CASSANDRA-9738 may not make it into 3.0rc1, but having an off-heap key-cache is still a goal, we should change the index file format to meet off-heap requirements (so I've set fixver to 3.0rc1).
Off-heap (and mmap'd index files) need the offsets of the individual IndexInfo objects and the at least the offset field of IndexInfo structures.
The format I propose is as follows:
(long) position (vint since 3.0, 64bit before)
(int) serialized size of data that follows (vint since 3.0, 32bit before)
-- following for indexed entries only (so serialized size > 0)
(long) header-length (vint since 3.0)
(int) DeletionTime.localDeletionTime (32 bit int)
(long) DeletionTime.markedForDeletionAt (64 bit long)
(int) number of IndexInfo objects (vint since 3.0, 32bit before)
(*) serialized IndexInfo objects, see below
(*) offsets of serialized IndexInfo objects, since version "ma" (3.0)
Each IndexInfo object's offset is relative to the first IndexInfo object.
(*) IndexInfo.firstName (ClusteringPrefix serializer, either Clustering.serializer.serialize or Slice.Bound.serializer.serialize)
(*) IndexInfo.lastName (ClusteringPrefix serializer, either Clustering.serializer.serialize or Slice.Bound.serializer.serialize)
(long) IndexInfo.offset (vint encoded since 3.0, 64bit int before)
(long) IndexInfo.width (vint encoded since 3.0, 64bit int before)
(bool) IndexInfo.endOpenMarker != null (if 3.0)
(int) IndexInfo.endOpenMarker.localDeletionTime (if 3.0 && IndexInfo.endOpenMarker != null)
(long) IndexInfo.endOpenMarker.markedForDeletionAt (if 3.0 && IndexInfo.endOpenMarker != null)
Regarding the IndexInfo.offset and .width fields there are two options.
- Serialize both of them or
- Serialize only the offset field plus a last byte offset to be able to recalculate the width of the last IndexInfo
The first option is probably the simpler one, the second saves a few bytes (those of the vint encoded width).
EDIT: update vint fields (as per
EDIT2: add header-length fields (as per