should
have at least one runner that can execute the complete model (may be a
direct runner)"
I want to highlight this, because whether an _SDK_ supports unbounded
data
is not particularly well-defined, and will evolve:
- With the Runner API, an SDK will need to support building a graph
with
unbounded constructs, as today with probably minimal changes.
- With the Fn API, if any part of the Fn API is specific to unbounded
data, the SDK will need to implement it. I think right now there is
no such
thing, and we don't want such a thing, so SDKs implementing the Fn API
automatically support unbounded data.
- There will also likely be an SDK-specific shim just as there is
today,
to leverage idiomatic deserialized representations. The richness of
this
shim will decrease so that it will need to "support" unbounded data
but
that will be a ~one liner.
Getting the Python SDK on master will accelerate our progress towards
the
Fn API - partly technical, partly community - which is the best path
towards support for unbounded data across multiple runners. I think
the
criteria are written with the completed portability framework in
mind. So
this exchange makes me actually more convinced we should merge
python-sdk
to master.
On Fri, Jan 20, 2017 at 9:53 AM, Robert Bradshaw <
rober...@google.com.invalid> wrote:
On Thu, Jan 19, 2017 at 11:56 PM, Dan Halperin
<dhalp...@google.com.invalid> wrote:
I do not think that Python SDK yet meets the bar [1] for implementing
the
Beam model -- supporting Unbounded data is very important. That said,
given
the committed and sustained set of contributors, it generally makes
sense
to me to make an exception in anticipation of these features being
fleshed
out soon; including potentially new users/contributors that would
arrive
once in master.
[1] https://lists.apache.org/thread.html/CAAzyFAxcmexUQnbF=Y
k0plmm3f5e5bqwjz4+c5doruclnxo...@mail.gmail.com
That is a valid point. The Python SDK supports all the unbounded
parts
of the model except for unbounded sources, which was deferred while
seeing how https://s.apache.org/splittable-do-fn played out. I've
been
working with the team and merging/reviewing most of their code, and
have full confidence this will be coming (and on that note can vouch
for a healthy community and support which are much harder to add
later).
In short, I think it has the required maturity, and I'm in favor of
merging soonish.
On Wed, Jan 18, 2017 at 12:24 AM, Ahmet Altay
<al...@google.com.invalid
wrote:
Thank you all for the comments so far. I would follow the process as
suggested by Davor and others in this thread.
Ahmet
On Tue, Jan 17, 2017 at 11:47 PM, Sergio Fernández <
wik...@apache.org
wrote:
Hi
On Tue, Jan 17, 2017 at 5:22 PM, Ahmet Altay
<al...@google.com.invalid
wrote:
tl;dr: I would like to start a discussion about merging
python-sdk
branch
to master branch. Python SDK is mature enough and merging it to
master
will
accelerate its development and adoption.
Good point, Ahmet!
I've following closed the development since it was imported in
June.
For
the prototypes I've implemented so far it works quite well; I guess
we'd
just need to focus the next months in bringing more runners support.
With a great effort from a lot of contributors(*), Python SDK [1]
is
now
a
mostly complete, tested, performant Python implementation of the
Beam
model. Since June, when we first started with Python SDK in Apache
Beam
we
have been continuously improving it.
I wouldn't merge during the preparation of 0.5.0 release, but
after
that
could be a good time to merge back into master.
** Python SDK currently supports:
* Model: All main concepts are present (ParDo, GroupByKey,
Windowing
etc.).
* IO: There are extensible APIs for writing new bounded sources
and
sinks.
Implementations are provided for Text, Avro, BigQuery, and
Datastore.
* Runners: Python SDK has an extensible base runner module that
allows
building specific runners on top of it. The SDK comes with two
pipeline
runners: DirectRunner and DataflowRunner; and it is possible to
add
more.
The existing runners are currently limited to bounded execution
and
otherwise equivalent to their Java SDK counterparts in
functionality.
What would the effort of porting, and maintaining, parallel
versions
of
the
Java runners? I guess I'd need to dig deeper in the model, but this
may
represent a major effort for the project, right?
It is somewhat higher for DirectRunner because DirectRunner also
implements
the code for execution. It is not that high for DataflowRunner
because
the
base runner module has a lot of helpers with the right hooks for
implementing a generic runner. I would _expect_ the experience in
general
would be similar to the latter.
* Testing: Python SDK implements ValidatesRunner test framework
for
implementing integration test for current and future runners.
There
is
unit
test coverage for all modules, and a number of integrations test
for
validating existing runners.
* Documentation and examples: Documentation work has started on
Python
SDK.
Beam Programming Guide page has been updated to include Python
[2].
The
code comes with many ready to use examples and we are in a good
place
to
start documenting those on the website.
** We are not done yet, next on the roadmap we have:
* Streaming: Both of the existing runners lack support for
streaming
execution, and currently there is work going on for adding
streaming
support to DirectRunner [3].
* Documentation: Filling the rest of the Beam documentations with
Python
SDK specific information and examples.
* SDK consistency: Making Python SDK consistent with the Java
SDK.
We
have
come a long way on this and have only a few items left [4].
* Beamifying: We have been working on removing Dataflow-specific
references
both from the documentation and from the code. There is some work
left,
and
we are currently working on those as well [5].
** Steps and implications of merging to master:
* Master branch is merged to python-sdk branch at regular
intervals
and
the
last merge was on 12/22. All the past merges were uneventful
because
there
is a minimal overlap in modified files between branches.
Integrating
python-sdk to master will similarly touch a small number of
existing
files.
* Python SDK is using the same tools for building and testing. It
is
already integrated with Maven, Jenkins and Travis. Specifically
the
impact
to the testing infrastructure would be:
- There will be two additional test configurations in Travis.
Since
Travis
runs all configurations in parallel there should not be a
noticeable
change
in the Travis run time.
- Jenkins pre-commit test will start running the Python SDK
tests.
It
will
add an additional 5 minutes to the completion time of pre-commit
test.
Historically Python SDK tests were not flaky and did not cause any
random
failures.
- Jenkins Python post-commit test is already separated from the
other
post-commit tests and will continue to exist. It would not change
the
testing time for any other test.
* The release process needs to be updated to accommodate
releasing
Python
artifacts. Python SDK would fit in the existing release schedule
and
could
be released along with the Java SDK. The additional steps would
include:
- Generating Python artifacts. This could be done with a single
command
using Maven today.
- Publishing the artifacts to a central repository such as PyPI.
I'm more than happy to help on this. We left on purpose some
things
open
when we added Maven support to the Python build.
That would be awesome. We can coordinate on that post-merge.
- Updating the release guide to reflect the changes above.
* Users: There are existing users using the Python SDK. To give a
rough
estimate, a distribution of the Beam Python SDK had a total of 23K
downloads in the past 6 months [6]. Some of those users are
already
engaged
with the community (e.g. [7]). There might be an increased amount
engagement from the rest of them after the merge.
Python 3 support is something we definitively need to look ahead.
I'd
try
to make the codebase compatible with both 2.7.x and 3.6.x, rather
than
using other solutions like 2to3.
I agree with you. I think it makes more sense to make codebase
compatible
with both. As you mentioned Python 3 support is not a short-term goal
in
the roadmap, and we can discuss it more as we approach that.
Looking forward to hearing your thoughts and comments on
“graduating”
python-sdk to the master.
Thank you,
Ahmet
(*) Python SDK branch currently has a diverse group of
contributors.
Regular contributors include Charles Chen, Chamikara Jayalath,
María
García
Herrero, Mark Liu, Pablo Estrada, Robert Bradshaw (Apache Beam
PMC),
Sourabh Bajaj, and Vikas Kedigehalli. We have also had
contributions
from
Abdullah Bashir, Marco Buccini, Sergio Fernández, Seunghyun Lee,
and
Younghee Kwon.
[1] https://github.com/apache/beam/tree/python-sdk/sdks/python
[2] https://beam.apache.org/documentation/programming-guide/
[3] https://issues.apache.org/jira/browse/BEAM-1265
[4]
https://issues.apache.org/jira/issues/?jql=status%20%3D%20Op
en%20AND%20labels%20%3D%20sdk-consistency
[5] https://issues.apache.org/jira/browse/BEAM-1218
[6] https://pypi.python.org/pypi/google-cloud-dataflow/json
[7] https://issues.apache.org/jira/browse/BEAM-1251
Great summary, Ahmet. Thanks.
Cheers,
--
Sergio Fernández
Partner Technology Manager
Redlink GmbH
m: +43 6602747925
e: sergio.fernan...@redlink.co
w: http://redlink.co
