borgbase/vykar v0.6.3

on GitHub

Memory Optimization

This release dramatically reduces peak memory usage across all operations through mmap-backed storage and streaming processing:

• Mmap-backed pack assembly: Replace heap pack buffers with mmap-backed assembly, reducing backup heap usage from 384 MiB to ~4 MiB
• Pack buffer pool memory reduction: ~40% reduction in pack buffer pool memory (640→384 MiB) before mmap conversion
• Tiered dedup index: Add mmap cache and xor filter for memory-efficient chunk deduplication
• Index-free extract: Restore files without loading the full chunk index into memory, using mmap restore cache
• Streaming item serialization/deserialization: Eliminate temp allocations during backup and reduce extract peak memory
• Incremental index update: Reduce save_state() peak memory by flushing incrementally (~138M saved)
• Open-phase optimization: Reduce open-phase peak memory from ~227M to ~128M
• Restore planner compaction: Use SmallVec and filtered chunk index for ~40% better throughput

Features

• cache_dir config option: Control where mmap temp files and cache data are stored
• Reuse ZSTD compression context: Thread-local Compressor reuse reduces allocation churn
• Pre-size FileCache HashMap: Use old cache length for initial capacity

Bug Fixes

• Release pack buffer pool in save_state() to avoid retaining memory past its useful life
• Recalculate pack writer targets in load_chunk_index() for correct sizing

Infrastructure

• Add profiling build option to release workflow
• Add linux_only option for manual release workflow runs

Downloads

Verify checksums with sha256sum -c SHA256SUMS.