This is a simple guide that demonstrates how to use TileDB on Google Cloud Storage (GCS). After setting up TileDB to work with GCS, your TileDB programs will function properly without any API change! Instead of using local file system paths when creating/accessing groups, arrays, and VFS files: use URIs that start with
gcs://. For instance, if you wish to create (and subsequently write/read) an array on GCS, you use URI
gcs://<your-bucket>/<your-array-name> for the array name.
Navigate to the Google Cloud Console.
Create a service account.
Set your authentication environment variable.
Configure your GCP project id by setting the following key in a configuration object (see Configuration):
TileDB will automatically authenticate using the configuration file referenced by the environment variable set in step 3. Bucket operations will be performed within the namespace of your configured project id.
So far we explained that TileDB arrays and groups are stored as directories. There is no directory concept on GCS and other similar object stores. However, GCS uses character
/ in the object URIs which allows the same conceptual organization as a directory hierarchy in local storage. At a physical level, TileDB stores on GCS all the files it would create locally as objects. For instance, for array
gcs://bucket/path/to/array, TileDB creates array schema object
gcs://bucket/path/to/array/__array_schema.tdb, fragment metadata object
gcs://bucket/path/to/array/<fragment>/__fragment_metadata.tdb, and similarly all the other files/objects. Since there is no notion of a “directory” on GCS, nothing special is persisted on GCS for directories, e.g.,
gcs://bucket/path/to/array/<fragment>/ does not exist as an object.
TileDB writes the various fragment files as append-only objects using the insert object API of the Google Cloud C++ SDK. In addition to enabling appends, this API renders the TileDB writes to GCS particularly amenable to optimizations via parallelization. Since TileDB updates arrays only by writing (appending to) new files (i.e., it never updates a file in-place), TileDB does not need to download entire objects, update them, and re-upload them to GCS. This leads to excellent write performance.
TileDB reads utilize the range GET request API of the GCS SDK, which retrieves only the requested (contiguous) bytes from a file/object, rather than downloading the entire file from the cloud. This results in extremely fast subarray reads, especially because of the array tiling. Recall that a tile (which groups cell values that are stored contiguously in the file) is the atomic unit of IO. The range GET API enables reading each tile from GCS in a single request. Finally, TileDB performs all reads in parallel using multiple threads, which is a tunable configuration parameter.