Edit this page | Blame

Genotype database

GeneNetwork has been using a plain text file format to store genotypes. We are now moving to a fast read-optimized genotype database file format built on LMDB.

To convert a plain text GeneNetwork genotype file `BXD.geno` to a genodb genotype database `bxd`, use the `genodb` CLI tool from cl-gn like so.

./genodb import BXD.geno bxd

genenetwork3 includes a tiny Python library to read the built genodb database. Here is a sample invocation reading the entire matrix, row 17 and column 13 from a database at `/tmp/bxd`.

from gn3 import genodb

with genodb.open('/tmp/bxd') as db:
    matrix = genodb.matrix(db)
    print(genodb.row(matrix, 17))
    print(genodb.column(matrix, 13))

Note: Pjotr has also written an implementation with zig for mgamma.

The rest of this document describes the design and layout of genodb.

Database layout

genodb is an immutable functional database built on the LMDB key-value store. An immutable database may sound like an oxymoron, but is indeed possible and practical. More precisely, in an immutable database, values once put in are never mutated. When a value needs to be changed, a new modified copy of the value is created. The old value is never touched. This immutability means that we can happily pass along and operate on values without worrying whether it may have changed in the underlying database.

To learn the basics of functional databases, see the following talks:

Being a functional database, genodb can store multiple versions of the genotype matrix. These versions are stored efficiently on disk optimizing for disk usage. Two additional copies of the most recent version of the matrix are stored in read-optimized form for fast retrieval.

Row order

Probably caused by a mismatch between cases in GN database and the genotype file. It would be great to have code that would automatically update geno files when we add a sample of known type.

Currently we are juggling 4 genotype formats and soon a 5th. I agree that dynamic genotypes would be nice and we will probably have to look at versioning lmdb for that. If the markers are fixed and we store individuals as rows that could work.


LMDB maps octet vector keys to octet vector values. Any data we put into a LMDB database needs to be encoded to octets (effectively aka bytes). genodb supports the following three data types with their respective encodings.

  • integer: little-endian encoded 64-bit unsigned integer
  • string: UTF-8 encoded without a terminating null character
  • octet vector: no encoding required, written verbatim


The basic unit of storage in the database is a blob. A BLOB is an octet vector PAYLOAD with associated METADATA. To store a blob in the database, we first compute its HASH, and then put PAYLOAD into the database as a <HASH, PAYLOAD> key-value pair. HASH is the SHA256 hash of BLOB (both the octet vector payload and its associated metadata). To compute HASH, we first serialize BLOB into a series of octets, and then hash the resulting octet vector. Precisely, if BLOB contains PAYLOAD and is associated with (KEY, VALUE),... pairs of metadata, then hash(BLOB) is given by

BLOB = blob(payload=PAYLOAD,
            metadata=[(KEY, VALUE),...])
hash(BLOB) = SHA256(concatenate(length(BLOB.payload), BLOB.payload,
                                [length(BLOB.metadata.KEY), BLOB.metadata.KEY,
                                 length(BLOB.metadata.VALUE), BLOB.metadata.VALUE],...))

This encoding of BLOB into octets is one-to-one. So, assuming there are no hash collisions, every BLOB is uniquely mapped to a HASH.

Each piece (KEY, VALUE) of METADATA is stored as a <concatenate(HASH, ":", KEY), VALUE> key-value pair. KEY is a string identifying that piece of metadata. VALUE is a string, an integer, or an octet vector representing the value of that piece of metadata, and is encoded accordingly.

Matrix storage

We store every version of the genotype matrix in the database, each version as a blob. To construct a MATRIX blob, we first store each ROW and COLUMN of the MATRIX as a separate blob. Each row and column is encoded into an octet vector with one octet corresponding to one element of the genotype matrix. Then, MATRIX is constructed as a blob whose octet vector is the concatenation of the hashes of all the row and column blobs. In addition, the number of rows and columns of MATRIX are also stored as metadata.

ROW = blob(payload=ROW-VECTOR, metadata=[])
COLUMN = blob(payload=COLUMN-VECTOR, metadata=[])
MATRIX = blob(payload=concatenate(concatenate(hash(ROW1), hash(ROW2),...),
                                  concatenate(hash(COLUMN1), hash(COLUMN2),...)),
              metadata=[("nrows", NUMBER-OF-ROWS),
                        ("ncols", NUMBER-OF-COLUMNS)])

We repeat this for every version of the genotype matrix, and associate the concatenated hashes of all the matrix blobs with the "all-versions" key by mutation.

    value=concatenate(hash(MATRIX1), hash(MATRIX2),...))

Fast storage for the current matrix

We store two additional copies of the current matrix for fast retrieval. This read-optimized version of the matrix is, essentialy, the matrix in its row-major and column-major forms. The row-major form facilitates fast row reads, and the column-major form facilitates fast column reads. If MATRIX0 is the most recent matrix, then the blob CURRENT_MATRIX stored in the database is given by the following.

CURRENT_MATRIX = blob(payload=concatenate(row-major-encoding(MATRIX0),
                      metadata=[("matrix", hash(MATRIX0))])

The hash of CURRENT_MATRIX is associated with the "current" key by mutation.


Design notes

Note that though even though genodb is a functional immutable database, the setting of the "all-versions" and "current" keys are done by mutation. This is unavoidable. Even a functional database needs to have a tiny amount of state. The trick is to manage and isolate this state so that it is manageable.

The attentive reader familiar with Guix might note the similarities between the layout of the genodb database and that of Guix's /gnu/store. Indeed, both genodb and the Guix store are functional databases. genodb happens to be realized on LMDB, and the Guix store happens to be realized on the filesystem.

Storing both rows and columns of older versions of the genotype matrix is redundant since the columns can be entirely derived from the rows. This is a happenstance due to the evolution of the genotype database layout, and may be removed in the future. Indeed, in the future, the older versions of the matrix could also be stored in compressed form for more efficient storage.

More thoughts

The choice of lmdb for column/row based storage is an extremely good idea! I have been writing code against it and its performance is great.

The storage of both columns and rows for all *versions* of data are perhaps not necessary (as people tend to use the latest and greatest). Having one final matrix blob may also be overkill and may not work for really large datasets. I suggest to store older versions as rows (vectors) only and the current matrix as cols + rows.

The hashing is a good idea, though its value may be limited on updates where changes are sparse. Nevertheless I want to retain this feature.

I will modify the format to retain metadata as more free-flow JSON records. This is useful when switching between languages and allows easy adding of various attributes. There will be a global metadata record and a per matrix metadata record.

We should make it a point that the RDF graph store gets updated when we change one of these files. So they can easily be found with their metadata.

(made with skribilo)