Fedimint uses a simple key-value store as its database. In theory any such KV store with the following features can be used:
- insert, update, delete actions
- transactions
- key prefix search
- optimistic transactions
We currently use RocksDB on both the client and the server.
The database is logically partitioned based on the module instance id of the module. Each module's keyspace is split up based on prefixing. Each prefix within a module identifies a logical entity. Entity prefixes between modules can overlap because they are logically partitioned by the module instance id. The key value pairs that belong to consensus do not belong to any specific module, so they are not prepended with a module prefix. Below is the format for prefixes within a module:
<2 BYTE MODULE ID>
Example for the Mint Module: 0xFF 0x00 0x00 0x10 In the above example, the module instance id = 0 and the entity it identifies is a NoteNonce, because it uses the 0x10 byte. 0xFF is the global prefix byte that identifies this as module data.
Example for consensus data: 0x01 In the above example, because the consensus data does not apply to any specific module, the global prefix byte and module instance id prefixes are missing.
The client uses the same isolation mechanism as fedimintd
to store data for each module.
In Fedimint, all interactions with the database use a database transaction. Database transactions are an abstraction for accessing the database in an atomic, consistent, and isolated way. Underneath, Fedimint uses RocksDb's optimistic transactions which means database transactions are allowed to read and write to the database concurrently. If there are two concurrent transactions that modify the same key, RocksDb's optimistic transactions will detect this "write-write" conflict and cause the transaction that commits second to fail.
Fedimint has defined a number of different structs for implementing the necessary functionality for database transactions. These structs follow the adapter pattern to wrap and isolate the features. At the bottom is an explanation of each interface/struct.
In order to avoid breaking changes, fedimintd
, gatewayd
, and the client must know of the structure of the data written to disk. If a code upgrade
has occurred, it is possible that the new version of the code expects the data written to disk to be structured differently. When this happens, a database
migration must occur to maintain backwards compatibility. Migrations are defined on a per-module basis in the get_database_migrations
function and applied
using apply_migrations
.
Since introducing a database breaking change is easy (just modifying a struct), tests have been introduced to catch DB breaking changes. just prepare_db_migration_snapshots
will prepare a database backup of dummy data for a module. test_migrations
will try to read from this database backup. If the
structure of the data has changed and the backup cannot be reading, this test will fail.
There are some times when making a DB breaking change (not backwards compatible) is intentional. In that case, to fix the migration tests, just prepare_db_migration_snapshot
needs to be updated
to reflect the new structure of the data. Then, the db/ folder at the root of the repository needs to be deleted. Then just prepare_db_migration_snapshot
can
be run to re-generate the database backup. test_migrations
will need to be updated to read the newly added/modified data.
IRawDatabase
andIRawDatabaseTransaction
- The interfaces raw database crates implement.IDatabase
andIDatabaseTransaction
- The interfaces including key subscribe & notify functionality, added on top of databases.IDatabaseTransactionOps
andIDatabaseTransactionCoreOps
- The interfaces of database transaction operationsIDatabaseTransactionOpsCoreTyped
- LikeIDatabaseTransactionOps
but with typed keys and values. Implemented generically over everything implementsIDatabaseTransactionOps
.
Database
andDatabaseTransaction
- Public facing newtypes overIDatabase
andIDatabaseTransaction
that also holdsdecoders
and minor helper logic.DatabaseTransactionRef
- A logical reference toDatabaseTransaction
that does not expose the commit operation.
BaseDatabase
andBaseDatabaseTransaction
- Adapter implementingIDatabase
forIRawDatabase
PrefixDatabase
andPrefixDatabaseTransaction
- Adapter overIDatabase
andIDatabaseTransaction
implementing key prefix handling to provide database partitioning/isolation.
MemDatabase
andMemDatabaseTransaction
- Base implementation of an in-memory database transaction.RocksDbDatabase
andRocksDbDatabaseTransaction
- Base implementation of a Rocksdb database. Uses optimistic transaction internally.RocksDbReadOnly
andRocksDbReadOnlyTransaction
- Base implementation of a Rocksdb read-only database. Will panic on writes.
classDiagram
DatabaseTransaction ..* IDatabaseTransaction : wraps
DatabaseTransactionRef ..* DatabaseTransaction : wraps
PrefixDatabaseTransaction ..|> IDatabaseTransaction : implements
PrefixDatabaseTransaction ..* IDatabaseTransaction : wraps
BaseDatabaseTransaction ..|> IDatabaseTransaction : implements
BaseDatabaseTransaction ..* IRawDatabaseTransaction : wraps
MemTransaction ..|> IRawDatabaseTransaction : implements
RocksDbTransaction ..|> IRawDatabaseTransaction : implements
DatabaseTransactionRef ..|> IDatabaseTransactionOpsCore : implements
DatabaseTransaction ..|> IDatabaseTransactionOpsCore : implements
BaseDatabaseTransaction ..|> IDatabaseTransactionOpsCore : implements
PrefixDatabaseTransaction ..|> IDatabaseTransactionOpsCore : implements
MemTransaction ..|> IDatabaseTransactionOpsCore : implements
RocksDbTransaction ..|> IDatabaseTransactionOpsCore : implements
class IDatabaseTransactionOpsCore {
<<interface>>
+ raw_insert_bytes()
+ raw_get_bytes()
+ raw_remove()
}
class IDatabaseTransaction {
<<interface>>
+ commit_tx()
}
class DatabaseTransaction {
- IDatabaseTransaction
}
class DatabaseTransactionRef {
- &DatabaseTransaction
}
class PrefixDatabaseTransaction {
- IDatabaseTransaction
}
class BaseDatabaseTransaction {
- IRawDatabaseTransaction
}
class RocksDbTransaction {
}
class MemTransaction {
}