Thanks. That helps answer my question. What if a batch fails? Could records in 
that batch get written to disk by one HDFS state but not another?

From: Bobby Evans <[email protected]>
Reply-To: "[email protected]" <[email protected]>
Date: Thursday, August 3, 2017 at 1:27 PM
To: "[email protected]" <[email protected]>
Subject: Re: HDFS state and commits and only-once semantics

Writing to a state in storm in not atomic.  Storm guarantees that once the 
batch completes that the data is written out to all of the states that expect 
to receive it.  The HDFS state guarantees that the dat will have been flushed 
out the the data nodes when a batch completes and if the topology keeps running 
eventually the files will be rotated and made available for others to process, 
but there is no guarantee that the files will rotate at the same time or 
anything like that.


- Bobby



On Thursday, August 3, 2017, 1:17:39 PM CDT, Geoffrey Holmes 
<[email protected]> wrote:



I read STORM-837 (https://issues.apache.org/jira/browse/STORM-837) and have a 
question. How does this work if I have more than one HDFS state in my Trident 
topology? Can I ensure that a record ends up written to both HDFS states or 
none but not just one or the other?




Reply via email to