The purpose of the spec would be to provide the names, type and descriptions of the options. We don't need anything beyond the JSON types (string, number, bool, object, list) because the only ambiguity we get is how do we parse command line string into the JSON type (and that ambiguity is actually only between string and non-string since all the other JSON types are unambiguous).
Also, I believe the flow would be
1) Parse options
a) Find the URL from args specified and/or additional methods on PipelineOptions that exposes a programmatic way to set the URL during parsing.
b) Query URL for option specs
c) Parse the remainder of the options
2) Construct pipeline
3) Choose runner
4) Submit job to runner
Note that I believe that we will want to have multiple URLs to support cross language pipelines since we will want to be able to ask other SDK languages/versions for their "list" of supported PipelineOptions.
There's two questions here:
(A) What do we do in the short term?
I think adding every runner option to every SDK is not sustainable
(n*m work, assuming every SDK knows about every runner), and having a
patchwork of options that were added as one-offs to SDKs is not
desirable either. Furthermore, it seems difficult to parse unknown
options as if they were valid options, so my preference here would be
to just use a special runner_option flag. (One could also pass a set
of unparsed/unvalidated runner options to the runner, even if they're
not distinguished for the user, and runners (or any intermediates)
could run a "promote" operation that promotes any of these unknowns
that they recognize to real options before further processing. The
parsing would be done as repeated-string, and not be intermingled with
the actually validated options. This is essential a variant of option 1.)
(B) What do do in the long term? While the JobServer approach sounds
nice, I think it introduces a lot of complexity (we have too much of
that already) and still doesn't completely solve the problem. In
particular, it changes the flow from
1. Parse options
2. Construct pipeline
3. Choose runner
4. Submit job to runner
1. Parse options
2. Construct pipeline
3. Choose runner
4a. Query runner for option specs
4b. Re-parse options
4c. Submit job to runner
In particular, doing 4b in the SDK rather than just let the runner
itself do the validation as part of (4) doesn't save much and forces
us to come up with a (probably incomplete) spec as to how to define
options, their types, and their validations. It also means that a
delegating runner must choose and interact with its downstream
runner(s) synchronously, else we haven't actually solved the issue.
For these reasons, I don't think we even want to go with the JobServer
approach in the long term, which has bearing on (A).
On Wed, Nov 7, 2018 at 8:50 PM Maximilian Michels <mxm@xxxxxxxxxx> wrote:
> If the preferred approach is to eventually have the JobServer serve the
> options, then the best intermediate solution is to replicate common
> options in the SDKs.
> If we went down the "--runner_option" path, we would end up with
> multiple ways of specifying the same options. We would eventually have
> to deprecate "runner options" once we have the JobServer approach. I'd
> like to avoid that.
> For the upcoming release we can revert the changes again and add the
> most common missing options to the SDKs. Then hopefully we should have
> fetching implemented for the release after.
> Do you think that is feasible?
> On 30.10.18 23:00, Lukasz Cwik wrote:
> > I still like #3 the most, just can't devote the time to get it done.
> > Instead of going with a fully implemented #3, we could hardcode the a
> > subset of options and types within each SDK until the job server is
> > ready to provide this information and then migrate to the "full" list.
> > This would be an easy path for SDKs to take on. They could "know" of a
> > few well known options, and if they want to support all options, they
> > implement the integration with the job server.
> > On Fri, Oct 26, 2018 at 9:19 AM Maximilian Michels <mxm@xxxxxxxxxx
> > <mailto:mxm@xxxxxxxxxx>> wrote:
> > > I would prefer we don't introduce a (quirky) way of passing
> > unknown options that forces users to type JSON into the command line
> > (or similar acrobatics)
> > Same here, the JSON approach seems technically nice but too bulky
> > for users.
> > > To someone wanting to run a pipeline, all options are equally
> > important, whether they are application specific, SDK specific or
> > runner specific.
> > I'm also reluctant to force users to use `--runner_option=` because the
> > division into "Runner" options and other options seems rather arbitrary
> > to users. Most built-in options are also Runner-related.
> > > It should be possible to *optionally* qualify/scope (to cover
> > cases where there is ambiguity), but otherwise I prefer the format
> > we currently have.
> > Yes, namespacing is a problem. What happens if the user defines a
> > custom
> > PipelineOption which clashes with one of the builtin ones? If both are
> > set, which one is actually applied?
> > Note that PipelineOptions so far has been treating name equality to mean
> > option equality and the Java implementation has a bunch of strict checks
> > to make sure that default values aren't used for duplicate definitions,
> > they have the same type, etc...
> > With 1), you fail the job if the runner can't understand your option
> > because its not represented the same way. User then needs to fix-up
> > their declaration of the option name.
> > With 2), there are no name conflicts, the SDK will need to validate that
> > the option isn't set in both formats and error out if it is before
> > pipeline submission time.
> > With 3), you can prefetch all the options and error out to the user
> > during argument parsing time.
> > Here is a summary of the possible paths going forward:
> > 1) Validate PipelineOptions at Runner side
> > ==========================================
> > The main issue raised here was that we want to move away from parsing
> > arguments which look like options without validating them. An easy fix
> > would be to actually validate them on the Runner side. This could be
> > done by changing the deserialization code of PipelineOptions which so
> > far ignores unknown JSON options.
> > See: PipelineOptionsTranslation.fromProto(Struct protoOptions)
> > Actually, this wouldn't work for user-defined PipelineOptions because
> > they might not be known to the Runner (if they are defined in Python).
> > 2) Introduce a Runner-Option Flag
> > =================================
> > In this approach we would try to add as many pipeline options for a
> > Runner to the SDK, but allow additional Runner options to be passed
> > using the `--runner-option=key=val` flag. The Runner, like in 1), would
> > have to ensure validation. I think this has been the most favored
> > way so
> > far. Going forward, that means that `--parallelism=4` and
> > `--runner-option=parallelism=4` will have the same effect for the Flink
> > Runner.
> > 3) Implement Fetching of Options from JobServer
> > ===============================================
> > The options are retrieved from the JobServer before submitting the
> > pipeline. I think this would be ideal but, as mentioned before, it
> > increases the complexity for implementing new SDKs and might overall
> > just not be worth the effort.
> > What do you think? I'd implement 2) for the next release, unless there
> > are advocates for a different approach.
> > Cheers,
> > Max