This test throws NPE:

public class TestSlowAsyncMapper {
    @Test
    public void foo() {
        var hz = Hazelcast.newHazelcastInstance(newConfig());
        hz.getJet().newJob(newPipeline()).join();
    }

    private Pipeline newPipeline() {
        var p = Pipeline.create();
        p.readFrom(TestSources.itemStream(Integer.MAX_VALUE))
                .withNativeTimestamp

Similar to dask/dask#7800, we can replace our tmpfile context manager

https://github.com/dask/distributed/blob/5e5b9835a5e6929bb97acf94f0a2eecccc0b6a2f/distributed/utils.py#L862-L863

with tempfile.TemporaryFile from the standard library.

It seems that the number of joining clients (not the num of computing clients) is fixed in fedml_api/data_preprocessing/**/data_loader and cannot be changed except CIFAR10 datasets.

Here I mean that it seems the total clients is decided by the datasets, rather the input from run_fedavg_distributed_pytorch.sh.

https://github.com/FedML-AI/FedML/blob/3d9fda8d149c95f25ec4898e31df76f035a33b5d/fed

Summary

The current implementation does not allow creating templates specifying the parallelism field as per the Argo spec.

Use Cases

Many possible use cases, my current one is limiting the amount of parallel nodes running a certain template so as to save

If enter_data() is called with the same train_path twice in a row and the data itself hasn't changed, a new Dataset does not need to be created.

We should add a column which stores some kind of hash of the actual data. When a Dataset would be created, if the metadata and data hash are exactly the same as an existing Dataset, nothing should be added to the ModelHub database and the existing