Thank you Stephan for spotting the problem. In hindsight it’s obvious that this
can never work. I’ll figure something out :)
> Am 24.03.2017 um 10:28 schrieb Stephan Ewen :
>
> The code will not work properly, sorry.
>
> The value returned by the state is whatever is stored under the key for wh
The code will not work properly, sorry.
The value returned by the state is whatever is stored under the key for
which the function was called the last time.
In addition, the unsynchronized access is most likely causing the RocksDB
fault.
TL:DR - The "ValueState" / "ListState" / etc in Flink are n
Hi,
@Robert: I have uploaded all the log files that I could get my hands on to
https://www.dropbox.com/sh/l35q6979hy7mue7/AAAe1gABW59eQt6jGxA3pAYaa?dl=0. I
tried to remove all unrelated messages logged by the job itself. In
flink-root-jobmanager-0-micardo-dev.log I kept the Flink startup messag
Florian, can you post the log of the Taskmanager where the segfault
happened ?
On Wed, Mar 22, 2017 at 6:19 PM, Stefan Richter wrote:
> Hi,
>
> for the first checkpoint, from the stacktrace I assume that the backend is
> not accessed as part of processing an element, but by another thread. Is
>
Hi,
for the first checkpoint, from the stacktrace I assume that the backend is not
accessed as part of processing an element, but by another thread. Is that
correct? RocksDB requires accessing threads to hold the task’s checkpointing
lock, otherwise they might call methods on an instance that i
Hi Stephen,
you are right, the second stack trace is indeed from a run of Flink 1.1.4.
Sorry, my bad.
That leaves us with the first trace of a segfault for which I can guarantee
that it brought down a 1.2.0 instance. Unfortunately I cannot reproduce the
problem. It has happened twice so far, b
Hi!
It looks like you are running the RocksDB state backend 1.1 (is still an
old version packaged into your JAR file?)
This line indicates that: org.apache.flink.contrib.streaming.state.
RocksDBStateBackend.performSemiAsyncSnapshot (the method does not exist in
1.2 any more)
Can you try and run
Hi,
I have experienced two crashes of Flink 1.2 by SIGSEGV, caused by something in
RocksDB. What is the preferred way to report them? All I got at the moment are
two hs_err_pid12345.log files. They are over 4000 lines long each. Is there
anything significant that I should extract to help you gu