[
https://issues.apache.org/jira/browse/FLINK-17723?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17123931#comment-17123931
]
Stephan Ewen commented on FLINK-17723:
--------------------------------------
One more thought:
Some NullPointerExceptions and related checks come not from
multi-threaded-ness, but from lazy initialization.
Because Functions in Flink can be shipped to the distributed processing threads
using Java Serialization, some parts of them need to be lazily initialized (via
the open() method). That brings the problem that "closing before lazy
initialization finished" (during early cancellation) may find uninitialized
fields.
This does not impact correctness, but it can pollute the log and make
cancellation more rough (lead to errors that cause the TaskManager to assume an
unclean situation and exit, relying on a restart of the JVM process).
In various parts of the code you will find "serializable factory" patterns that
help work around this and make the code use more "eager initialization" to
simplify this (both behavior and readability).
> Written design for flink threading model and guarantees made to the various
> structual components
> ------------------------------------------------------------------------------------------------
>
> Key: FLINK-17723
> URL: https://issues.apache.org/jira/browse/FLINK-17723
> Project: Flink
> Issue Type: Improvement
> Components: Documentation
> Reporter: John Lonergan
> Priority: Major
>
> I enjoy using Flink but ...
> Do we have a written design for the threading model including the guarantees
> made by the core framework in terms of threading and concurrency.
> Looking at various existing components such as JDBC and file sinks and other
> non-core facilities.
> Having some difficulty understanding the intended design.
> Want to understand the assumptions I can make about when certain functions
> will be called (for example JDBCOutputFormat open vs flush vs writeRecord vs
> close) and whether this will always be from the same thread or some other
> thread, or whether they might be called concurrently, in order to verify the
> correctness of the code.
> What guarantees are there?
> Does a certain reference need a volatile or even a synchronisation or not.
> What's the design for threading?
> If the intended design is not written down then we have to infer it from the
> code and we will definitiely come to different conclusions and thus bugs and
> leaks. and other avoidable horrors.
> It's really hard writing good MT code and a strong design is necessary to
> provide a framework for the code.
> Some info here
> https://flink.apache.org/contributing/code-style-and-quality-common.html#concurrency-and-threading
> , but this isn't a design and doesn't say how it's meant to work. However
> that page does agree that writing MT code is very hard and this just
> underlines the need for a strong and detailed design for this aspect.
> ==
> Another supporting example.
> When I see code like this ...
> FileOutputFormat
>
> {code:java}
> public void close() throws IOException {
> final FSDataOutputStream s = this.stream;
> if (s != null) {
> this.stream = null;
> s.close();
> }
> }
> {code}
> My feeling is that someone else wasn't sure what the right approach was.
> I can only guess that the author was concerned that someone else was going to
> call the function concurrently, or mess with the class state by some other
> means. And, if that were true then would this code even be MT safe - who
> knows? Ought there be a volatile in there or a plain old sync? Or perhaps
> none of the caution is needed at all (framework guarantees preventing the
> need)?
> Or take a look at the extensive sychronisation efforts in
> https://github.com/apache/flink/blob/master/flink-connectors/flink-hadoop-compatibility/src/main/java/org/apache/flink/api/java/hadoop/mapred/HadoopOutputFormatBase.java
> is this code correct? Not to mention that fact that this close() method
> might throw an NPE if there is any possiblity that 'this.outputCommitter'
> might not have been initialised in open OR is the framework can ever call
> close() without open() having completed.
> I find if worrying that I see a lot of code in the project that is similarly
> uncertain and inconsistent syncronisation and resource management.
> I would have hoped that the underlying core framework provided guarantees
> that avoided the need to have extensive synchronisation effort in derived or
> auxiliary classes.
> What's the design.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
