I've had an offline discussion with Till and we came to the following
conclusion:
All out-lined approaches work perfectly when Flink is built in it's
entirety with maven.
None of the out-lined approach work perfectly when Flink is _not_ built
in it's entirety.
The source-file generation approach has similar problems as the
properties-file one; if I build flink-runtime, switch to another branch,
build flink-dist, then the information that will be displayed in the UI
when I start the cluster may be incorrect.
Put another way, you can still easily run into the issue of "doesn't
work on /my/ machine", the solution to which always is to rebuilt Flink
completely.
Frustratingly enough, with both approaches you can even run into the
issue where the information _seems_ correct, when it actually isn't; but
this isn't something that we can really solve.
As such, this is no longer really about correctness, but one about
convenience and maintenance overhead.
As for maintenance, I think a good argument can be made that the
generation approach would be easier to maintain, in parts because it is
the more standard approach for this kind of thing.
The properties-file approach however clearly wins in convenience as it
neither requires an additional command to be run and doesn't cause
compile errors. The compile error is quite a problem indeed since it
provides no information as to what the actual problem is; or rather how
to solve it.
Overall we concluded that we would like to stick to the existing
approach of having a properties file (or JSON or whatever).
I saw that you updated the PR, and will take a look shortly.
On a personal note, I'm sympathetic to your idea, but people are quite
sensitive to things interrupting their work-flow (which I ran into a few
times already :/ ), so we put a lot of emphasis on that aspect.
On 16/04/2020 16:05, Niels Basjes wrote:
Hi,
Apparently the IDEs (like IntelliJ) are imperfect at supporting the DEVs in
these kinds of use cases.
The solutions you see are like you have in IntelliJ: a "generate-sources"
button.
It is a way of working I see in almost all projects I'm involved with:
Almost all either use a parser generator (like javacc or Antlr) or generate
code from a specification (like Avro or Protobuf).
In all of these cases there is no "default" value you can use so this
discussion never arose in those projects.
All other options (other than 'hard' generating a file/class without having
defaults) involve having some kind of "default" (regardless if it is a
property file, a Json file or a Java interface).
This will lead to the illusion of "it works on my machine" or "why doesn't
it work on my machine".
What I also see is that these options which include "defaults" also have
quite a bit of extra code to handle all the situations related to loading
the settings (either through resource loading or classloading).
The goal I was working on is that the docker image that is loaded is tagged
with a flink+scala version.
If the DEV does not generate the right values or has values from a
different branch then a different docker image may be used then what is
expected leading to "why doesn't it work on my machine" .
So yes it is an additional learning step for the DEVs and yes it takes a
bit of getting used to.
Yet I truly believe it is the best direction.
Niels
On Wed, Apr 15, 2020 at 5:26 PM Till Rohrmann <trohrm...@apache.org> wrote:
I'm not advocating for a specific approach. The point I wanted to make is
that there are solutions which allow us to get rid of the problematic
parsing and not disrupting the workflow. If the Jackson JSON file approach
works for Niels, then I'm fine with that as well.
Cheers,
Till
On Wed, Apr 15, 2020 at 5:21 PM Chesnay Schepler <ches...@apache.org>
wrote:
It doesn't have to be a properties file, nor do we necessarily have to
do any manual parsing.
It could just be a JSON file that we point Jackson at. Naturally we
could also generate it with Jackson.
You'd have a POJO for all the fields with sane defaults (an analogue to
the proposed generated POJO) and 5-6 lines at most for loading the file
and handling errors.
I don't think this would be such a problem.
If I understand Till correctly he's proposing a service-loader approach,
which seems overkill to me. It also introduces some inconsistency in
what one needs to do before working on stuff which I'm apprehensive
about.
On 15/04/2020 16:31, Till Rohrmann wrote:
Hi everyone,
thanks for starting this discussion Niels.
I like the idea of getting rid of parsing a Properties instance. On the
other hand, I also understand that people are concerned about
disrupting
the workflows of our devs.
Maybe we can find a compromise between both approaches. For example,
one
could define an EnvironmentInformationProvider interface which is
loaded
at
runtime. The implementation of this interface could be generated by
`mvn
generate-sources`. Moreover, if we cannot find it, then we fail with a
descriptive exception saying to run `mvn generate-sources`. That way
only
devs who rely on the EnvironmentInformation have to run `mvn
generate-sources` when testing/developing and we could still get rid of
all
the parsing and logic of a properties file. What do you think?
Cheers,
Till
On Mon, Apr 13, 2020 at 10:53 AM Niels Basjes <ni...@basjes.nl> wrote:
Hi,
On Mon, Apr 13, 2020 at 8:30 AM Yang Wang <danrtsey...@gmail.com>
wrote:
Although the latter option is more stable,
i think it is not acceptable for all the developers to execute `mvn
generate-sources` first.
Otherwise, the Flink project is just broken and could not run tests,
Flink
jobs in IDE.
It is important to realize that this discussion is essentially around
the
fact that systems like IntelliJ (what I use) execute various ways of
generating code that have been written in a maven pom.xml in different
ways. Simply put: They don't execute all instructions that have been
defined in the pom.xml automatically when needed.
- With the properties file variant (which uses resource filtering)
IntelliJ does resource filtering automatically yet no maven
plugins
are
run that affect the properties that are replaced.
So the git variables are NOT populated when running in IntelliJ
and
I
have seen several values in the propersites file are nonsense.
As a consequence the Java code reading those needs to catch things
like
"missing properties file" and various kinds of "nonsense values"
and
replace them with a workable default.
So when running this code you are actually running something
different
from what will be run when doing a `mvn clean install`, yet the
developer
is under the illusion it all works because of the code that
catches
all
of
those problems.
Depending on the way these variables are used this may lead to "It
fails
in production but it works fine in IntelliJ" situations.
In my mind the code that catches all of those exceptional
situations in
poorly generated properties only exists so that developers can run
the
(otherwise perfectly fine) software in a "broken"
build/development
environment.
- With the way I propose to generate the code the effect is indeed
slightly harder: If you do not run the pom.xml based code
generation it
will not work.
I understand that this requires the developers to think a bit more
about
the code they are working on.
The upside is that the code either works perfectly or does not
compile.
There is no "it compiles but is really nonsense".
I prefer this.
Also at this point in time this is already true for Flink: There
are
quite a few modules where the developer MUST run mvn
generate-sources
for
all tests to run successfully.
Like I indicated there is a javacc SQL parser and there are a lot
of
tests that rely on generating Avro code before they are able to
run.
I have several projects where systems like Avro and Antlr force me
in
this direction, there is simply no other way to do this.
So i think the version properties file is enough for now. +1 for the
first
option.
Like I said. I'm fine with either choice by the committers.
I would choose the `mvn generate-sources` code variant as it is much
simpler and much more reliable.
Niels
Niels Basjes <ni...@basjes.nl> 于2020年4月9日周四 下午4:47写道：
Hi,
I'm working on https://issues.apache.org/jira/browse/FLINK-16871 to
make
more build time variables (like the scala version) into the code
available
at runtime.
During the review process there was discussion around a basic
question:
*Is
generating java code during the build ok?*
See
-
https://github.com/apache/flink/pull/11245#discussion_r400035133
- https://github.com/apache/flink/pull/11592
-
https://github.com/apache/flink/pull/11592#issuecomment-610282450
As suggested by Chesnay Schepler I'm putting this question to the
mailing
list.
https://github.com/apache/flink/pull/11592#issuecomment-610963947
The main discussion was around the ease of use when running in an
IDE
like
IntelliJ.
So essentially we have two solution directions available:
1. *Generate a properties file and then use the classloader to
load
this
file as a resource and then parse it as a property file.*
This is the currently used solution direction for this part of
the
code.
A rough version of this (to be improved) :
https://github.com/apache/flink/commit/47099f663b7644056e9d87b262cd4dba034f513e
This method has several effects:
1. The developer can run the project immediately from within
the
IDE
as fallback values are provided if the 'actual' values are
missing.
2. This property file (with stuff that should never be
overwritten)
can be modified by placing a different one in the classpath.
In
fact it IS
modified in the flink-dist as it generates a new file with
the
same
name
into the binary distribution (I consider this to be bad).
3. Loading resources means loading, parsing and a lot of
error
handling. Lots of things "can be null" or be a default
value.
So
the
values are unreliable and lots of code needs to handle this.
In
fact
when
running from IntelliJ the properties file is generated poorly
most
of the
time, only during a normal maven build will it work
correctly.
2. *Generate a Java source file and then simply compile this and
make
it
part of the project.*
Essentially the same model as you would have when using Apache
Avro,
Protobuf, Antlr 4 and javacc (several of those are used in
Flink!).
A rough version of this (to be improved) :
https://github.com/apache/flink/commit/d215e4df60dc9d647dcee1aa9a2114cbf49d0566
This method has several effects:
1. The developer MUST run 'mvn generate-sources' before the
actual
the
project immediately from within the IDE as fallback values
are
provided if
the 'actual' values are missing.
2. The code/test will not run until this step is done.
3. Because the file is generated by a plugin it is always
correct.
As
a consequence all variables are always available and the
downstream
users
no longer have to handle the "can be null" or "default value"
situations.
So is generating code similar to what I created a desired change?
My opinion is that it is the better solution, the data available is
more
reliable and as a consequence the rest of the code is simpler.
It would probably mean that during development of flink you should
be
aware
of this and do an 'extra step' to get it running.
What do you guys think?
--
Best regards / Met vriendelijke groeten,
Niels Basjes
--
Best regards / Met vriendelijke groeten,
Niels Basjes
