Thanks for sharing! The first thing that comes to my mind is the Flink Kubernetes Operator [1]. If you're deploying on Kubernetes, it's worth considering the operator to manage your SQL jobs. If you want an end-to-end solution, you may also take a look at StreamPark[2].
Best, Shengkai [1] https://github.com/apache/flink-kubernetes-operator/tree/main/examples/flink-sql-runner-example [2] https://github.com/apache/streampark 王召 <[email protected]> 于2026年6月17日周三 11:21写道： > Hi Flink users, > > Sharing something that solved a recurring problem for us in production, > in case others are hitting the same wall. > > 1. THE PROBLEM > > In Flink 1.20, there is no supported way to deploy a .sql file containing > both DDL and DML statements via "flink run" in Application or Per-Job mode. > SQL Client's -f flag only works in Session mode; the Application mode > equivalent (FLIP-480) is implemented in Flink 2.x but not backported. > > For teams not ready to migrate to Flink 2.x which is a significant > breaking-change upgrade. This means either writing Java wrappers around > each SQL statement, or living with Session mode’s resource-sharing > limitations. > > 2. WHAT WE BUILT > > A small launcher JAR that fills this gap: > > $FLINK_HOME/bin/flink run \ > --target yarn-application \ > flink-sql-bootstrap.jar \ > --script-file hdfs://warehouse/jobs/dwd_orders.sql > > A single command, with DDL + DML in one file, running in Application mode. > It also supports: > > - Catalog snapshots: pre-register tables, views, and UDFs from a JSON file > so SQL scripts contain zero DDL > - Per-operator resource tuning: set parallelism, CPU, and memory per > operator via a JSON config, filling the gap between Flink SQL and > DataStream-level resource control > - Dry-run modes: --validate (syntax check, ~2s, no cluster needed) and > --compile (outputs the optimized plan JSON), useful for CI/CD pipelines > > Verified on Flink 1.20.4, 2.0.2, 2.1.1, and 2.2.0. > > Repo: https://github.com/tonyabasy/flink-sql-bootstrap > > I'm curious whether others have encountered the same deployment > challenges, and what workarounds you've been using. Also happy to discuss > if this approach could be useful in your setup. > > Best, > Zhao Wang >
