can it run use ParameterServerStrategy

[Coder-Yifan]

I found all examples use MultiWorkerMirroredStrategy

[leewyang]

Yes, in theory, it should be supported, especially when using InputMode.TENSORFLOW (where Spark is really just being used as a cluster/resource manager to start up the nodes of your TF cluster).

That said, the ParameterServerStrategy has been marked experimental since the DistributionStrategies were initially introduced several years ago, so there hasn't been a lot of emphasis on it, and I would proceed with caution when using it.

[Coder-Yifan]

But tf.distribute.experimental.ParameterServerStrategy needs parameter “cluster_resolver”，how can i pass the parameter in main_fuc

cluster = TFCluster.run(sc, main_fun, args, args.cluster_size, num_ps=2, tensorboard=args.tensorboard,input_mode=TFCluster.InputMode.TENSORFLOW, master_node='chief')

def main_fun(args, ctx):
    """Example demonstrating loading TFRecords directly from disk (e.g. `HDFS)` without tensorflow_datasets."""
    strategy = tf.distribute.experimental.ParameterServerStrategy()#cluster_resolver```

[leewyang]

Pretty sure this should work, since we use TF_CONFIG to set up the cluster.