Version 3 (modified by flip, 8 years ago) (diff)


Array BLOBs in the Database

Some of our arrays (e.g. the ppms, areas and phases arrays in the simulations table) are stored as SQLite BLOBs. The BLOBs are written in a language-neutral format, meaning that you can read them not only with Python, but with C, MATLAB, Perl, etc. This document describes the format in which they're read and written.

In this context, "array" always means "numpy array".

Python Code

The Python code we use to convert numpy arrays to/from BLOBs are in Look for the functions _numpy_array_to_blob() and _blob_to_numpy_array().

The Raw Data

Each BLOB contains an array of unspecified length. We construct the BLOB with the following algorithm:

  1. Flatten the array into a 1D list.
  2. Use XDR to turn the array into an architecture- and language-neutral format.
  3. Call Python's sqlite3.Binary() method to make the XDR-ed data palatable to SQLite.
  4. Insert the data.

The arrays are represented as XDR variable-length arrays.

The Metadata

Observant readers will note that the description above makes no provision for two important pieces of metadata that one would require to reconstruct a numpy array from the BLOB. Those pieces of metadata are shape and datatype.

As of this writing, we don't explicitly store this information anywhere in the database because our current BLOBs don't need it. (Those BLOBs are the ppms, areas and phases columns of the simulations table.) Those arrays are always 1D and have a datatype of Python float (equivalent to numpy.float64 or a C double). For 1D arrays, the shape is implied by the length which we know from XDR.

In the future, if we have more complicated arrays to store, we might have to consider metadata storage. However, how we'll store it will be decided on a case-by-case basis.

For instance, if we have arrays that always have a shape of (1, 4, N) where N is the only number that varies, then we still won't need to store shape information in the database because it can be inferred from the length of the array (which XDR provides) and the two fixed dimensions. On the other hand, if we have an array where the shape varies with each row, for example (2, X, Y, Z) then we'd need to add a shape column to the table.

Complex Numbers

XDR doesn't understand complex numbers, so when writing complex numbers to BLOBs we first split them into (real, imaginary) pairs.