Get dask-gateway scheduler address

[AlbertDeFusco]

When connecting to a dask-gateway the client.scheduler_address is a proxy address

>>>client.scheduler.address
'gateway://dask.training.anaconda.com:8786/4fd53916f0214703934701aa7a7eaf85'

I was able to solve this with the following in core::_train with client.scheduler_info()['address'])

    # Start the XGBoost tracker on the Dask scheduler
    host, port = parse_host_port(client.scheduler_info()['address'])
    env = yield client._run_on_scheduler(
        start_tracker, host.strip("/:"), len(worker_map)
    )

However, I get the following warning.

>>> from dask_xgboost import XGBRegressor
>>> xgb = XGBRegressor()
>>> xgb.fit(X, y)

/Users/adefusco/Applications/miniconda3/envs/xgb/lib/python3.7/site-packages/distributed/client.py:3299: RuntimeWarning: coroutine 'Client._update_scheduler_info' was never awaited
  self.sync(self._update_scheduler_info)
RuntimeWarning: Enable tracemalloc to get the object allocation traceback

I have verified that his update works correctly on a 9m row training set and scales linearly from 4 to 8 workers (2cores/worker). Is this the correct approach to get the actual scheduler address?

[TomAugspurger]

I'll look into this soon. I'm planning a refactor to move this logic into distributed itself.

[jrdzha]

Hi, is there any update on this issue? I'm using the Dask implementation in XGBoost itself, rather than this library, so my feeling is this may be a bug with Dask rather than XGBoost?

I'm using LocalCluster with Dask 2.28.0.

dtrain = xgb.dask.DaskDMatrix(client, X, y)

output = xgb.dask.train(
    client,
    {
        'verbosity': 2,
         'tree_method': 'hist',
         'objective': 'binary:logistic'
    },
    dtrain,
    num_boost_round=4, 
    evals=[(dtrain, 'train')]
)

/root/anaconda3/lib/python3.7/site-packages/distributed/client.py:3530: RuntimeWarning: coroutine 'Client._update_scheduler_info' was never awaited
  self.sync(self._update_scheduler_info)