Volatile on-disk database of blocks

Logic

The VolatileDB is a key-value store of blocks indexed by their hashes. It is parameterised by the block type blk.

The "volatile" in the name refers to the fact that the blocks stored in it make up the volatile part of the chain, i.e., the last k blocks of the chain, which can still be rolled back. Not only the last k blocks of the current chain are stored in this database, but also blocks of forks which we have switched from or will switch to.

The VolatileDB appends new blocks sequentially to a file. When volMaxBlocksPerFile are stored in the current file, a new file is started.

The VolatileDB provides four main operations:

1. Adding blocks with putBlock
2. Get blocks or information about them with getBlockComponent
3. Accessing the in-memory indices using getBlockInfo and filterByPredecessor
4. Garbage collecting blocks older than a given slot using garbageCollect

Garbage collection will only delete a file from the VolatileDB when all blocks in it have a slot older than the one passed to garbageCollect.

Errors

When an exception occurs while modifying the VolatileDB, we close the database as a safety measure, e.g., in case a file could not be written to disk, as we can no longer make sure the in-memory indices match what's stored on the file system. When reopening, we validate the blocks stored in the file system and reconstruct the in-memory indices.

NOTE: this means that when a thread modifying the VolatileDB is killed, the database will be closed. This is an intentional choice to simplify things.

The in-memory indices can always be reconstructed from the file system. This is important, as we must be resilient against unexpected shutdowns, power losses, etc.

We achieve this by only performing basic operations on the VolatileDB: * putBlock only appends a new block to a file. Losing an update means we only lose a block, which is not a problem, it can be redownloaded. * garbageCollect only deletes entire files. * There is no operation that modifies a file in-place. This means we do not have to keep any rollback journals to make sure we are safe in case of unexpected shutdowns.

We only throw VolatileDBError. File-system errors are caught, wrapped in a VolatileDBError, and rethrown. We make sure that all calls to HasFS functions are properly wrapped. This wrapping is automatically done when inside the scope of modifyOpenState and withOpenState. Otherwise, we use wrapFsError.

Concurrency

A single folder should only be used by a single VolatileDB. Naturally, a VolatileDB can be accessed concurrently by multiple threads.

File-system layout:

The on-disk representation is as follows:

dbFolder/
  blocks-0.dat
  blocks-1.dat
  ...

Files not fitting the naming scheme are ignored. The numbering of these files does not correlate to the blocks stored in them.

Each file stores a fixed number of blocks, specified by volMaxBlocksPerFile. When opening the VolatileDB, it will start appending to the file with the highest number that is not yet full. If all are full or none exist, a new file will be created.

There is an implicit ordering of block files, which is NOT alpharithmetic. For example, blocks-20.dat < blocks-100.dat.

Recovery

The VolatileDB will always try to recover to a consistent state even if this means deleting all of its contents. In order to achieve this, it truncates the files containing blocks if some blocks fail to parse, are invalid, or are duplicated.