h5py chunking structured structured numpy arrays

Question

I'm getting data from sensors and storing it in HDF5 files using h5py. The sensordata comes in as a bytes object, I use numpy to convert it into a structured array. Then I write the structured array into the HDF5 file. This all works as intended.

Now I want to read the data back from the HDF5 file and I'm only interested in some parts of it. For example if I only want to read one column. The problem is that directly writing the structured numpy array to the HDF5 file writes all the data as a single block, for example with shape (10,). The default chunksize is set to (256,). This means I read 256 rows and all columns of data for each chunk. However, this gets really slow when the number of columns increases.

Is there a way to modify the data or change the chunking parameters so I can read a single column of data instead of a whole block in each chunk?

A minimal example of what I'm using is shown below:

import h5py
import ctypes
import numpy as np

class SensorStruct(ctypes.Structure):
    _pack_ = 4
    _fields_ = [('tc_time',ctypes.c_int64),
                ('pc_time',ctypes.c_double),
                ('nSample', ctypes.c_ushort),
                ('fMean', ctypes.c_float),
                ('fLowerbound', ctypes.c_float),
                ('fUpperbound', ctypes.c_float)]

def CreateFile(filename):
    #Create a new HDF5 file
    with h5py.File(filename, 'w', libver='latest') as f:
        f.swmr_mode = True

def AddDataset(filename, dsetname, struct):
    #Add a dataset to an existing HDF5 file
    with h5py.File(filename, 'r+', libver='latest', swmr=True) as f:
        f.create_dataset(dsetname, 
                         dtype = struct, 
                         shape = (0,), #Shape will update each time data is added
                         maxshape = (60480000,),
                         chunks = True, #Need to modify this somehow
                         compression = 'gzip')

def WriteData(filename, data):
    #Append an existing dataset with new data
    with h5py.File(filename, 'r+', libver='latest', swmr=True) as f:
        dset = f[dsetname]

        length = dset.shape[0]
        maxlength = dset.maxshape[0]

        newlength = length + len(data)
        if newlength < maxlength:
            dset.resize((newlength,))
            dset[length:newlength] = data

filename = 'TESTFILE.h5'
dsetname = 'Sensor1'
struct_dt = np.dtype(SensorStruct)

#Rawdata comes in from a sensor every few seconds, returns as bytes object
rawdata1 = b"\x15\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x00\x00\x00\x00'u\x1fA\xf4Q\x01AbY?A\x16\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x01\x00\x00\x00\x1c\xe4&A[\x85\x0bA\x97\x96=A\x17\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x02\x00\x00\x00\xf6\x8b\x02A\xe5\xd5\xd5@\xc1Y\x1bA\x18\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x03\x00\x00\x00 \xec9A?W\x17A\xd0vRA\x19\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x04\x00\x00\x00\xf2\t/A\x83U\x19A\r&[A\x1a\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x05\x00\x00\x00s\x8a\x18A\xb0\x19\x04A\xc6\xb51A\x1b\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x06\x00\x00\x00P\xb6>A6\xc5 A)erA\x1c\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x07\x00\x00\x00\xe5e\x11A\x17\x9c\xff@^\xbf5A\x1d\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\x08\x00\x00\x00*\xbe\x19At\xd5\x04AN\x919A\x1e\xcd[\x07\x00\x00\x00\x00 x\x81BA\x02\xd7A\t\x00\x00\x00\xa2* A(-\x03AE\xedFA"
rawdata2 = b'\x1f\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\n\x00\x00\x00\xb6\x89&A\xd7\x8f\x07A\xe9\x00SA \xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x0b\x00\x00\x00\x91I\xfd@*\\\xdc@<\x17!A!\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x0c\x00\x00\x00,q\x12A\x81\x1f\xfa@\x81\xfe(A"\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\r\x00\x00\x00\x04@\x1cA\x03p\x05A\x05\xb03A#\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x0e\x00\x00\x00\xad\x89:A8h#A\xab\x88SA$\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x0f\x00\x00\x00I\x0f\xf5@\x15\xaa\xca@Rk\x0cA%\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x10\x00\x00\x00\xab\xeb\x1dA\x86 \x05A\x1807A&\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x11\x00\x00\x00Q\xda3A\xdc\xa6\x1cAT)ZA\'\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x12\x00\x00\x00U\xb3=A\xae\xcb\x1aA\xebmQA(\xcd[\x07\x00\x00\x00\x00 x\x01EA\x02\xd7A\x13\x00\x00\x00\x8f\x82\x0cA\x11\x15\xf3@$]&A'
data1 = np.frombuffer(rawdata1, dtype=SensorStruct)
data2 = np.frombuffer(rawdata2, dtype=SensorStruct)

CreateFile(filename)
AddDataset(filename, dsetname, SensorStruct)
WriteData(filename, data1)
WriteData(filename, data2)

Here I'm trying to read a single column of data:

import time
t0 = time.time()

with h5py.File(filename, 'r', libver='latest', swmr=True) as f:
    dset = f[dsetname]

    #Optimize chunking so I can read one column
    #My real dataset contains hundreds of columns and milions of rows
    #So this minimal example may look slightly trivial
    print('Chunksize: {}'.format(dset.chunks))
    t = dset['pc_time']

print('Reading the time column took {} seconds'.format(time.time()-t0))

Answer 1

In [551]: dt = np.dtype([('a',int),('b','uint8'),('c','float32'),('d','float64')])
In [552]: x = np.ones(10, dt)
In [553]: x.dtype
Out[553]: dtype([('a', '<i8'), ('b', 'u1'), ('c', '<f4'), ('d', '<f8')])
In [554]: x.itemsize
Out[554]: 21
In [555]: x.__array_interface__
Out[555]: 
{'data': (40185408, False),
 'strides': None,
 'descr': [('a', '<i8'), ('b', '|u1'), ('c', '<f4'), ('d', '<f8')],
 'typestr': '|V21',
 'shape': (10,),
 'version': 3}

Each record this array takes up 21 bytes, 'V21'.

In [557]: f = h5py.File('vtype.h5','w')
In [558]: ds = f.create_dataset('data', data=x)
In [559]: ds
Out[559]: <HDF5 dataset "data": shape (10,), type "|V21">
In [560]: ds.dtype
Out[560]: dtype([('a', '<i8'), ('b', 'u1'), ('c', '<f4'), ('d', '<f8')])

In h5dump this dataset displays as

  DATATYPE  H5T_COMPOUND {
     H5T_STD_I64LE "a";
     H5T_STD_U8LE "b";
     H5T_IEEE_F32LE "c";
     H5T_IEEE_F64LE "d";
  }

The docs on chunking show a chunking tuple with the same number of elements as the array's shape.

http://docs.h5py.org/en/stable/high/dataset.html#chunked-storage

Here I created a 1d array, so chunking, if specified, only applies to that dimension, not to the Compound datatype.

For numpy arrays, accessing a single field of a structured array is relatively fast, comparable to accessing a column of a 2d array, or a stand along 1d array. It is a view.

But loading from a h5 dataset is a copy. With this small example, loading ds[:] is faster than ds['a']. And ds[:n]['a'] is faster than ds['a'][:n].

I don't have a sense of how these timings compare with column access of a simple 2d array. And I don't know if the timings depend on the size of the dtype.