Threading

Background

Several bug reports have been filed in the past by the users regarding problems related to the impossibility to use PyTables in multi-thread programs.

The problem was mainly related to an internal registry that forced the sharing of HDF5 file handles across multiple threads.

In PyTables 3.1.0 the code for file handles management has been completely redesigned (see the Backward incompatible changes section in Changes from 3.1.0 to 3.1.1) to be more simple and transparent and to allow the use of PyTables in multi-thread programs.

Citing the Changes from 3.1.0 to 3.1.1:

It is important to stress that the new implementation still has an
internal registry (implementation detail) and it is still
**not thread safe**.
Just now a smart enough developer should be able to use PyTables in a
muti-thread program without too much headaches.

A common schema for concurrency

Although it is probably not the most efficient or elegant solution to solve a certain class of problems, many users seems to like the possibility to load a portion of data and process it inside a thread function using multiple threads to process the entire dataset.

Each thread is responsible of:

  • opening the (same) HDF5 file for reading,
  • load data from it and
  • close the HDF5 file itself

Each file handle is of exclusive use of the thread that opened it and file handles are never shared across threads.

In order to do it in a safe way with PyTables some care should be used during the phase of opening and closing HDF5 files in order ensure the correct behaviour of the internal machinery used to manage HDF5 file handles.

Very simple solution

A very simple solution for this kind of scenario is to use a threading.Lock around part of the code that are considered critical e.g. the open_file() function and the File.close() method:

import threading

lock = threading.Lock()

def synchronized_open_file(*args, **kwargs):
    with lock:
        return tb.open_file(*args, **kwargs)

def synchronized_close_file(self, *args, **kwargs):
    with lock:
        return self.close(*args, **kwargs)

The synchronized_open_file() and synchronized_close_file() can be used in the thread function to open and close the HDF5 file:

import numpy as np
import tables as tb

def run(filename, path, inqueue, outqueue):
    try:
        yslice = inqueue.get()
        h5file = synchronized_open_file(filename, mode='r')
        h5array = h5file.get_node(path)
        data = h5array[yslice, ...]
        psum = np.sum(data)
    except Exception as e:
        outqueue.put(e)
    else:
        outqueue.put(psum)
    finally:
        synchronized_close_file(h5file)

Finally the main function of the program:

  • instantiates the input and output queue.Queue,
  • starts all threads,
  • sends the processing requests on the input queue.Queue
  • collects results reading from the output queue.Queue
  • performs finalization actions (threading.Thread.join())
import os
import queue
import threading

import numpy as np
import tables as tb

SIZE = 100
NTHREADS = 5
FILENAME = 'simple_threading.h5'
H5PATH = '/array'

def create_test_file(filename):
    data = np.random.rand(SIZE, SIZE)

    with tb.open_file(filename, 'w') as h5file:
        h5file.create_array('/', 'array', title="Test Array", obj=data)

def chunk_generator(data_size, nchunks):
    chunk_size = int(np.ceil(data_size / nchunks))
    for start in range(0, data_size, chunk_size):
        yield slice(start, start + chunk_size)

def main():
    # generate the test data
    if not os.path.exists(FILENAME):
        create_test_file(FILENAME)

    threads = []
    inqueue = queue.Queue()
    outqueue = queue.Queue()

    # start all threads
    for i in range(NTHREADS):
        thread = threading.Thread(
            target=run, args=(FILENAME, H5PATH, inqueue, outqueue))
        thread.start()
        threads.append(thread)

    # push requests in the input queue
    for yslice in chunk_generator(SIZE, len(threads)):
        inqueue.put(yslice)

    # collect results
    try:
        mean_ = 0.

        for i in range(len(threads)):
            out = outqueue.get()
            if isinstance(out, Exception):
                raise out
            else:
                mean_ += out

        mean_ /= SIZE * SIZE

    finally:
        for thread in threads:
            thread.join()

    # print results
    print('Mean: {}'.format(mean_))

if __name__ == '__main__':
    main()

The program in the example computes the mean value of a potentially huge dataset splinting the computation across NTHREADS (5 in this case) threads.

The complete and working code of this example (Python 3 is required) can be found in the examples directory: simple_threading.py.

The approach presented in this section is very simple and readable but has the drawback that the user code have to be modified to replace open_file() and File.close() calls with their safe version (synchronized_open_file() and synchronized_close_file()).

Also, the solution showed in the example does not cover the entire PyTables API (e.g. although not recommended HDF5 files can be opened using the File constructor) and makes it impossible to use pythonic constructs like the with statement:

with tb.open_file(filename) as h5file:
    do_something(h5file)

Monkey-patching PyTables

An alternative implementation with respect to the Very simple solution presented in the previous section consists in monkey-patching the PyTables package to replace some of its components with a more thread-safe version of themselves:

import threading

import tables as tb
import tables.file as _tables_file

class ThreadsafeFileRegistry(_tables_file._FileRegistry):
    lock = threading.RLock()

    @property
    def handlers(self):
        return self._handlers.copy()

    def add(self, handler):
        with self.lock:
            return super().add(handler)

    def remove(self, handler):
        with self.lock:
            return super().remove(handler)

    def close_all(self):
        with self.lock:
            return super().close_all(handler)

class ThreadsafeFile(_tables_file.File):
    def __init__(self, *args, **kargs):
        with ThreadsafeFileRegistry.lock:
            super().__init__(*args, **kargs)

    def close(self):
        with ThreadsafeFileRegistry.lock:
            super().close()

@functools.wraps(tb.open_file)
def synchronized_open_file(*args, **kwargs):
    with ThreadsafeFileRegistry.lock:
        return _tables_file._original_open_file(*args, **kwargs)

# monkey patch the tables package
_tables_file._original_open_file = _tables_file.open_file
_tables_file.open_file = synchronized_open_file
tb.open_file = synchronized_open_file

_tables_file._original_File = _tables_file.File
_tables_file.File = ThreadsafeFile
tb.File = ThreadsafeFile

_tables_file._open_files = ThreadsafeFileRegistry()

At this point PyTables can be used transparently in example program presented in the previous section. In particular the standard PyTables API (including with statements) can be used in the thread function:

def run(filename, path, inqueue, outqueue):
    try:
        yslice = inqueue.get()
        with tb.open_file(filename, mode='r') as h5file:
            h5array = h5file.get_node(path)
            data = h5array[yslice, ...]
        psum = np.sum(data)
    except Exception as e:
        outqueue.put(e)
    else:
        outqueue.put(psum)

The complete code of this version of the example can be found in the examples folder: simple_threading.py. Python 3 is required.