hey everyone, we are working on migrating our Flink deployments from AWS EMR to Kubernetes (using the operator). On EMR, we are using emrfs to upload checkpoints to S3 whereas in Kubenetes, we are using Presto. We are trying to figure out a rollback plan where-in EMR is able to restore checkpoints created in Kubernetes in case we run into problems. When we tested this, the job in EMR is failing due to rate limit errors because it's using the Presto plugin and that needs some fine tuning. Perhaps it's performing some form of auto-discovery via

s3://

prefix passed for checkpoint location. Is there a way we can indicate to Flink to use emrfs like by changing the s3 prefix?

• Flink conf (/etc/flink/conf/flink-conf.yaml):

fs.allowed-fallback-filesystems: s3;s3p

• Hadoop core site configuration (/etc/hadoop/conf/core-site.xml):

<property>
    <name>fs.s3p.impl</name>
    <value>com.amazon.ws.emr.hadoop.fs.EmrFileSystem</value>
</property>

With the above configuration Flink will fallback on the hadoop filesystem for the

s3p

scheme. The Hadoop filesystem, is then instructed to leverage EMRFS as implementation for the

s3p

scheme.