Support for compaction filtering

[cb1kenobi]

RocksDB allows us to specify a custom compaction filter that is set when the database is opened. This compaction filter would need to read the record's value and determine if it's expired or if the record is a deletion record and the version is older than the retention time.

options.compaction_filter = new CustomCompactionFilter();
// or
options.compaction_filter_factory.reset(new CustomCompactionFilterFactory());

Docs: https://github.com/facebook/rocksdb/wiki/Compaction-Filter

[kriszyp]

Proposed API:
The Database class adds a new setCompactionFilter(filter: CompactionFilterSet): void method. Type is defined as:

// CompactionFilterSet is an array of conditions and if any of the elements (OR) match a record that is being compacted, the record should be discarded
type CompactionFilterSet = CompactionFilterConditions[];

// CompactionFilterConditions is an array of conditions that all must be matched (AND) for the record to be discarded 
type CompactionFilterConditions = CompactionFilterCondition[];

// defines a sequence of bytes that must match, according to the operator, the record, at the position specified.
type CompactionFilterCondition = {
  position: number; // this offset/position in the record to attempt to match
  operator: 'equal' | 'greater_than' | 'less_than' | 'bitmap';
  value: Uint8Array; 
};

Example:

const db = RocksDatabase.open(path);
db.setCompactionFilter([
  // first condition that can trigger discarding:
  [
    { position: 8, operation: 'equal', value: new Uint8Array([0, 0, 0, 0]) }, bytes 8-11 must be zero AND...
    { position: 0, operation: 'less_than', value: new Uint8Array([66,3,2,3,4,5,6,3]) }, // bytes 0-7 must be less than provided bytes 
  ], // OR, second condition that trigger discarding:
  [
    { position: 9, operation: 'bitmap', value: new Uint8Array([8]) }, if byte 9 has bit 3 set (& 8), AND 
    { position: 12, operation: 'less_than', value: new Uint8Array([66,3,2,3,4,5,6,3]) }, // bytes 12-19 must be less than provided bytes 
  ]
]);

All comparisons should be done as big-endian.

[cb1kenobi]

Nice! I like this.

To implement this using a db.setCompactionFilter() method, rocksdb-js would always need to wire up a custom compaction filter regardless if it is being used since this method would be called after the database is opened. Furthermore, what happens if this method is called more than once? Does it reset the existing rules? Do we want/need to support changing rules at runtime?

Maybe we can take the simple route and go with an RocksDatabaseOptions option? Then we don't need to worry about initializing the compaction filter if it's not used.

const myFilter = [ [ /* rule 1 */ ], [ /* rule 2 */ ] ];

const db = RocksDatabase.open('/path/to/db', {
  compactionFilter: myFilter
});

While the database is being opened, rocksdb-js initializes the custom compaction filter on the native side where the rules are baked into what is essentially a vector of vectors.

Would we ever need to disable the compaction filter at runtime? (e.g. return true regardless the record)