git.net

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Openstack] Can any user add or delete OpenStack Swift middleware?


Kota,

Sorry for the late response, see more below:

On Fri, Sep 21, 2018 at 2:59 AM Kota TSUYUZAKI
<tsuyuzaki.kota at lab.ntt.co.jp> wrote:
>
> Hi Qiao,
>
> > Thanks! I'm interested and would like to join, as well as contribute!
> >
>
> One example, that is how the multi-READ works, is around [1], the storlets middleware can make a subrequest against to the backend Swift then,
> attach the request input to the application in the Docker container by passing the file descriptor to be readable[2][3][4]*.
> After all of the prepartion for the invocation, the input descriptors will be readable in the storlet app as the InputFile.
>
> * At first, the runtime prepares the extra source stub at [2], then creates a set of pipes for each sources to be communicated with the app
> inside the docker daemon[3], then, the runtime module reads the extra data from Swift GET and flushes all buffers into the descriptor [4].
>
> 1: https://github.com/openstack/storlets/blob/master/storlets/swift_middleware/handlers/proxy.py#L294-L305
> 2: https://github.com/openstack/storlets/blob/master/storlets/gateway/gateways/docker/runtime.py#L571-L581
> 3: https://github.com/openstack/storlets/blob/master/storlets/gateway/gateways/docker/runtime.py#L665-L666
> 4: https://github.com/openstack/storlets/blob/master/storlets/gateway/gateways/docker/runtime.py#L833-L840
>
> Following the mechanism, IMO, what we can do to enable multi-out is
>
> - add the capability to create some PUT subrequest at swift_middleware module (define the new API header too)
> - create the extra communication write-able fds in the storlets runtime (perhaps, storlets daemon is also needed to be changed)
> - pass all data from the write-able fds as to the sub PUT request input
>
>
> If you have any nice idea rather than me, it's always welcome tho :)

I think your approach is clear and straightforward. One quick question:
> - create the extra communication write-able fds in the storlets runtime (perhaps, storlets daemon is also needed to be changed)
So the Storlet app will write to those fds? Are these fds temporary
and need to be destroyed after PUT requests in step 3?

>
> > Another potential use case: imagine I want to compress objects upon
> > PUTs using two different algorithms X and Y, and use the future
> > 'multi-write' feature to store three objects upon any single PUT
> > (original copy, X-compressed copy and Y-compressed copy). I can
> > install two Storlets which implement X and Y respectively. However,
> > seems Storlets engine can only invoke one per PUT, so this is still
> > not feasible. Is that correct?
> >
>
> It sounds interesting. As you know, yes, one Storlet application can be invoked per one PUT.
> On the other hand, Storlets has been capable to run several applications as you want.
> One idea using the capability, if you develop an application like:
>
> - Storlet app invokes several multi threads with their output descriptor
> - Storlet app reads the input stream, then pushes the data into the threads
> - Each thread performs as you want (one does as X compression, the other does as Y compression)
>   then, writes its own result to the output descriptor
>
> It might work for your use case.

Sounds great, I guess it should work as well.

I'm also concerned with "performance isolation" in Storlets. For
instance, is it possible for a user to launch several very
heavy-loaded Storlets apps to consume lots of CPU/memory resources to
affect other users? Does Storlets do performance/resource isolation?

Thanks,
Qiao

>
> Thanks,
> Kota
>
>
> (2018/09/19 5:52), Qiao Kang wrote:
> > Dear Kota,
> >
> > On Mon, Sep 17, 2018 at 11:43 PM Kota TSUYUZAKI
> > <tsuyuzaki.kota at lab.ntt.co.jp> wrote:
> >>
> >> Hi Quio,
> >>
> >>> I know Storlets can provide user-defined computation functionalities,
> >>> but I guess some capabilities can only be achieved using middleware.
> >>> For example, a user may want such a feature: upon each PUT request, it
> >>> creates a compressed copy of the object and stores both the original
> >>> copy and compressed copy. It's feasible using middlware but I don't
> >>> think Storlets provide such capability.
> >>
> >> Interesting, exactly currently it's not supported to write to multi objects for a PUT request but as well as other middlewares we could adopt the feasibility into Storlets if you prefer.
> >> Right now, the multi read (i.e. GET from multi sources) is only available and I think we would be able to expand the logic to PUT requests too. IIRC, in those days, we had discussion on sort of the
> >> multi-out use cases and I'm sure the data structure inside Storlets are designed to be capable to that expantion. At that time, we called them "Tee" application on Storlets, I could not find the
> >> historical discussion logs about how to implement tho, sorry. I believe that would be an use case for storlets if you prefer the user-defined application flexibilities rather than operator defined
> >> Swift middleware.
> >>
> >> The example of multi-read (GET from multi sources) are here:
> >> https://github.com/openstack/storlets/blob/master/tests/functional/python/test_multiinput_storlet.py
> >>
> >> And if you like to try to write multi write, please join us, I'm happy to help you anytime.
> >>
> >
> > Thanks! I'm interested and would like to join, as well as contribute!
> >
> > Another potential use case: imagine I want to compress objects upon
> > PUTs using two different algorithms X and Y, and use the future
> > 'multi-write' feature to store three objects upon any single PUT
> > (original copy, X-compressed copy and Y-compressed copy). I can
> > install two Storlets which implement X and Y respectively. However,
> > seems Storlets engine can only invoke one per PUT, so this is still
> > not feasible. Is that correct?
> >
> >>
> >>> Another example is that a user may want to install a Swift3-like
> >>> middleware to provide APIs to a 3rd party, but she doesn't want other
> >>> users to see this middleware.
> >>>
> >>
> >> If the definition can be made by operators, perhaps one possible solution that preparing different proxy-server endpoint for different users is available. i.e. an user uses no-s3api available proxy,
> >> then the others use a different proxy-server endpoint that has the s3api in the pipeline.
> >>
> >> Or, it sounds like kinda defaulter middleware[1], I don't think it has the scope turning on/off the middlewares for now.
> >>
> >> 1: https://review.openstack.org/#/c/342857/
> >
> > I see, thanks for pointing out the defaulter project!
> >
> > Best,
> > Qiao
> >
> >>
> >> Best,
> >> Kota
> >>
> >> (2018/09/18 11:34), Qiao Kang wrote:
> >>> Kota,
> >>>
> >>> Thanks for your reply, very helpful!
> >>>
> >>> I know Storlets can provide user-defined computation functionalities,
> >>> but I guess some capabilities can only be achieved using middleware.
> >>> For example, a user may want such a feature: upon each PUT request, it
> >>> creates a compressed copy of the object and stores both the original
> >>> copy and compressed copy. It's feasible using middlware but I don't
> >>> think Storlets provide such capability.
> >>>
> >>> Another example is that a user may want to install a Swift3-like
> >>> middleware to provide APIs to a 3rd party, but she doesn't want other
> >>> users to see this middleware.
> >>>
> >>> Regards,
> >>> Qiao
> >>>
> >>> On Mon, Sep 17, 2018 at 9:19 PM Kota TSUYUZAKI
> >>> <tsuyuzaki.kota at lab.ntt.co.jp> wrote:
> >>>>
> >>>> With Storlets, users will be able to create their own applications that are able to run like as a Swift middeleware. The application (currently Python and Java are supported as the language but the
> >>>> apps can calls any binaries in the workspace) can be uploaded as a Swift object, then, users can invoke them with just an extra header that specifies your apps.
> >>>>
> >>>> To fit your own use case, we may have to consider to invole or to integrate the system for you but I believe Storlets could be a choice for you.
> >>>>
> >>>> In detail, Storlets documantation is around there,
> >>>>
> >>>> Top Level Index: https://docs.openstack.org/storlets/latest/index.html
> >>>> System Overview: https://docs.openstack.org/storlets/latest/storlet_engine_overview.html
> >>>> APIs: https://docs.openstack.org/storlets/latest/api/overview_api.html
> >>>>
> >>>> Thanks,
> >>>>
> >>>> Kota
> >>>>
> >>>> (2018/09/17 8:59), John Dickinson wrote:
> >>>>> You may be interested in Storlets. It's another OpenStack project, maintained by a Swift core reviewer, that provides this sort of user-defined middleware functionality.
> >>>>>
> >>>>> You can also ask about it in #openstack-swift
> >>>>>
> >>>>> --John
> >>>>>
> >>>>>
> >>>>>
> >>>>> On 16 Sep 2018, at 9:25, Qiao Kang wrote:
> >>>>>
> >>>>>> Hi,
> >>>>>>
> >>>>>> I'm wondering whether Swift allows any user (not the administrator) to
> >>>>>> specify which middleware that she/he wants his data object to go throught.
> >>>>>> For instance, Alice wants to install a middleware but doesn't want Bob to
> >>>>>> use it, where Alice and Bob are two accounts in a single Swift cluster.
> >>>>>>
> >>>>>> Or maybe all middlewares are pre-installed globally and cannot be
> >>>>>> customized on a per-account basis?
> >>>>>>
> >>>>>> Thanks,
> >>>>>> Qiao
> >>>>>> _______________________________________________
> >>>>>> Mailing list: http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
> >>>>>> Post to     : openstack at lists.openstack.org
> >>>>>> Unsubscribe : http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
> >>>>>
> >>>>> _______________________________________________
> >>>>> Mailing list: http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
> >>>>> Post to     : openstack at lists.openstack.org
> >>>>> Unsubscribe : http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
> >>>>
> >>>>
> >>>> --
> >>>> ----------------------------------------------------------
> >>>> Kota Tsuyuzaki(é?²ï¨?ã??浩太)  <tsuyuzaki.kota at lab.ntt.co.jp>
> >>>> NTT Software Innovation Center
> >>>> Distributed Computing Technology Project
> >>>> Phone  0422-59-2837
> >>>> Fax    0422-59-2965
> >>>> -----------------------------------------------------------
> >>>>
> >>>>
> >>>> _______________________________________________
> >>>> Mailing list: http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
> >>>> Post to     : openstack at lists.openstack.org
> >>>> Unsubscribe : http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
> >>>
> >>>
> >>
> >>
> >> --
> >> ----------------------------------------------------------
> >> Kota Tsuyuzaki(é?²ï¨?ã??浩太)  <tsuyuzaki.kota at lab.ntt.co.jp>
> >> NTT Software Innovation Center
> >> Distributed Computing Technology Project
> >> Phone  0422-59-2837
> >> Fax    0422-59-2965
> >> -----------------------------------------------------------
> >>
> >
> >
>
>
>