Architecture¶

Nessie builds on the recent ecosystem developments around table formats. The rise of very large metadata and eventually consistent cloud data lakes (S3 specifically) drove the need for an updated model around metadata management. Where consistent directory listings in HDFS used to be sufficient, there were many features lacking. This includes snapshotting, consistency and fast planning. Apache Iceberg was created to help alleviate those problems.

For more insight into why we created Nessie, you can read the founding blog post by one of Nessie’s creators.

Inspiration¶

The Iceberg format relies on a set of metadata files stored with (or near) the actual data tables. This allows Iceberg to fulfill the same role as the Hive Metastore for transactions without the need for expensive metadata scans or centralized planning (see Iceberg performance). This includes things such as partitioning (including hidden partitions), schema migrations, appends and deletes. It does however require a pointer to the active metadata set to function. This pointer allows the Iceberg client to acquire and read the current schema, files and partitions in the dataset. Iceberg currently relies on the Hive metastore or hdfs to perform this role. The requirements for this root pointer store is it must hold (at least) information about the location of the current up-to-date metadata file, and it must be able to update this location atomically. In Hive this is accomplished by locks and in hdfs by using atomic file swap operations. These operations don’t exist in eventually consistent cloud object stores, necessitating a Hive metastore for cloud data lakes. The Nessie system is designed to store the root metadata pointer and perform atomic updates to this pointer, obviating the need for a Hive metastore. Removing the need for a Hive metastore simplifies deployment and broadens the reach of tools that can work with Iceberg tables.

The Nessie service is a lightweight Java-based REST API server. It uses a standard optimistic locking strategy to ensure atomic transactions. This relies on every operation carrying an expected hash state for the store and allows for a very light weight and scalable implementation. The implementation uses configurable authentication (e.g. IAM on AWS, JWT elsewhere) and a configurable backend (currently supporting RocksDB for single-node, and Apache Cassandra, ScyllaDB, Google BigTable, Amazon DynamoDB or MongoDB) and uses the optimistic locking features of cloud based key value stores to ensure scalability across servers. This architecture allows for Nessie to run in a docker container or in a number of other configurations.