partd 0.2.0

API

1. Create a Partd:

   >>> import partd
   >>> p = partd.File('/path/to/new/dataset/'')

2. Append key-byte pairs to dataset:

   >>> p.append({'x': b'Hello ', 'y': b'123'})
   >>> p.append({'x': b'world!', 'y': b'456'})

3. Get all bytes associated to a set of keys:

   >>> p.get(['y', 'x'])
   [b'123456', b'Hello world!']

4. Idempotently set single key-value pair (no append, no update):

   >>> p.iset('z', b'metadata'])

4. Destroy partd dataset:

   >>> p.drop()

That’s it.

There is no in-memory state.

Implementations

The reference implementation uses file-based locks. This works surprisingly well as long as you don’t do many small writes.

If you do many small writes then you probably want to cache in memory; this is hard to do in parallel while also maintaining consistency. For this we have a centralized server (see partd.Shared) that caches data in memory and writes only large chunks to disk when necessary

• Server Process:

  >>> server = p.Server('/path/to/dataset', 'ipc://server')

• Worker processes:

  >>> p = Shared('ipc://server')
  >>> p.append(...)

Encodings and Compression

Once we can robustly and efficiently append bytes we move on to encoding various things as bytes either with serialization systems like Pickle or MSGPack or with compression routines like zlib, snappy, or blosc. In principle we want to compose all of these choice together

1. Write policy: partd.File, partd.Shared

2. Encoding: partd.Pickle, partd.Numpy

3. Compression: partd.Blosc, partd.Snappy, …

Partd objects compose by nesting for example here we make a shared server that writes snappy compressed numpy arrays:

>>> p = partd.Numpy(partd.Snappy(partd.Shared('foo')))

And here a partd that writes pickle encoded BZ2 compressed bytes directly to disk:

>>> p = partd.Pickle(partd.BZ2(partd.File('foo')))