Hi, It will only affect the states. Because Flink will serialize these states when doing snapshot, therefore, we need to make sure that the states can be deserialized if the program is restarted. As for the stateless operators, it’s not affected because Flink won’t do a snapshot on them.
Best, Jiayi Liao Original Message Sender:Jayant ametawittyam...@gmail.com Recipient:bupt_ljybupt_...@163.com Cc:useru...@flink.apache.org Date:Wednesday, Nov 28, 2018 17:50 Subject:Re: Flink operator UUID and serialVersionUID Thanks, I'll look into the bravo project. Will it just impact the MapState or all the operators? If I have a map operator which converts DataStreamMyObject to DataStreamTuple2String, MyObject. Will this fail to recover as well if a field is added to MyObject? Jayant Ameta On Wed, Nov 28, 2018 at 3:08 PM bupt_ljy bupt_...@163.com wrote: Hi, It’ll fail because Flink can’t successfully deserialize the data in savepoint into the new “MyObject” class. There is no official way to fix this problem. However, you can take a look at the bravo projecthttps://github.com/king/bravo, which can help to reconstruct the savepoint, but only with the RocksDBStateBackend now. Best, Jiayi Liao Original Message Sender:Jayant ametawittyam...@gmail.com Recipient:bupt_ljybupt_...@163.com Cc:useru...@flink.apache.org Date:Wednesday, Nov 28, 2018 17:14 Subject:Re: Flink operator UUID and serialVersionUID If I upgrade my flink job, and add a field in "MyObject" class. Will the restore fail? If so, how to handle such scenarios? Should I convert the "MyObject" instance in json and store the string? Jayant Ameta On Wed, Nov 28, 2018 at 1:26 PM bupt_ljy bupt_...@163.com wrote: Hi Jayant, If you change the “MyObject” class and influence the serialized value, then the “MyObject” instance can’t be deserialized when restoring, which causes a failure of restoring. You can just use the default serialVersionUID instead defining it explicitly(it makes no difference if you do it). Best, Jiayi Liao Original Message Sender:Jayant ametawittyam...@gmail.com Recipient:bupt_ljybupt_...@163.com Cc:useru...@flink.apache.org Date:Wednesday, Nov 28, 2018 15:46 Subject:Re: Flink operator UUID and serialVersionUID Thanks for clarifying Jiayi. If there is a change in "MyObject" class, would it help to have a serialVersionUID defined? Thanks, Jayant On Wed, Nov 28, 2018 at 12:52 PM bupt_ljy bupt_...@163.com wrote: Hi, Jayant 1. The uuid is an unique identifier for a specific operator, which means that Flink uses the uuid to recognize the operator when restoring. 2. The operator has already implemented the Serializable interface so you don’t need to do it explicitly. 3. The type information of “MyObject” should be defined in the MapStateDescriptor, so no need to worry about it. Best, Jiayi Liao Original Message Sender:Jayant ametawittyam...@gmail.com Recipient:useru...@flink.apache.org Date:Wednesday, Nov 28, 2018 15:09 Subject:Flink operator UUID and serialVersionUID Hi all, I've a few questions regarding serial version: 1. The production ready checklist mentions using uuidsfor operators. How is it different from setting a serialVersionUID on an operator? 2. Which operators need to have a serialVersionUID present (or implement Serializable interface)? 3. If I have a MapStateString, MyObject, does MyObject need to have a serialVersionUID and does it need to implement Serializable interface? Thanks, Jayant
