[lttng-dev] lost trace events

[Jonathan Rajotte-Julien]

Hi,

On Wed, Oct 31, 2018 at 02:58:37PM -0700, Alok Priyadarshi wrote:
> We are tracing a multi-process application that generates high-frequency
> user-space trace events.
> 
> Environment:
> - Ubuntu 16.04 LTS
> - Lttng 2.10.4 - KeKriek
> 
> Session:
> - Snapshot mode
> - One channel with default config:
> overwrite=1, subbuf_size=524288, num_subbuf=4
> 
> We have noticed that a few events are missing. I understand that in

How did you notice those "missing" events?

Would you be able to share a small reproducer with us so that we can assess the
situation?

> overwrite mode we can lose entire sub-buffers but IIUC we would only lose
> events at the beginning, not in-between events in a sub-buffer.
> 
> Is it not true that only the oldest sub-buffer is overwritten and when
> overwrite happens, none of the latest events are discarded?

This is how it should work yes.

> 
> Few more things to note:
> - Before lttng-snapshot-record, we stop tracing as recommended by the
> documentation

Good.

> - Packet context field events_discarded is always set to 0
> - Event loss seems to happen around packet transition. I have noticed this
> in at least two traces

Please provide us with how you arrived to this conclusion and a small
reproducer if possible.

> 
> The snapshot mode seems perfect for our use case. It is OK to lose events
> from either ends of the snapshot, but it is essential to not lose any
> in-between events. I was under the impression that overwrite mode
> guarantees that.

It should.

Cheers

-- 
Jonathan Rajotte-Julien
EfficiOS