Library and CLI for storing numeric data frames in HDF5

pip install h5df==0.1.5



Python library and CLI for storing numeric data frames in HDF5.


Pandas has utilities for storing data frames in HDF5, but it uses PyTables under the hood, which means it is limited to frames with a relatively low number of columns (low 1000s).

This library is intended for storing and querying arbitrarily large numeric matrices which have row and column names. It has a CLI which can export/import to/from delimited text, or it can be used from within Python with tight integration with Pandas.

This library stores only numeric matrices, so it cannot handle data frames with mixed types (e.g., some strings and some numbers).


From PyPI:

pip install h5df

Latest version:

git clone
cd h5df
python install --user

This installs the CLI script "h5df", and a Python module with the same name.


$ cat in.txt

    A   B   C
X   1   2   3
Y   4   5   5

$ h5df load foo.h5 /my/path < in.txt
$ h5df dump foo.h5 /my/path

    A   B   C

To select an individual row or column, use "h5py row|column":

$ h5df row foo.h5 X

CLI flags

Use h5df <cmd> --help for a full listing of options, but a few useful ones:

  • h5df load -v : will output progress as a matrix is loaded (every 100 rows)
  • h5df <any output command> -p N will output values with decimal precision N


The two main classes are h5df.Store and h5df.Frame, representing a HDF5 file and individual data frame, respectively. Here is some example usage:

>> from h5df import Store
>> import pandas as pd
>> import numpy as np
>> np.random.seed(0)

# Create a Store object; the default mode is read-only.
# See for available modes
>> store = Store("test.h5df", mode="a")
>> index = ["A","B","C"]
>> columns = ["V","W","X","Y","Z"]
>> mkdf = lambda: pd.DataFrame(np.random.random((3,5)), index=index, columns=columns)
>> store.put("/frames/1", mkdf())
>> store.put("/frames/2", mkdf())

# Iterate through HDF5 paths corresponding to Frame objects
>> for key in store: print(key)

>> df1 = store["/frames/1"]

# Various selection options

# returns pandas.Series
>> df1.column("W")
>> df1.row("A")

# returns a pandas.DataFrame
>> df1.rows(["A","C"])
>> df1.cols(["W","Y"])

# Returns the whole Frame as a pandas.DataFrame
>> df1.to_frame()

The full list of methods supported by h5df.Frame is:

  • Frame.row(key) and Frame.col(key) - return a pandas.Series corresponding to the row/column
  • Frame.rows(keys) and Frame.cols(keys) - given a list of row/column index names, return an in-memory pandas.DataFrame corresponding to the subset of the overall Frame containing the desired rows or columns
  • Frame.shape - returns a tuple of (# rows, # columns)
  • Frame.to_frame() - return the entire Frame as an in-memory pandas.DataFrame. Make sure you have enough memory!
  • Frame.add(key, data) - add a new row to the matrix with the given unique key. Due to the way of

Storage format

Each h5df.Frame is stored as an HDF5 Group containing 3 Datasets: index and columns (both are 1D arrays of 8-byte integers or UTF-8 encoded binary strings), and data (a 2D double array).

The Group also contains a few HDF5 attributes: - h5df.index_type and h5df.columns_type : a string, either "str" or "int",

marking the data type of each of the corresponding indices
  • h5df.is_frame : a boolean, always set to true, which indicates that this Group contains valid Frame data

Because of this design, it is possible to store a Frame "inside" the Group containing another Frame, but is not recommended in case of future format changes (and because it is confusing).

Performance notes

Data is indexed row-major. Thus row-based queries will be much faster. Generally you should pre-transpose your matrix before putting it into the Store to ensure that the most frequently queried axis will be on the rows.

The h5df.Store() constructor takes a keyword argument, "driver". The full description of available drivers is at . For Linux systems, the default stdio-based driver is "sec2", whereas "core" will memory-map the whole HDF5 file. If your system supports it and the file is frequently used (and therefore will be in your OS page cache), "core" may be faster, especially for reads.


Currently there is no way to select rows by numeric index location (i.e., the equivalent to pandas.DataFrame.iloc).

Encoding and decoding indices (from unicode to binary) is a little slow, meaning that quick queries are slower than they could be.

Iterating through the frames in a HDF5 file, Store.__iter__ is quite inefficient if the file contains large numbers of frames.

For Frame.dump(), output formatting is not vectorized (slower than necessary).

String indices are stored as np.dtype("|S100") encoded as "utf-8". This has several practical consequences:

  1. index and column names are currently limited to 100 UTF-8 characters
  2. UTF-8 encoding is hardcoded and other encodings are not supported (thus, characters from other encodings that will fail str.encode("utf-8") will cause an error.

There are plans to fix these limitations in future versions.

Potential gotchas

When matrices are renamed or deleted using Store.rename() or Store.delete(), existing Frame objects based on this data will not be notified of this change and their behavior is undefined. Most likely an attempt to use dangling Frame objects will result in an error, but may return erroneous results for some methods.

It is up to the user to avoid this situation. When renaming a Frame, the user should subsequently get a new Frame object from the destination path using Store.__getitem__ if they plan to continue to use the data.