Weighing in late, as usual, but I'm not sure what Leif means by
"ts/mgmt.h". Did you mean "ts/mgmtapi.h"? If so, that's a rather heavy
weight solution for the desired action here. AFAICT using that requires
setting up a full API client, not just making a single call to send a
message. You'd also have the problem that it wouldn't work without
`traffic_manager` which could be problematic for testing.
The main goal here is to address a chronic issue which is the ability for
plugins to shutdown ATS in a standard way, certainly better than calling
"abort()" or "TSAssert(false)".
On Wed, Jun 19, 2019 at 3:04 PM Fei Deng <duke8...@gmail.com> wrote:
> I think it is already logging it right now, leif had a new PR for it just
> recently to remove the Emergency call.
>
> Regards,
> Fei Deng
>
>
> On Wed, Jun 19, 2019 at 11:18 AM Sudheer Vinukonda
> <sudheervinuko...@yahoo.com.invalid> wrote:
>
> > >>> ....while on a config reload it should not call it and fall back to
> > the previous configuration.
> > Is there a way we can alert this (at a minimum, an Error log?)?
> > Falling back to old config (definitely preferable), but, silently is a
> bit
> > misleading and confusing.
> >
> > - Sudheer
> >
> > On Wednesday, June 19, 2019, 9:13:44 AM PDT, Fei Deng <
> > duke8...@gmail.com> wrote:
> >
> > Adding some notes here from the discussion we had just now:
> >
> > - *Fatal* and *Emergency* both need to be exposed to plugins, where
> > *Emergency* keeps traffic server down and *Fatal* will try and restart
> > (just like the core functions).
> > - When reading configs for the first time and fails, traffic server
> > should call *Emergency*, while on a config reload it should not call it
> > and fall back to the previous configuration.
> >
> > If I forgot something we just talked about please add to this note @Alan
> > @Bryan.
> >
> > Regards,
> > Fei Deng
> >
> >
> > On Tue, Jun 18, 2019 at 10:11 AM Fei Deng <duke8...@gmail.com> wrote:
> >
> > > So this is what it looks like now:
> > >
> > > tsapi void
> > > TSEmergency(const char *fmt, ...)
> > > {
> > > va_list args;
> > >
> > > va_start(args, fmt);
> > > EmergencyV(fmt, args);
> > > va_end(args);
> > > }
> > >
> > > Regards,
> > > Fei Deng
> > >
> > >
> > > On Tue, Jun 18, 2019 at 9:41 AM Fei Deng <duke8...@gmail.com> wrote:
> > >
> > >> Yes, this is suggested by Alan and Susan, since sometimes we need
> > plugins
> > >> to be able to abort and not try to restart (say if some important
> files
> > are
> > >> missing). I'll change the name to TSEmergency().
> > >>
> > >> And I'll look at TSDebug and TSError, and make it similar to those.
> > >>
> > >> Regards,
> > >> Fei Deng
> > >>
> > >>
> > >> On Mon, Jun 17, 2019 at 5:59 PM Bryan Call <bc...@apache.org> wrote:
> > >>
> > >>> Looking how Emergency() vs ink_emergency() works. The only
> difference
> > >>> is Emergency() calls Diags::cleanup_func() with is only defined in
> > >>> traffic_manager.
> > >>>
> > >>> Here is the call stack for Emergency():
> > >>> Emergency > DiagsError > diags->error > error_va > ink_emergency_va >
> > >>> ::exit(UNRECOVERABLE_EXIT)
> > >>>
> > >>> Vs
> > >>> ink_emergency > ink_emergency_va > ::exit(UNRECOVERABLE_EXIT)
> > >>>
> > >>> -Bryan
> > >>>
> > >>>
> > >>> > On Jun 17, 2019, at 3:27 PM, Bryan Call <bc...@apache.org> wrote:
> > >>> >
> > >>> > I would assume he is going to be calling ink_emergency(), so it
> > >>> wouldn’t be a management API.
> > >>> >
> > >>> > Fei, can you add more details about what "This API is a wrapper for
> > >>> the "Emergency” call” is really going to do?
> > >>> >
> > >>> > -Bryan
> > >>> >
> > >>> >
> > >>> >> On Jun 17, 2019, at 3:22 PM, Leif Hedstrom <zw...@apache.org>
> > wrote:
> > >>> >>
> > >>> >>
> > >>> >>
> > >>> >>> On Jun 17, 2019, at 4:16 PM, Bryan Call <bc...@apache.org>
> wrote:
> > >>> >>>
> > >>> >>> I would recommend having it like TSDebug and TSError.
> > >>> >>>
> > >>> >>> tsapi void TSEmergency(const char *fmt, ...) TS_PRINTFLIKE(1, 2);
> > >>> >>>
> > >>> >>> TSDebug and TSError:
> > >>> >>> tsapi void TSDebug(const char *tag, const char *format_str, ...)
> > >>> TS_PRINTFLIKE(2, 3);
> > >>> >>> tsapi void TSError(const char *fmt, ...) TS_PRINTFLIKE(1, 2);
> > >>> >>>
> > >>> >>
> > >>> >>
> > >>> >> What I don’t like with the original idea is that we’re putting
> > >>> management APIs (which we have ts/mgmt.h for) into the plugin APIs.
> > There
> > >>> should be nothing stopping a plugin from using the management APIs I
> > don’t
> > >>> think? If there is, that’s yet another issue.
> > >>> >>
> > >>> >> I feel reasonably strongly about not merging ts/hmgmt.h
> > functionality
> > >>> into ts/ts.h.
> > >>> >>
> > >>> >> — Leif
> > >>> >>
> > >>> >>>
> > >>> >>> -Bryan
> > >>> >>>
> > >>> >>>
> > >>> >>>> On Jun 17, 2019, at 2:13 PM, Fei Deng <duke8...@apache.org>
> > wrote:
> > >>> >>>>
> > >>> >>>> void TSEmergencyShutdown(const char *log_msg);
> > >>> >>>>
> > >>> >>>> This API is a wrapper for the "Emergency" call, which signals
> > >>> >>>> traffic_manager to not restart traffic_server after the
> shutdown,
> > >>> i.e. this
> > >>> >>>> call should be used when something has gone wrong and cannot be
> > >>> recovered
> > >>> >>>> by a restart.
> > >>> >>>
> > >>> >>
> > >>> >
> > >>>
> > >>>
>
