Ankur Goenka noticed through debugging that multiple instances of BatchFlinkExecutableStageContext.BatchFactory are loaded for a given job when executing in standalone cluster mode. This context factory is responsible for maintaining singleton state across a TaskManager (JVM) in order to share SDK Environments across workers in a given job. The multiple-loading breaks singleton semantics and results in an indeterminate number of Environments being created.
It turns out that the Flink classloading mechanism is determined by deployment mode. Note that "user code" as referenced by this link is actually the Flink job server jar. Actual end-user code lives inside of the SDK Environment and uploaded artifacts.
In order to maintain singletons without resorting to IPC (for example, using file locks and/or additional gRPC servers), we need to force non-dynamic classloading. For example, this happens when jobs are submitted to YARN for one-off deployments via `flink run`. However, connecting to an existing (Flink standalone) deployment results in dynamic classloading.
We should investigate this behavior and either document (and attempt to enforce) deployment modes that are consistent with our requirements, or (if possible) create a custom classloader that enforces singleton loading.