Version updates - Development| Alibaba Cloud Documentation Center

This topic describes the latest version updates of Mars, including the new and enhanced features in a specific version of Mars.

The following table describes the latest version updates of Mars. For more information, see the URL of the related version.


Version	Change type	Description
v0.6.3	New feature	DataFrame: More features are added for `md.Index`. For more information, see Add more functionalities for `md.Index`(#1864). `{DataFrame,Series}.rename_axis` is supported. For more information, see Implements `{DataFrame,Series}.rename_axis`(#1870).
v0.6.3	Enhanced feature	JSON can be used for internal serialization. For more information, see Allow internal serialization to use JSON(#1882). The performance of `{md.read_csv()` and `md.read_parquet()}.head()` is optimized. For more information, see Optimize performance of {md.read_csv(), md.read_parquet()}.head()(#1883). The performance of `df.sort_values().head()` is optimized. For more information, see Optimize performance of `df.sort_values().head()`(#1888). Column pruning is supported for the data sources that use `groupby().agg()`. For more information, see Support column pruning for groupby().agg() on data sources(#1889). `named_{dataframe, series, tensor}` is optimized to obtain more metadata. For more information, see Improve `named_{dataframe, series, tensor}` that it's able to get more meta(#1897).
v0.6.2	New feature	DataFrame `df.groupby().head()` is supported. For more information, see Implements head() on groupby objects(#1851). Learn `mars.learn.preprocessing.{MinMaxScaler, minmax_scale}` is supported. For more information, see Implements `mars.learn.preprocessing.{MinMaxScaler, minmax_scale}'(#1858).
v0.6.2	Enhanced feature	The `recall_by_id` computation method of Proxima is optimized. For more information, see Improve proxima `recall_by_id` computation method(#1805). The bug that exists in reading from and writing to vineyard of tensors and DataFrame is fixed. For more information, see Revise to/from vineyard, of Tensor and DataFrame(#1806). A memory prediction mechanism similar to `read_csv` is added for `read_parquet`. For more information, see Add memory estimation for `read_parquet` as well as `read_csv`(#1815). Compound aggregate functions can be used in Lambda. For more information, see Support using compound agg function in lambda(#1819). The incremental_index field is added to `reset_index`. The default value of this field is False. For more information, see Add `incremental_index` argument to `reset_index` which by default is False(#1842). The `to_pandas` method of DataFrame and Series obtains data in batches. For more information, see Support `to_pandas` in a batch way for DataFrame and Series(#1859). You can specify a memory scale in Kubernetes. For more information, see Support specifying memory scale in kubernetes(#1861).
v0.6.1	New feature	Tensor When a tensor is converted into a sparse tensor, values other than 0 can be specified for missing parameters. For more information, see Support 'missing' argument for 'tensor.tosparse()' and 'fill_value' argument for 'sparse_tensor.todense()'(#1802). DataFrame The replace API is added. For more information, see Implements {DataFrame,Series}.replace(#1765). The cartesian_chunk API is added to calculate the Cartesian product of two DataFrame objects and apply a function to each chunk. This API is an extension of the pandas API. For more information, see Add '{DataFrame, Series}.cartesian_chunk' support(#1777). String concatenation is optimized for aggregation and group aggregation. For more information, see Integrate 'str.cat' into reduction and groupby-aggregation(#1781). The level parameter is supported in aggregate functions. For more information, see Implements reduction with 'level' argument(#1784).
v0.6.0	Compatibility	Aggregation and group aggregation are rewritten. If the Mars client of an earlier version connects to the Mars server of this version, these features may fail.
	New feature	DataFrame: The `num_partitions` parameter is supported during DataFrame initialization. For more information, see Support 'num_partitions' argument for DataFrame initializers(#1733). Named aggregation is supported. For more information, see Add support for named aggregations(#1748).
	Enhanced feature	The `groupby()` and agg() functions are unified in ReductionCompiler. For more information, see Unify groupby.agg() using ReductionCompiler(#1739).
	Fixed issue	md.read_csv is fixed to prevent the error that may occur if both names and usecols are specified. For more information, see Fix 'md.read_csv' when names and usecols specified(#1738). The string dtype is supported for tensor reduction and PSRS balancing. For more information, see Support string dtype for tensor reductions(#1746). XGBoost and LightGBM on DataFrame are fixed. For more information, see Fix xgboost and lightgbm on DataFrames(#1751). The repeated running of the same code in distributed mode is fixed. For more information, see Fix repeated running of same code in distributed mode(#1753). A column in DataFrame can be set to a tensor of the scalar type. For more information, see Support setting scalar which is a tensor for DataFrame(#1758).