Hey all, I’m building a Flink app that pulls in messages from a Kafka topic and writes them out to disk using a custom bucketed sink. Each message needs to be parsed using a schema that is also needed when writing in the sink. This schema is read from a remote file on a distributed file system (it could also be fetched from a service). The schema will be updated very infrequently.
In order to support schema evolution, I have created a custom source that occasionally polls for updates and if it finds one parses the new schema and sends a message containing the serialized schema. I’ve connected these two streams and then use a RichCoFlatMapFunction to flatten them back into a single output stream (schema events get used to update the parser, messages get parsed using the parser and emitted). However, I need some way to communicate the updated schema to every task of the sink. Simply emitting a control message that is ignored when writing to disk means that only one sink partition will receive the message and thus update the schema. I thought about sending the control message as side output and then broadcasting the resulting stream to the sink alongside the processed event input but I couldn’t figure out a way to do so. For now, I’m bundling the schema used to parse each event with the event, storing the schema in the sink, and then checking every event’s schema against the stored schema but this is fairly inefficient. Also, I’d like to eventually increase the types of control messages I can send to the sink, some of which may not be idempotent. Is there a better way to handle this pattern? (Bonus question: ideally, I’d like to be able to perform an action when all sink partitions have picked up the new schema. I’m not aware of any way to emit metadata of this sort from Flink tasks beyond abusing the metrics system. This approach still leaves open the possibility of tasks picking up the new schema and then crashing for unrelated reasons thus inflating the count of tasks using a specific schema and moreover requires tracking at least the current level of parallelism and probably also Flink task state outside of Flink. Are there any patterns for reporting metadata like this to the job manager?) I’m using Flink 1.8.
