1.5 Checkpoint metadata location

[Bryant Baltes]

Hi All,

After upgrading from 1.3.2 to 1.5.2, one of our apps that uses checkpointing no longer writes metadata files to the state.checkpoints.dir location provided to the flink conf.  I see this email chain addressed this here: https://lists.apache.org/thread.html/922f77880eca2a7b279e153090da2388b54f19e89528a2a35937d9a8@%3Cuser.flink.apache.org%3E. 

I am still a bit unclear what the workaround is.  We use the metadata files to recover when the app goes down or gets restarted.  

Thanks,

Bryant

[vino yang]

Hi Bryant,

Maybe Stefan can answer your question, ping him for you.

Thanks, vino.

Bryant Baltes <[hidden email]> 于2018年9月25日周二 上午12:29写道：

Hi All,

After upgrading from 1.3.2 to 1.5.2, one of our apps that uses checkpointing no longer writes metadata files to the state.checkpoints.dir location provided to the flink conf.  I see this email chain addressed this here: https://lists.apache.org/thread.html/922f77880eca2a7b279e153090da2388b54f19e89528a2a35937d9a8@%3Cuser.flink.apache.org%3E. 

I am still a bit unclear what the workaround is.  We use the metadata files to recover when the app goes down or gets restarted.  

Thanks,

Bryant

[Till Rohrmann]

Hi Bryant,

I think if you explicitly define the StateBackend in your code (calling StreamExecutionEnvironment#setStateBackend), then you also define the checkpointing directory when calling the StateBackend's constructor. This is also the directory in which the metadata files are stored. You could also say that the job specific settings overwrite the cluster settings specified in the flink-conf.yaml.

I think the workaround was to change some Flink code and build a custom version. For more details, you have to ask Gyula.

Cheers,

Till

On Tue, Sep 25, 2018 at 4:59 AM vino yang <[hidden email]> wrote:

Hi Bryant,

Maybe Stefan can answer your question, ping him for you.

Thanks, vino.

Bryant Baltes <[hidden email]> 于2018年9月25日周二 上午12:29写道：

Hi All,

After upgrading from 1.3.2 to 1.5.2, one of our apps that uses checkpointing no longer writes metadata files to the state.checkpoints.dir location provided to the flink conf.  I see this email chain addressed this here: https://lists.apache.org/thread.html/922f77880eca2a7b279e153090da2388b54f19e89528a2a35937d9a8@%3Cuser.flink.apache.org%3E. 

I am still a bit unclear what the workaround is.  We use the metadata files to recover when the app goes down or gets restarted.  

Thanks,

Bryant

[Gyula Fóra]

Yes, the only workaround I found at the end was to restore the previous behavior where metadata files are written separately.

But for this you need a custom Flink build with the changes to the check pointing logic.

Gyula

On Tue, 25 Sep 2018 at 16:45, Till Rohrmann <[hidden email]> wrote:

Hi Bryant,

I think if you explicitly define the StateBackend in your code (calling StreamExecutionEnvironment#setStateBackend), then you also define the checkpointing directory when calling the StateBackend's constructor. This is also the directory in which the metadata files are stored. You could also say that the job specific settings overwrite the cluster settings specified in the flink-conf.yaml.

I think the workaround was to change some Flink code and build a custom version. For more details, you have to ask Gyula.

Cheers,

Till

On Tue, Sep 25, 2018 at 4:59 AM vino yang <[hidden email]> wrote:

Hi Bryant,

Maybe Stefan can answer your question, ping him for you.

Thanks, vino.

Bryant Baltes <[hidden email]> 于2018年9月25日周二 上午12:29写道：

Hi All,

After upgrading from 1.3.2 to 1.5.2, one of our apps that uses checkpointing no longer writes metadata files to the state.checkpoints.dir location provided to the flink conf.  I see this email chain addressed this here: https://lists.apache.org/thread.html/922f77880eca2a7b279e153090da2388b54f19e89528a2a35937d9a8@%3Cuser.flink.apache.org%3E. 

I am still a bit unclear what the workaround is.  We use the metadata files to recover when the app goes down or gets restarted.  

Thanks,

Bryant