danny0405 edited a comment on pull request #2899:
URL: https://github.com/apache/hudi/pull/2899#issuecomment-837659380


   > > each write task write out these records buffer (grouping by file group 
id) using a write client there.
   > 
   > Understood. So each task manager is, sending RPC calls to the timeline 
server running locally? This is very different from Spark. What's the life 
cycle of the `write client` in each JVM process? i.e how often do we 
re-initialize it? (it can't be per record, so wondering)
   > 
   > I get why you want a singleton now. Let's work together to figure 
something out.
   
   The life cycle of the flink `write client` is: it is living until the job 
ends, we start a write client for each task and we use that client always. In 
my thoughts, if i request the embedded timeline server with a new instant time, 
the server should sync itself first, it is more easy for server to keep the 
freshness than client(i.e each client should figure out how and when to 
re-initialize the fs view server).


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to