instances work
in HA
>> > mode. If SB leader and active ovn-northd instance got killed by system
power
>> > outage, standby ovn-northd instance would never detect the failure.
>> >
>>
>> Thanks Zhen! I added the Renat and Numan who worked on the reverted
commit to CC, so that they can comment if this is ok.
>>
>> For the commit message, I think it may be decoupled from the HA scenario
that is supposed to be fixed by the other patch in this series. The issue
this patch fixes is that before the initial NB downloading is complete the
northd will not send probe, so if the DB server is down (ungracefully)
before the northd reads the NB_Global options, the northd would never
probe, thus never reconnect to the new leader. (it is related to RAFT, but
whether it is multiple northds is irrelevant)
>>
>> As to the original commit that is reverted by this one:
>>
>>     northd: Don't poll ovsdb before the connection is fully established
>>
>>     Set initial SB and NB DBs probe interval to 0 to avoid connection
>>     flapping.
>>
>>     Before configured in northd_probe_interval value is actually applied
>>     to southbound and northbound database connections, both connections
>>     must be fully established, otherwise ovnnb_db_run() will return
>>     without retrieving configuration data from northbound DB. In cases
>>     when southbound database is big enough, default interval of 5 seconds
>>     will kill and retry the connection before it is fully established, no
>>     matter what is set in northd_probe_interval. Client reconnect will
>>     cause even more load to ovsdb-server and cause cascade effect, so
>>     northd can never stabilise. We have more than 2000 ports in our lab,
>>     and northd could not start before this patch, holding at 100% CPU
>>     utilisation both itself and ovsdb-server.
>>
>>     After connections are established, any value in
northd_probe_interval,
>>     or default DEFAULT_PROBE_INTERVAL_MSEC is applied correctly.
>>
>> I am not sure how would the commit help. There are at most 3 - 5 northds
(in practice), and suppose there are tens or hundreds of ovn-controllers
that makes SB busy, it is just 3 - 5 more clients retrying reconnect SB for
several times, and if NB is not that busy (most likely), these northd
clients should get the proper probe settings applied soon without causing
more issues at all. So I don't think the default probe 5 sec would cause
cascade effect for the initial period. @Renat @Numan please correct me if I
am wrong.
>>
>> Thanks,
>> Han
>>
>> > Signed-off-by: zhen wang <zhewang at nvidia.com>
>> > ---
>> >  northd/northd.c | 4 ++--
>> >  1 file changed, 2 insertions(+), 2 deletions(-)
>> >
>> > diff --git a/northd/northd.c b/northd/northd.c
>> > index 688a6e4ef..b7e64470f 100644
>> > --- a/northd/northd.c
>> > +++ b/northd/northd.c
>> > @@ -74,8 +74,8 @@ static bool use_ct_inv_match = true;
>> >
>> >  /* Default probe interval for NB and SB DB connections. */
>> >  #define DEFAULT_PROBE_INTERVAL_MSEC 5000
>> > -static int northd_probe_interval_nb = 0;
>> > -static int northd_probe_interval_sb = 0;
>> > +static int northd_probe_interval_nb = DEFAULT_PROBE_INTERVAL_MSEC;
>> > +static int northd_probe_interval_sb = DEFAULT_PROBE_INTERVAL_MSEC;
>> >  #define MAX_OVN_TAGS 4096
>> >
>> >  /* Pipeline stages. */
>> > --
>> > 2.20.1
>> >


More information about the dev mailing list