Iteratively writing to HDF5 Stores in Pandas

Question

Pandas has the following examples for how to store Series, DataFrames and Panelsin HDF5 files:

Prepare some data:

In [1142]: store = HDFStore('store.h5')

In [1143]: index = date_range('1/1/2000', periods=8)

In [1144]: s = Series(randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [1145]: df = DataFrame(randn(8, 3), index=index,
   ......:                columns=['A', 'B', 'C'])
   ......:

In [1146]: wp = Panel(randn(2, 5, 4), items=['Item1', 'Item2'],
   ......:            major_axis=date_range('1/1/2000', periods=5),
   ......:            minor_axis=['A', 'B', 'C', 'D'])
   ......:

Save it in a store:

In [1147]: store['s'] = s

In [1148]: store['df'] = df

In [1149]: store['wp'] = wp

Inspect what's in the store:

In [1150]: store
Out[1150]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df            frame        (shape->[8,3])  
/s             series       (shape->[5])    
/wp            wide         (shape->[2,5,4])

Close the store:

In [1151]: store.close()

Questions:

1. In the code above, when is the data actually written to disk?

2. Say I want to add thousands of large dataframes living in .csv files to a single .h5 file. I would need to load them and add them to the .h5 file one by one since I cannot afford to have them all in memory at once as they would take too much memory. Is this possible with HDF5? What would be the correct way to do it?

3. The Pandas documentation says the following:

   "These stores are not appendable once written (though you simply remove them and rewrite). Nor are they queryable; they must be retrieved in their entirety."

   What does it mean by not appendable nor queryable? Also, shouldn't it say once closed instead of written?

Solution 1

1. As soon as the statement is exectued, eg store['df'] = df. The close just closes the actual file (which will be closed for you if the process exists, but will print a warning message)

2. Read the section http://pandas.pydata.org/pandas-docs/dev/io.html#storing-in-table-format

   It is generally not a good idea to put a LOT of nodes in an .h5 file. You probably want to append and create a smaller number of nodes.

   You can just iterate thru your .csv and store/append them one by one. Something like:

   for f in files:
     df = pd.read_csv(f)
     df.to_hdf('file.h5',f,df)
   

   Would be one way (creating a separate node for each file)

3. Not appendable - once you write it, you can only retrieve it all at once, e.g. you cannot select a sub-section

   If you have a table, then you can do things like:

   pd.read_hdf('my_store.h5','a_table_node',['index>100'])
   

   which is like a database query, only getting part of the data

   Thus, a store is not appendable, nor queryable, while a table is both.

Solution 2

Answering question 2, with pandas 0.18.0 you can do:

store = pd.HDFStore('compiled_measurements.h5')
for filepath in file_iterator:
    raw = pd.read_csv(filepath)
    store.append('measurements', raw, index=False)

store.create_table_index('measurements', columns=['a', 'b', 'c'], optlevel=9, kind='full')
store.close()

Based on this part of the docs.

Depending on how much data you have, the index creation can consume enormous amounts of memory. The PyTables docs describes the values of optlevel.