Implement concurrent compaction

[rmind]

Currently, the index structures are generally append-only as the document removal uses tombstones (special markings) to indicate deletions. Therefore, many deletions would produce gaps in the index files which would waste space. We want to address this problem by implementing compaction.

The following is a proposal for a concurrent compaction algorithm:

new-dtmap = exclusive-open-create
lock new-dtmap

// Initial concurrent sync (captures most of the data)
dtmap-sync from current-dtmap

lock current-dtmap
  // Sync any remaining data with the lock held if there was a race
  dtmap-sync from current-dtmap

  // Make the new index globally visible
  atomic-posix-rename new-dtmap.filename to current-dtmap.filename

  // Publish the compaction offset for the active index references
  atomic-store-release current-dtmap.compaction-offset <= last offset in new-dtmap
unlock current-dtmap
unlock new-dtmap

• The active index references check for a compaction-offset change, re-open the index (picking up the new file) and sync from this offset. The existing references to the memory-mapped file itself, primarily idxdoc_t::offset, would require a sequential scan to be adjusted.
• The next compaction may be triggered only when all active index references have synced. The index deletion and re-creation should be separately handled with the timestamp/generation number to avoid the ABA problem.