[
https://issues.apache.org/jira/browse/PIO-30?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15495931#comment-15495931
]
ASF GitHub Bot commented on PIO-30:
-----------------------------------
GitHub user Ziemin opened a pull request:
https://github.com/apache/incubator-predictionio/pull/295
[PIO-30] Set up a cross build for Scala 2.10 (Spark 1.6.2) and Scala …
This PR introduces a simple profile-based build of PredictionIO, which
comes along with some upgrades including a new version of Spark.
### The key changes include:
* build.sbt - here I created two profiles with different sets of artifacts
to be included. Their names are _scala-2.11_ and _scala-2.10_, where the
former is chosen by default. In order to set a desired profile for sbt command
`-Dbuild.profile=<profile_name>` property has to be provided.
You can print the description of some profile by e.g.
```sbt -Dbuild.profile=scala-2.10 printProfile```
The _scala-2.11_ settings include Spark version 2.0.0, while _scala-2.11_
sets it to 1.6.2. This can be configured, by adding a dedicated property:
``` sbt -Dbuild.profile=scala-2.11 -Dspark.version=1.6.0 -D
Dhadoop.version=2.6.4 <sbt task> ```
This command will set a build profile to _scala-2.11_, but will use a
different versions of Spark and Hadoop. It makes the configuration more
flexible, especially if someone wants to build the project according to their
own needs.
Very important thing to note is that versions of spark before 1.6.x are no
longer supported and scala 2.10.x is deprecated.
*
`data/src/main/spark-1/org/apache/predictionio/data/SparkVersionDependent.scala`
and
`data/src/main/spark-2/org/apache/predictionio/data/SparkVersionDependent.scala`
- are the only examples of version dependent code. They are solely for
providing a proper type of an object for Spark sql related actions. Sbt is
configured to include version specific source paths like these.
* make_distribution.sh - in order to create an archive for Scala 2.10 one
has to provide it with an argument in the same way as sbt
(./make_distribution.sh -Dbuild.profile=scala-2.10). By default it will build
for _scala-2.11_.
* integration tests - The docker image is updated, I pushed it with a tag
spark_2.0.0 not to interfere with the current build. It contains both versions
of Spark and on startup sets up environment according to dependencies that
predictionIO was built with. It uses a simple Java program
`tests/docker-files/BuildInfoPrinter.java` linked with the assembly of
PredictionIO to acquire necessary information. Travis CI makes use of the setup
and runs 8 parallel builds, the number doubled because of introducing two
different build profiles.
I also noticed that there are many places hardcoding different version
numbers and links to packages to be downloaded. Maintaining the project and
keeping everything consistent gets only more difficult, therefore I came up
with a few small scripts getting proper versions of dependencies from the build
config and setting some variables accordingly. An example is `conf/vendors.sh`,
which provided that some variables are set (e.g. by `dev/set-build-profile.sh`)
initializes some other useful variables, i.e. `SPARK_DOWNLOAD, SPARK_ARCHIVE,
SPARK_DIRNAME`. They are used in travis configuration as well as in the
Dockerfile, which now should be built with a dedicated script
`tests/docker-build.sh`. Such approach makes it easier to keep everything
coherent while bumping version numbers.
### Some problems encountered during upgrade
Updating Spark caused some troubles
* The classpath has to be extended to run the unit tests successfully for
some sub-packages. (see `build.sbt`)
* Column names have to be handled differently for Postgres in JDBCPevents,
as Spark surrounds them with "..." what breaks the current schema in this case
* `tests/pio_tests/utils.py` - has a special Spark pass through argument to
set `spark.sql.warehouse.dir`, because the defaults cause runtime exceptions.
See ->
[here](https://mail-archives.apache.org/mod_mbox/spark-user/201608.mbox/%3ccamassd+efz+uscmnzvkfp00qbr9ynv8lrfhvz9lrmnwh2vk...@mail.gmail.com%3E)
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/Ziemin/incubator-predictionio upgrade
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/incubator-predictionio/pull/295.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #295
----
commit 39879f04639b123510f0668e0de7b33fd7418784
Author: Marcin Ziemiński <ziem...@gmail.com>
Date: 2016-08-09T21:20:54Z
[PIO-30] Set up a cross build for Scala 2.10 (Spark 1.6.2) and Scala 2.11
(Spark 2.0.0).
Changes also include updating travis integration tests, which run now
eight parallel builds.
----
> Cross build for different versions of scala and spark
> -----------------------------------------------------
>
> Key: PIO-30
> URL: https://issues.apache.org/jira/browse/PIO-30
> Project: PredictionIO
> Issue Type: Improvement
> Reporter: Marcin Ziemiński
>
> The present version of Scala is 2.10 and Spark is 1.4, which is quite old.
> With Spark 2.0.0 come many performance improvements and features, that people
> will definitely like to add to their templates. I am also aware that past
> cannot be ignored and simply dumping 1.x might not be an option for other
> users.
> I propose setting up a crossbuild in sbt to build with scala 2.10 and Spark
> 1.6 and a separate one for Scala 2.11 and Spark 2.0. Most of the files will
> be consistent between versions including API. The problematic ones will be
> divided between additional source directories: src/main/scala-2.10/ and
> src/main/scala-2.11/. The dockerized tests should also take the two versions
> into consideration
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
