I wonder if we could do this _only_ over the FnApi. The FnApi already does
batching I believe. What if we made schemas a fundamental part of our
protos, and had no SchemaCoder. The FnApi could then batch up a bunch of
rows an encode using Arrow before sending over the wire to the harness.
Of course this still turns back into individual records before it goes back
to user code. However well-known combiners can be executed directly in the
harness, which means aggregations like "sum a field" can be run inside the
harness over the columnar data. Moving these combiners into the harness
might itself be a huge perf benefit for Python, as we could then run them
in a more-performant language.
Reuven
On Tue, Jan 8, 2019 at 7:44 AM Robert Bradshaw <rober...@google.com> wrote:
> On Tue, Jan 8, 2019 at 4:32 PM Reuven Lax <re...@google.com> wrote:
> >
> > I agree with this, but I think it's a significant rethinking of Beam
> that I didn't want to couple to schemas. In addition to rethinking the API,
> it might also require rethinking all of our runners.
>
> We're already marshaling (including batching) data over the FnApi, so
> it might not be that big of a change. Also, the choice of encoding
> over the data channel is already parametrizable via a coder, so it's
> easy to make this an optional feature that runners and SDKs can opt
> into. I agree that we don't want to couple it to schemas (though
> that's where it becomes even more useful).
>
> > Also while columnar can be a large perf win, I suspect that we currently
> have lower-hanging fruit to optimize when it comes to performance.
>
> It's probably a bigger win for Python than for Java.
>
> >
> > Reuven
> >
> > On Tue, Jan 8, 2019 at 5:25 AM Robert Bradshaw <rober...@google.com>
> wrote:
> >>
> >> On Fri, Jan 4, 2019 at 12:54 AM Reuven Lax <re...@google.com> wrote:
> >> >
> >> > I looked at Apache Arrow as a potential serialization format for Row
> coders. At the time it didn't seem a perfect fit - Beam's programming model
> is record-at-a-time, and Arrow is optimized for large batches of records
> (while Beam has a concept of "bundles" they are completely non
> deterministic, and records might bundle different on retry). You could use
> Arrow with single-record batches, but I suspect that would end up adding a
> lot of extra overhead. That being said, I think it's still something worth
> investigating further.
> >>
> >> Though Beam's model is row-oriented, I think it'd make a lot of sense
> >> to support column-oriented transfer of data across the data plane
> >> (we're already concatenating serialized records lengthwise), with
> >> Arrow as a first candidate, and (either as part of the public API or
> >> as an implementation detail) columnar processing as well (e.g.
> >> projections, maps, filters, and aggregations can often be done more
> >> efficiently in a columnar fashion). While this is often a significant
> >> win in C++ (and presumably Java), it's essential for doing
> >> high-performance computing in Python (e.g. Numpy, SciPy, Pandas,
> >> Tensorflow, ... all have batch-oriented APIs and avoid representing
> >> records as individual objects, something we'll need to tackle for
> >> BeamPython at least).
> >>
> >> >
> >> > Reuven
> >> >
> >> >
> >> >
> >> > On Fri, Jan 4, 2019 at 12:34 AM Gleb Kanterov <g...@spotify.com>
> wrote:
> >> >>
> >> >> Reuven, it sounds great. I see there is a similar thing to Row
> coders happening in Apache Arrow, and there is a similarity between Apache
> Arrow Flight and data exchange service in portability. How do you see these
> two things relate to each other in the long term?
> >> >>
> >> >> On Fri, Jan 4, 2019 at 12:13 AM Reuven Lax <re...@google.com> wrote:
> >> >>>
> >> >>> The biggest advantage is actually readability and usability. A
> secondary advantage is that it means that Go will be able to interact
> seamlessly with BeamSQL, which would be a big win for Go.
> >> >>>
> >> >>> A schema is basically a way of saying that a record has a specific
> set of (possibly nested, possibly repeated) fields. So for instance let's
> say that the user's type is a struct with fields named user, country,
> purchaseCost. This allows us to provide transforms that operate on field
> names. Some example (using the Java API):
> >> >>>
> >> >>> PCollection users = events.apply(Select.fields("user")); // Select
> out only the user field.
> >> >>>
> >> >>> PCollection joinedEvents =
> queries.apply(Join.innerJoin(clicks).byFields("user")); // Join two
> PCollections by user.
> >> >>>
> >> >>> // For each country, calculate the total purchase cost as well as
> the top 10 purchases.
> >> >>> // A new schema is created containing fields total_cost and
> top_purchases, and rows are created with the aggregation results.
> >> >>> PCollection purchaseStatistics = events.apply(
> >> >>> Group.byFieldNames("country")
> >> >>> .aggregateField("purchaseCost", Sum.ofLongs(),
> "total_cost"))
> >> >>> .aggregateField("purchaseCost",
> Top.largestLongs(10), "top_purchases"))
> >> >>>
> >> >>>
> >> >>> This is far more readable than what we have today, and what unlocks
> this is that Beam actually knows the structure of the record instead of
> assuming records are uncrackable blobs.
> >> >>>
> >> >>> Note that a coder is basically a special case of a schema that has
> a single field.
> >> >>>
> >> >>> In BeamJava we have a SchemaRegistry which knows how to turn user
> types into schemas. We use reflection to analyze many user types (e.g.
> simple POJO structs, JavaBean classes, Avro records, protocol buffers,
> etc.) to determine the schema, however this is done only when the graph is
> initially generated. We do use code generation (in Java we do bytecode
> generation) to make this somewhat more efficient. I'm willing to bet that
> the code generator you've written for structs could be very easily modified
> for schemas instead, so it would not be wasted work if we went with schemas.
> >> >>>
> >> >>> One of the things I'm working on now is documenting Beam schemas.
> They are already very powerful and useful, but since there is still nothing
> in our documentation about them, they are not yet widely used. I expect to
> finish draft documentation by the end of January.
> >> >>>
> >> >>> Reuven
> >> >>>
> >> >>> On Thu, Jan 3, 2019 at 11:32 PM Robert Burke <r...@google.com>
> wrote:
> >> >>>>
> >> >>>> That's an interesting idea. I must confess I don't rightly know
> the difference between a schema and coder, but here's what I've got with a
> bit of searching through memory and the mailing list. Please let me know if
> I'm off track.
> >> >>>>
> >> >>>> As near as I can tell, a schema, as far as Beam takes it is a
> mechanism to define what data is extracted from a given row of data. So in
> principle, there's an opportunity to be more efficient with data with many
> columns that aren't being used, and only extract the data that's meaningful
> to the pipeline.
> >> >>>> The trick then is how to apply the schema to a given serialization
> format, which is something I'm missing in my mental model (and then how to
> do it efficiently in Go).
> >> >>>>
> >> >>>> I do know that the Go client package for BigQuery does something
> like that, using field tags. Similarly, the "encoding/json" package in the
> Go Standard Library permits annotating fields and it will read out and
> deserialize the JSON fields and that's it.
> >> >>>>
> >> >>>> A concern I have is that Go (at present) would require pre-compile
> time code generation for schemas to be efficient, and they would still
> mostly boil down to turning []bytes into real structs. Go reflection
> doesn't keep up.
> >> >>>> Go has no mechanism I'm aware of to Just In Time compile more
> efficient processing of values.
> >> >>>> It's also not 100% clear how Schema's would play with protocol
> buffers or similar.
> >> >>>> BigQuery has a mechanism of generating a JSON schema from a proto
> file, but that's only the specification half, not the using half.
> >> >>>>
> >> >>>> As it stands, the code generator I've been building these last
> months could (in principle) statically analyze a user's struct, and then
> generate an efficient dedicated coder for it. It just has no where to put
> them such that the Go SDK would use it.
> >> >>>>
> >> >>>>
> >> >>>> On Thu, Jan 3, 2019 at 1:39 PM Reuven Lax <re...@google.com>
> wrote:
> >> >>>>>
> >> >>>>> I'll make a different suggestion. There's been some chatter that
> schemas are a better tool than coders, and that in Beam 3.0 we should make
> schemas the basic semantics instead of coders. Schemas provide everything a
> coder provides, but also allows for far more readable code. We can't make
> such a change in Beam Java 2.X for compatibility reasons, but maybe in Go
> we're better off starting with schemas instead of coders?
> >> >>>>>
> >> >>>>> Reuven
> >> >>>>>
> >> >>>>> On Thu, Jan 3, 2019 at 8:45 PM Robert Burke <rob...@frantil.com>
> wrote:
> >> >>>>>>
> >> >>>>>> One area that the Go SDK currently lacks: is the ability for
> users to specify their own coders for types.
> >> >>>>>>
> >> >>>>>> I've written a proposal document, and while I'm confident about
> the core, there are certainly some edge cases that require discussion
> before getting on with the implementation.
> >> >>>>>>
> >> >>>>>> At presently, the SDK only permits primitive value types (all
> numeric types but complex, strings, and []bytes) which are coded with beam
> coders, and structs whose exported fields are of those type, which is then
> encoded as JSON. Protocol buffer support is hacked in to avoid the type
> anaiyzer, and presents the current work around this issue.
> >> >>>>>>
> >> >>>>>> The high level proposal is to catch up with Python and Java, and
> have a coder registry. In addition, arrays, and maps should be permitted as
> well.
> >> >>>>>>
> >> >>>>>> If you have alternatives, or other suggestions and opinions, I'd
> love to hear them! Otherwise my intent is to get a PR ready by the end of
> January.
> >> >>>>>>
> >> >>>>>> Thanks!
> >> >>>>>> Robert Burke
> >> >>>>
> >> >>>>
> >> >>>>
> >> >>>> --
> >> >>>> http://go/where-is-rebo
> >> >>
> >> >>
> >> >>
> >> >> --
> >> >> Cheers,
> >> >> Gleb
>
