Known Issues¶
PsiTri is under active development
The following are known issues and limitations. They are tracked here for transparency.
File Size Management¶
PsiTri's copy-on-write architecture means dead copies accumulate between compaction passes. The background compactor continuously reclaims space, but the file size can be 2-4x the reachable data size during sustained write workloads.
Use compact_and_truncate() after bulk operations to reclaim disk space:
Snapshot Pressure on Fan-Out¶
Long-lived snapshots compete with live nodes for cacheline slots in inner nodes. Under heavy snapshot pressure, inner node fan-out can degrade from 256 to as low as 16 branches per node. The compactor can relocate object data but cannot reassign ptr_address slots to new cacheline positions.
Fan-out recovers when snapshots are released. See Scaling Limits for analysis.
Recovery Time¶
Recovery time is proportional to database size (must scan segments + walk reachable tree), not proportional to uncommitted work. For very large databases, recovery after a crash may take significant time. The deferred_cleanup mode provides fast restart by deferring the expensive leak reclamation.
Multi-Writer Concurrency¶
Multi-writer support is in progress. Currently, concurrent writes to the same root index are serialized via mutex. Writes to different root indices (out of 512 available) are fully concurrent.