Lazy Evaluation

You need to call .execute() on Mars tensors, DataFrames and remote functions to trigger execution.

>>> import mars.tensor as mt
>>> import mars.dataframe as md
>>> df = md.DataFrame(mt.random.rand(3, 3))
>>> df
DataFrame <op=DataFrameFromTensor, key=182b756be8a9f15c937a04223f11ffba>
>>> df.execute()
          0         1         2
0  0.167771  0.568741  0.877450
1  0.037518  0.796745  0.072169
2  0.052900  0.936048  0.307194

Calling .execute() will return Mars object itself, .fetch() could be called on executed objects to get the result.

>>> import mars.remote as mr
>>> f = mr.spawn(lambda x: x + 1, args=(10,))
>>> f.execute()
Object <op=RemoteFunction, key=8a9ef53cb93cd7698d71512ec813682b>
>>> f.fetch()
11

However, there are exceptions that some functions will trigger execution intermediately.

• Iterating over DataFrame, including mars.dataframe.DataFrame.iterrows() and mars.dataframe.DataFrame.itertuples().

• All plot functions for DataFrame and Series, including mars.dataframe.DataFrame.plot(), mars.dataframe.DataFrame.plot.bar() and so forth.

• All functions in Mars learn like fit, predict and so forth.

Asynchronous Execution

Note

New in version 0.5.2

Specifying wait=False can make the execution asynchronous, it will return a Future object.

>>> import mars.tensor as mt
>>> a = mt.random.rand(100, 10)
>>> future = a.mean().execute(wait=False)
>>> future.done()
True
>>> future.result()
0.49123541512823077

Mars Remote Eager Execution