I think adding hints for runners is reasonable, though hints should
always be assumed to be optional - they shouldn't change semantics of the
program (otherwise you destroy the portability promise of Beam). However
there are many types of hints that some runners might find useful (e.g. this
step needs more memory. this step runs ML algorithms, and should run on a
machine with GPUs. etc.)
Robert has mentioned in the past that we should try and keep PCollection
an immutable object, and not introduce new setters on it. We slightly break
this already today with PCollection.setCoder, and that has caused some
problems. Hints can be set on PTransforms though, and propagate to that
PTransform's output PCollections. This is nearly as easy to use however, as
we can implement a helper PTransform that can be used to set hints. I.e.
pc.apply(AddHints.of(hint1, hint2, hint3))
Is no harder than called pc.addHint()
Reuven
On Tue, Jan 30, 2018 at 11:39 AM, Jean-Baptiste Onofré <[email protected]
<mailto:[email protected]>> wrote:
Maybe I should have started the discussion on the user mailing list:
it would be great to have user feedback on this, even if I got your
points.
Sometime, I have the feeling that whatever we are proposing and
discussing, it doesn't go anywhere. At some point, to attract more
people, we have to get ideas from different perspective/standpoint.
Thanks for the feedback anyway.
Regards
JB
On 30/01/2018 20:27, Romain Manni-Bucau wrote:
2018-01-30 19:52 GMT+01:00 Kenneth Knowles <[email protected]
<mailto:[email protected]> <mailto:[email protected]
<mailto:[email protected]>>>:
I generally like having certain "escape hatches" that are
well
designed and limited in scope, and anything that turns out
to be
important becomes first-class. But this one I don't really
like
because the use cases belong elsewhere. Of course, they
creep so you
should assume they will be unbounded in how much gets
stuffed into
them. And the definition of a "hint" is that deleting it
does not
change semantics, just performance/monitor/UI etc but this
does not
seem to be true.
"spark.persist" for idempotent replay in a sink:
- this is already @RequiresStableInput
- it is not a hint because if you don't persist your
results are
incorrect
- it is a property of a DoFn / transform not a
PCollection
Let's put this last point aside since we'll manage to make it
working wherever we store it ;).
schema:
- should be first-class
Except it doesn't make sense everywhere. It is exactly like
saying "implement this" and 2 lines later "it doesn't do
anything for you". If you think wider on schema you will want to
do far more - like getting them from the previous step etc... -
which makes it not an API thing. However, with some runner like
spark, being able to specifiy it will enable to optimize the
execution. There is a clear mismatch between a consistent and
user friendly generic and portable API, and a runtime, runner
specific, implementation.
This is all fine as an issue for a portable API and why all EE
API have a map to pass properties somewhere so I don't see why
beam wouldn't fall in that exact same bucket since it embraces
the drawback of the portability and we already hit it since
several releases.
step parallelism (you didn't mention but most runners need
some
control):
- this is a property of the data and the pipeline
together, not
just the pipeline
Good one but this can be configured from the pipeline or even a
transform. This doesn't mean the data is not important - and you
are more than right on that point, just that it is configurable
without referencing the data (using ranges is a trivial example
even if not the most efficient).
So I just don't actually see a use case for free-form hints
that we
haven't already covered.
There are several cases, even in the direct runner to be able to
industrialize it:
- use that particular executor instance
- debug these infos for that transform
etc...
As a high level design I think it is good to bring hints to beam
to avoid to add ad-hoc solution each time and take the risk to
loose the portability of the main API.
Kenn
On Tue, Jan 30, 2018 at 9:55 AM, Romain Manni-Bucau
<[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>>
wrote:
Lukasz, the point is that you have to choice to either
bring all
specificities to the main API which makes most of the
API not
usable or implemented or the opposite, not support
anything.
Introducing hints will allow to have eagerly for some
runners
some features - or just some very specific things - and
once
mainstream it can find a place in the main API. This is
saner
than the opposite since some specificities can never
find a good
place.
The little thing we need to take care with that is to
avoid to
introduce some feature flipping as support some feature
not
doable with another runner. It should really be about
runing a
runner execution (like the schema in spark).
Romain Manni-Bucau
@rmannibucau <https://twitter.com/rmannibucau
<https://twitter.com/rmannibucau>> | Blog
<https://rmannibucau.metawerx.net/
<https://rmannibucau.metawerx.net/>> | Old Blog
<http://rmannibucau.wordpress.com
<http://rmannibucau.wordpress.com>> | Github
<https://github.com/rmannibucau
<https://github.com/rmannibucau>> | LinkedIn
<https://www.linkedin.com/in/rmannibucau
<https://www.linkedin.com/in/rmannibucau>>
2018-01-30 18:42 GMT+01:00 Jean-Baptiste Onofré
<[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>>:
Good point Luke: in that case, the hint will be
ignored by
the runner if the hint is not for him. The hint can
be
generic (not specific to a runner). It could be
interesting
for the schema support or IOs, not specific to a
runner.
What do you mean by gathering
PTransforms/PCollections and
where ?
Thanks !
Regards
JB
On 30/01/2018 18:35, Lukasz Cwik wrote:
If the hint is required to run the persons
pipeline
well, how do you expect that the person we be
able to
migrate their pipeline to another runner?
A lot of hints like "spark.persist" are really
the user
trying to tell us something about the
PCollection, like
it is very small. I would prefer if we gathered
this
information about PTransforms and PCollections
instead
of runner specific knobs since then each runner
can
choose how best to map such a property on their
internal
representation.
On Tue, Jan 30, 2018 at 2:21 AM, Jean-Baptiste
Onofré
<[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>
<mailto:[email protected]
<mailto:[email protected]> <mailto:[email protected]
<mailto:[email protected]>>>> wrote:
Hi,
As part of the discussion about schema,
Romain
mentioned hint. I
think it's
worth to have an explanation about that
and
especially it could be
wider than
schema.
Today, to give information to the runner,
we use
PipelineOptions.
The runner can
use these options, and apply for all inner
representation of the
PCollection in
the runner.
For instance, for the Spark runner, the
persistence
storage level
(memory, disk,
...) can be defined via pipeline options.
Then, the Spark runner automatically
defines if
RDDs have to be
persisted (using
the storage level defined in the pipeline
options),
for instance if
the same
POutput/PCollection is read several time.
However, the user doesn't have any way to
provide
indication to the
runner to
deal with a specific PCollection.
Imagine, the user has a pipeline like
this:
pipeline.apply().apply().apply(). We
have three PCollections involved in this
pipeline.
It's not
currently possible
to give indications how the runner should
"optimized" and deal with
the second
PCollection only.
The idea is to add a method on the
PCollection:
PCollection.addHint(String key, Object
value);
For instance:
collection.addHint("spark.persist",
StorageLevel.MEMORY_ONLY);
I see three direct usage of this:
1. Related to schema: the schema
definition could
be a hint
2. Related to the IO: add headers for the
IO and
the runner how to
specifically
process a collection. In Apache Camel, we
have
headers on the
message and
properties on the exchange similar to
this. It
allows to give some
indication
how to process some messages on the Camel
component. We can imagine
the same of
the IO (using the PCollection hints to
react
accordingly).
3. Related to runner optimization: I see
for
instance a way to use
RDD or
dataframe in Spark runner, or even
specific
optimization like
persist. I had lot
of questions from Spark users saying: "in
my Spark
job, I know where
and how I
should use persist (rdd.persist()), but I
can't do
such optimization
using
Beam". So it could be a good improvements.
Thoughts ?
Regards
JB
--
Jean-Baptiste Onofré
[email protected] <mailto:[email protected]>
<mailto:[email protected] <mailto:[email protected]>>
<mailto:[email protected]
<mailto:[email protected]> <mailto:[email protected]
<mailto:[email protected]>>>
http://blog.nanthrax.net
Talend - http://www.talend.com
