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

Re: Bootstrapping Beam's Job Server



On Thu, Aug 23, 2018 at 6:47 AM Maximilian Michels <mxm@xxxxxxxxxx> wrote:
 > Going down this path may start to get fairly involved, with an almost
 > endless list of features that could be requested. Instead, I would
 > suggest we keep process-based execution very simple, and specify bash
 > script (that sets up the environment and whatever else one may want to
 > do) as the command line invocation.

Fair point. At the least, we will have to transfer the shell script to
the nodes. Anything else is up to the script.

That would be another artifact. But this can also be part of host provisioning (i.e. this worker execution model does not perform artifact staging). 


 > I would also think it'd be really valuable to provide a "callback"
 > environment, where an RPC call is made to trigger worker creation
 > (deletion?), passing the requisite parameters (e.g. the fn api
 > endpoints).

Aren't you making up more features now? :) Couldn't this be also handled
by the shell script?

On 23.08.18 14:13, Robert Bradshaw wrote:
> On Thu, Aug 23, 2018 at 1:54 PM Maximilian Michels <mxm@xxxxxxxxxx> wrote:
>>
>> Big +1. Process-based execution should be simple to reason about for
>> users.
>
> +1. In fact, this is exactly what the Python local job server does,
> with running Docker simply being a particular command line that's
> passed down here.
>
> https://github.com/apache/beam/blob/master/sdks/python/apache_beam/runners/portability/local_job_service_main.py
>
>> The implementation should not be too involved. The user has to
>> ensure the environment is suitable for process-based execution.
>>
>> There are some minor features that we should support:
>>
>> - Activating a virtual environment for Python / Adding pre-installed
>> libraries to the classpath
>>
>> - Staging libraries, similarly to the boot code for Docker
>
> Going down this path may start to get fairly involved, with an almost
> endless list of features that could be requested. Instead, I would
> suggest we keep process-based execution very simple, and specify bash
> script (that sets up the environment and whatever else one may want to
> do) as the command line invocation. We could even provide a couple of
> these. (The arguments to pass should be configurable).
>
> I would also think it'd be really valuable to provide a "callback"
> environment, where an RPC call is made to trigger worker creation
> (deletion?), passing the requisite parameters (e.g. the fn api
> endpoints). This could be useful both in a distributed system (where
> it may be desirable for an external entity to actually start up the
> workers) or for debugging/testing (where one could call into the same
> process that submitted the job, which would execute workers on
> separate threads with an already set up environment).
>
>> On 22.08.18 07:49, Henning Rohde wrote:
>>> Agree with Luke. Perhaps something simple, prescriptive yet flexible,
>>> such as custom command line (defined in the environment proto) rooted at
>>> the base of the provided artifacts and either passed the same arguments
>>> or defined in the container contract or made available through
>>> substitution. That way, all the restrictions/assumptions of the
>>> execution environment become implicit and runner/deployment dependent.
>>>
>>>
>>> On Tue, Aug 21, 2018 at 2:12 PM Lukasz Cwik <lcwik@xxxxxxxxxx
>>> <mailto:lcwik@xxxxxxxxxx>> wrote:
>>>
>>>      I believe supporting a simple Process environment makes sense. It
>>>      would be best if we didn't make the Process route solve all the
>>>      problems that Docker solves for us. In my opinion we should limit
>>>      the Process route to assume that the execution environment:
>>>      * has all dependencies and libraries installed
>>>      * is of a compatible machine architecture
>>>      * doesn't require special networking rules to be setup
>>>
>>>      Any other suggestions for reasonable limits on a Process environment?
>>>
>>>      On Tue, Aug 21, 2018 at 2:53 AM Ismaël Mejía <iemejia@xxxxxxxxx
>>>      <mailto:iemejia@xxxxxxxxx>> wrote:
>>>
>>>          It is also worth to mention that apart of the
>>>          testing/development use
>>>          case there is also the case of supporting people running in Hadoop
>>>          distributions. There are two extra reasons to want a process based
>>>          version: (1) Some Hadoop distributions run in machines with
>>>          really old
>>>          kernels where docker support is limited or nonexistent (yes, some of
>>>          those run on kernel 2.6!) and (2) Ops people may be reticent to the
>>>          additional operational overhead of enabling docker in their
>>>          clusters.
>>>          On Tue, Aug 21, 2018 at 11:50 AM Maximilian Michels
>>>          <mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>> wrote:
>>>           >
>>>           > Thanks Henning and Thomas. It looks like
>>>           >
>>>           > a) we want to keep the Docker Job Server Docker container and
>>>          rely on
>>>           > spinning up "sibling" SDK harness containers via the Docker
>>>          socket. This
>>>           > should require little changes to the Runner code.
>>>           >
>>>           > b) have the InProcess SDK harness as an alternative way to
>>>          running user
>>>           > code. This can be done independently of a).
>>>           >
>>>           > Thomas, let's sync today on the InProcess SDK harness. I've
>>>          created a
>>>           > JIRA issue: https://issues.apache.org/jira/browse/BEAM-5187
>>>           >
>>>           > Cheers,
>>>           > Max
>>>           >
>>>           > On 21.08.18 00:35, Thomas Weise wrote:
>>>           > > The original objective was to make test/development easier
>>>          (which I
>>>           > > think is super important for user experience with portable
>>>          runner).
>>>           > >
>>>           > >  From first hand experience I can confirm that dealing with
>>>          Flink
>>>           > > clusters and Docker containers for local setup is a
>>>          significant hurdle
>>>           > > for Python developers.
>>>           > >
>>>           > > To simplify using Flink in embedded mode, the (direct)
>>>          process based SDK
>>>           > > harness would be a good option, especially when it can be
>>>          linked to the
>>>           > > same virtualenv that developers have already setup,
>>>          eliminating extra
>>>           > > packaging/deployment steps.
>>>           > >
>>>           > > Max, I would be interested to sync up on what your thoughts are
>>>           > > regarding that option since you mention you also started to
>>>          work on it
>>>           > > (see previous discussion [1], not sure if there is a JIRA
>>>          for it yet).
>>>           > > Internally we are planning to use a direct SDK harness
>>>          process instead
>>>           > > of Docker containers. For our specific needs it will works
>>>          equally well
>>>           > > for development and production, including future plans to
>>>          deploy Flink
>>>           > > TMs via Kubernetes.
>>>           > >
>>>           > > Thanks,
>>>           > > Thomas
>>>           > >
>>>           > > [1]
>>>           > >
>>>          https://lists.apache.org/thread.html/d8b81e9f74f77d74c8b883cda80fa48efdcaf6ac2ad313c4fe68795a@%3Cdev.beam.apache.org%3E
>>>           > >
>>>           > >
>>>           > >
>>>           > >
>>>           > >
>>>           > >
>>>           > > On Mon, Aug 20, 2018 at 3:00 PM Maximilian Michels
>>>          <mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>
>>>           > > <mailto:mxm@xxxxxxxxxx <mailto:mxm@xxxxxxxxxx>>> wrote:
>>>           > >
>>>           > >     Thanks for your suggestions. Please see below.
>>>           > >
>>>           > >      > Option 3) would be to map in the docker binary and
>>>          socket to allow
>>>           > >      > the containerized Flink job server to start
>>>          "sibling" containers on
>>>           > >      > the host.
>>>           > >
>>>           > >     Do you mean packaging Docker inside the Job Server
>>>          container and
>>>           > >     mounting /var/run/docker.sock from the host inside the
>>>          container? That
>>>           > >     looks like a bit of a hack but for testing it could be
>>>          fine.
>>>           > >
>>>           > >      > notably, if the runner supports auto-scaling or
>>>          similar non-trivial
>>>           > >      > configurations, that would be difficult to manage
>>>          from the SDK side.
>>>           > >
>>>           > >     You're right, it would be unfortunate if the SDK would
>>>          have to deal with
>>>           > >     spinning up SDK harness/backend containers. For non-trivial
>>>           > >     configurations it would probably require an extended
>>>          protocol.
>>>           > >
>>>           > >      > Option 4) We are also thinking about adding process
>>>          based SDKHarness.
>>>           > >      > This will avoid docker in docker scenario.
>>>           > >
>>>           > >     Actually, I had started implementing a process-based
>>>          SDK harness but
>>>           > >     figured it might be impractical because it doubles the
>>>          execution path
>>>           > >     for UDF code and potentially doesn't work with custom
>>>          dependencies.
>>>           > >
>>>           > >      > Process based SDKHarness also has other applications
>>>          and might be
>>>           > >      > desirable in some of the production use cases.
>>>           > >
>>>           > >     True. Some users might want something more lightweight.
>>>           > >
>>>           >
>>>           > --
>>>           > Max
>>>
>>
>> --
>> Max

--
Max