We are still pretty new to Flink and I have a conceptual / DevOps question.
When a job is modified and we want to deploy the new version, what is the preferred method? Our jobs have a lot of keyed state. If we use snapshots we have old state that may no longer apply to the new pipeline. If we start a new job we can reprocess historical data from Kafka, but that can be very resource heavy for a while. Is there an option I am missing? Are there facilities to “patch” or “purge” selectively the keyed state? Michael |
Hi,
General solution for state/schema migration is under development and it might be released with Flink 1.6.0. Before that, you need to manually handle the state migration in your operator’s open method. Lets assume that your OperatorV1 has a state field “stateV1”. Your OperatorV2 defines field “stateV2”, which is incompatible with previous version. What you can do, is to add a logic in open method, to check: 1. If “stateV2” is non empty, do nothing 2. If there is no “stateV2”, iterate over all of the keys and manually migrate “stateV1” to “stateV2” In your OperatorV3 you could drop the support for “stateV1”. I have once implemented something like that here: Hope that helps! Piotrek
|
Hi Piotrek, It seems that this was implemented by `Operator` API, which is a more low level api compared to `Function` API. Since in `Function` API level we can only migrate state by event triggered, it is more convenient in this way to migrate state by foreach all keys in `open()` method. If I was implemented state operator by `ProcessFunction` API, is it possible to port it to `KeyedProcessOperator` and do the state migration that you mentioned? And are there something concerned and difficulties that will leads to restored state failed or other problems? Thank you! Best Regards, Tony Wei 2018-06-07 16:10 GMT+08:00 Piotr Nowojski <[hidden email]>:
|
What function are you implementing and how are you using it?
Usually it’s enough if your function implements RichFunction (or rather extend from AbstractRichFunction) and then you could use RichFunction#open in the similar manner as in the code that I posted in previous message. Flink in many places performs instanceof chekcs like: org.apache.flink.api.common.functions.util.FunctionUtils#openFunction public static void openFunction(Function function, Configuration parameters) throws Exception{
if (function instanceof RichFunction) { RichFunction richFunction = (RichFunction) function; richFunction.open(parameters); } } Piotrek
|
Hi Piotrek, I used `ProcessFunction` to implement it, but it seems that I can't call `getKeyedStateBackend()` like `WindowOperator` did. I found that `getKeyedStateBackend()` is the method in `AbstractStreamOperator` and `ProcessFunction` API didn't extend it. Dose that mean I can't look up all keys and migrate the entire previous states to the new states in `ProcessFunction#open()`? As I said, do I need to port `ProcessFunction` to `KeyedProcessOperator` to migration state like the manner showed in `WindowOperator`? Best Regards, Tony Wei 2018-06-07 20:28 GMT+08:00 Piotr Nowojski <[hidden email]>:
|
Hi,
Oh, I see now. Yes indeed getKeyedStateBackened() is not exposed to the function and you can not migrate your state that way. As far as I know yes, at the moment in order to convert everything at once (without getKeyes you still can implement lazy conversion) you would have to write your own operator. Piotrek
|
Hi Piotrek, So my question is: is that feasible to migrate state from `ProcessFunction` to my own operator then use `getKeyedStateBackend()` to migrate the states? If yes, is there anything I need to be careful with? If no, why and can it be available in the future? Thank you. Best Regards, Tony Wei 2018-06-07 21:43 GMT+08:00 Piotr Nowojski <[hidden email]>:
|
Hi,
Yes it should be feasible. As I said before, with Flink 1.6 there will be better way for migrating a state, but for now you either need to lazily convert the state, or iterate over the keys and do the job manually. Piotrek
|
In reply to this post by Michael Latta
Hi all,
I think I see a way to eagerly do full state migration without writing your own Operator, but it's kind of hacky and may have flaws I'm not aware of. In Flink 1.5 we now have the possibility to connect BroadcastStreams to KeyedStreams and apply a KeyedBroadcastProcessFunction. This is relevant because in the processBroadcastElement() method you can supply a KeyedStateFunction to the Context.applyToKeyedState() method, and this KeyedStateFunction will be applied every item of keyed state associated with the state descriptor you specify. I've been doing some experiments with this, and it's quite powerful in cases where it's useful to operate on all of your application's state. I believe this was intended for cases where an update to an item of broadcast state has implications for associated keyed state, but I see nothing that prevents you from essentially ignoring the broadcast stream and using this mechanism to implement keyed state migration. David On Fri, Jun 8, 2018 at 9:27 AM, Piotr Nowojski <[hidden email]> wrote:
David Anderson | Training Coordinator
-- Join Flink Forward - The Apache Flink Conference Stream Processing | Event Driven | Real Time |
Thank you all. This discussion is very helpful. It sounds like I can wait for 1.6 though given our development status.
Michael
|
Hi David, I have read the document for `Context.applyToKeyedState()`, but I still have some questions for using it to implement keyed state migration. `Context.applyToKeyedState()` can only be called in `processBoradcaseElement()`, so it won't have any key information. It looks like I can use `KeyedStateFunction` to get, update or clear my keyed states. Am I right? If I want to migrate to different type, e.g. change `string` type to `int` type, how do I archive by using this functionality? It seems that I can't use `key` parameter in `KeyedStateFunction` to access the other state, generated by another state descriptor. Please correct me if I misunderstood. Thank you. Best Regards, Tony Wei 2018-06-09 9:45 GMT+08:00 TechnoMage <[hidden email]>:
|
Tony, Sorry for that, David On Tue, Jun 12, 2018 at 9:56 AM, Tony Wei <[hidden email]> wrote:
David Anderson | Training Coordinator | data Artisans -- Join Flink Forward - The Apache Flink Conference Stream Processing | Event Driven | Real Time |
Free forum by Nabble | Edit this page |