----- On Nov 8, 2019, at 3:49 AM, Benjamin Poirier benjamin.poir...@gmail.com wrote:
> Hi, long time no see! > > Consider a system that's recording trace data continuously from a > userspace application. A problem is noticed and I'd like to investigate > it. What workflow would allow me to view past trace data to analyze the > cause of the problem and view current (live) trace data while working on > fixing the problem, all the while keeping a continuous record of events > (ie. not stopping the trace)? > > I thought of the following approaches but both seem to have > disadvantages: > 1) > Have one continuous tracing session in normal mode. When a problem is > noticed, use lttng-rotate to be able to read old trace data. Start a > second tracing session in live mode with the same event rules to view > live trace data. IIUC, there's no option to prevent lttng-relayd from > writing the traces to disk, so we end up writing two sets of identical > traces to disk during the time we want to look at live trace data. > 2) > Have one tracing session running in live mode all the time. When a > problem is noticed, use the viewer on whatever lttng-relayd has flushed > to disk to access old trace data (session rotation is not available). > Furthermore, this workflow has the disadvantage that the trace data is > going through lttng-relayd all the time (which I guess is less efficient > than writing it from lttng-consumerd) merely to support the rare case > where data needs to be seen live while analyzing a problem. > > Is there some better approach that I didn't think about? What's your > recommendation to support this workflow? > > I'm comparing lttng to the situation where the userspace application is > writing its logs directly to a file. In that case, it's simple to read > the file to access old and "live" data. Of course I realize that lttng > shines in other areas where simple file writing does not ;) Why do you need the "live" mode at all ? What do you intend to do with the live stream that cannot be done through notifiers invoked on rotation completion and use size-based or timer-based rotations ? Thanks, Mathieu > > Thank you, > -Benjamin > _______________________________________________ > lttng-dev mailing list > lttng-dev@lists.lttng.org > https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev -- Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com _______________________________________________ lttng-dev mailing list lttng-dev@lists.lttng.org https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
