Open Stack

Excerpts from Steve Baker's message of 2016-04-20 16:38:25 +1200:
> On 20/04/16 06:17, Monty Taylor wrote:
> > On 04/19/2016 10:16 AM, Daniel P. Berrange wrote:
> >> On Tue, Apr 19, 2016 at 09:57:56AM -0500, Dean Troyer wrote:
> >>> On Tue, Apr 19, 2016 at 9:06 AM, Adam Young <ayoung at redhat.com> wrote:
> >>>
> >>>> I wonder how much of that is Token caching.  In a typical CLI use 
> >>>> patter,
> >>>> a new token is created each time a client is called, with no 
> >>>> passing of a
> >>>> token between services.  Using a session can greatly decrease the 
> >>>> number of
> >>>> round trips to Keystone.
> >>>>
> >>>
> >>> Not as much as you think (or hope?).  Persistent token caching to 
> >>> disk will
> >>> help some, at other expenses though.  Using --timing on OSC will 
> >>> show how
> >>> much time the Identity auth round trip cost.
> >>>
> >>> I don't have current numbers, the last time I instrumented OSC there 
> >>> were
> >>> significant load times for some modules, so we went a good distance to
> >>> lazy-load as much as possible.
> >>>
> >>> What Dan sees WRT a persistent client process, though, is a 
> >>> combination of
> >>> those two things: saving the Python loading and the Keystone round 
> >>> trips.
> >>
> >> The 1.5sec overhead I eliminated doesn't actually have anything todo
> >> with network round trips at all. Even if you turn off all network
> >> services and just run 'openstack <somecmmand>' and let it fail due
> >> to inability to connect it'll still have that 1.5 sec overhead. It
> >> is all related to python runtime loading and work done during module
> >> importing.
> >>
> >> eg run 'unstack.sh' and then compare the main openstack client:
> >>
> >> $ time /usr/bin/openstack server list
> >> Discovering versions from the identity service failed when creating 
> >> the password plugin. Attempting to determine version from URL.
> >> Unable to establish connection to 
> >> http://192.168.122.156:5000/v2.0/tokens
> >>
> >> real    0m1.555s
> >> user    0m1.407s
> >> sys    0m0.147s
> >>
> >> Against my client-as-a-service version:
> >>
> >> $ time $HOME/bin/openstack server list
> >> [Errno 111] Connection refused
> >>
> >> real    0m0.045s
> >> user    0m0.029s
> >> sys    0m0.016s
> >>
> >>
> >> I'm sure there is scope for also optimizing network traffic / round
> >> trips, but I didn't investigate that at all.
> >>
> >>> I have (had!) a version of DevStack that put OSC into a subprocess and
> >>> called it via pipes to do essentially what Dan suggests.  It saves some
> >>> time, at the expense of complexity that may or may not be worth the 
> >>> effort.
> >>
> >> devstack doesn't actually really need any significant changes beyond
> >> making sure $PATH pointed to the replacement client programs and that
> >> the server was running - the latter could be automated as a launch on
> >> demand thing which would limit devstack changes.
> >>
> >> It actually doesn't technically need any devstack change - these
> >> replacement clients could simply be put in some 3rd party git repo
> >> and let developers who want the speed benefit simply put them in
> >> their $PATH before running devstack.
> >>
> >>> One thing missing is any sort of transactional control in the I/O 
> >>> with the
> >>> subprocess, ie, an EOT marker.  I planned to add a -0 option (think 
> >>> xargs)
> >>> to handle that but it's still down a few slots on my priority list.  
> >>> Error
> >>> handling is another problem, and at this point (for DevStack purposes
> >>> anyway) I stopped the investigation, concluding that reliability 
> >>> trumped a
> >>> few seconds saved here.
> >>
> >> For I/O I simply replaced stdout + stderr with a new StringIO handle to
> >> capture the data when running each command, and for error handling I
> >> ensured the exit status was fed back & likewise stderr printed.
> >>
> >> It is more than just a few seconds saved - almost 4 minutes, or
> >> nearly 20% of entire time to run stack.sh on my machine
> >>
> >>
> >>> Ultimately, this is one of the two giant nails in the coffin of 
> >>> continuing
> >>> to persue CLIs in Python.  The other is co-installability. (See that
> >>> current thread on the ML for pain points).  Both are easily solved with
> >>> native-code-generating languages.  Go and Rust are at the top of my
> >>> personal list here...
> >
> > Using entrypoints and plugins in python is slow, so loading them is 
> > slow, as is loading all of the dependent libraries. Those were choices 
> > made for good reason back in the day, but I'm not convinced either are 
> > great anymore.
> >
> > A pluginless CLI that simply used REST calls rather than the 
> > python-clientlibs should be able to launch in get to the business of 
> > doing work in 0.2 seconds - counting time to load and parse 
> > clouds.yaml. That time could be reduced - the time spent in occ 
> > parsing vendor json files is not strictly necessary and certainly 
> > could go faster. It's not as fast as 0.004 seconds, but with very 
> > little effort it's 6x faster.
> >
> > Rather than ditching python for something like go, I'd rather put 
> > together a CLI with no plugins and that only depended on keystoneauth 
> > and os-client-config as libraries. No?
> >
> 
> There is a middle ground between discovering plugins on every run and 
> not allowing plugins at all. There should be significant performance 
> gains by having a plugin registry which caches the result of discovery. 
> The cost is the complexity of managing registry rebuilds when plugins 
> change. This should be mitigated for most users by having registry 
> rebuild hooks triggered by package/pip installs. The needs of the 
> remaining users (mostly developers of plugins) can be handled by a 
> command which rebuilds the user registry which overrides any 
> system-built registry (or by triggering an automatic rebuild by deleting 
> the registry).

Cliff looks for commands on demand. If we modify it's command loader to
support some "built in" commands, and then implement the commands in OSC
that way, we can avoid scanning the real plugin system until we hit a
command that isn't built in.

The last time I proposed that, though, someone (Dean?) pointed out that
loading the plugins wasn't actually where OSC was spending its time. So,
maybe we should profile the thing before proposing code changes.

Doug

> 
> Then startup time is somewhat determined by registry parse time. If 
> python json is not fast enough may be faster options such as ujson:
> http://artem.krylysov.com/blog/2015/09/29/benchmark-python-json-libraries/
> 
> If this were to be implemented it might best live in stevedore.
>